Store command center
OPS · 34 projects · 1 agent · 2 flows
TSK-Bench · Task completion benchmark for AI models
TSK-1 Benchmark
Same request. Every AI. Scored on what it actually finished.
GPT follows long requests word for word. Claude finishes apps best, on your own key in automations and in agents on Pro and up. DeepSeek leads on design. Kimi builds rich apps, but slowly. Not sure? TSK-1 Auto picks for you.
TSK-1 Benchmark at a glance
- Test days
- 27
- AI labs tested
- 11
- Versions tested
- 36
- Newest test
| Date | GPT | Claude | DeepSeek | Kimi | Grok | GLM | Qwen | Gemini | MiniMax | Mistral |
|---|---|---|---|---|---|---|---|---|---|---|
| 1, tested this day | 1, tested this day | 0 | 0 | 0 | 1, tested this day | 0 | 0 | 0 | 0 | |
| 1 | 2, tested this day | 0 | 1, tested this day | 0 | 1 | 0 | 0 | 1, tested this day | 0 | |
| 2, tested this day | 3, tested this day | 1, tested this day | 1 | 0 | 2, tested this day | 0 | 1, tested this day | 1 | 0 | |
| 2 | 3 | 2, tested this day | 1 | 0 | 2 | 0 | 1 | 1 | 0 | |
| 3, tested this day | 4, tested this day | 3, tested this day | 1 | 0 | 2 | 0 | 2, tested this day | 1 | 0 | |
| 4, tested this day | 4 | 3 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 4 | 4 | 4, tested this day | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 4 | 5, tested this day | 5, tested this day | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 5, tested this day | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 6, tested this day | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 7, tested this day | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 7 | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | |
| 8, tested this day | 6, tested this day | 6, tested this day | 2, tested this day | 1, tested this day | 3, tested this day | 1, tested this day | 2 | 2, tested this day | 0 | |
| 8 | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 0 | |
| 9, tested this day | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 0 | |
| 10, tested this day | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 0 | |
| 11, tested this day | 6 | 6 | 2 | 1 | 4, tested this day | 1 | 2 | 2 | 0 | |
| 11 | 6 | 6 | 2 | 1 | 4 | 1 | 2 | 2 | 0 | |
| 12, tested this day | 7, tested this day | 7, tested this day | 3, tested this day | 2, tested this day | 5, tested this day | 1 | 2 | 2 | 0 | |
| 13, tested this day | 8, tested this day | 8, tested this day | 3 | 2 | 5 | 1 | 2 | 2 | 0 | |
| 14, tested this day | 9, tested this day | 9, tested this day | 4, tested this day | 2 | 5 | 1 | 2 | 2 | 0 | |
| 15, tested this day | 10, tested this day | 10, tested this day | 4 | 3, tested this day | 6, tested this day | 1 | 2 | 2 | 1, tested this day | |
| 16, tested this day | 10 | 10 | 4 | 3 | 6 | 1 | 2 | 2 | 1 | |
| 16 | 11, tested this day | 10 | 4 | 4, tested this day | 6 | 1 | 2 | 2 | 1 | |
| 17, tested this day | 12, tested this day | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 | |
| 18, tested this day | 12 | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 | |
| 19, tested this day | 12 | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 |
How a TSK-1 test runs
- One requestEvery model gets the same words.
- BuildThe model builds the app in Taskade Genesis.
- Use itWe enter data like a customer and check that it saved.
- Change itWe ask for one follow-up change.
- Grade itInterface, Task, Memory and Adapt.
How we grade these apps: read the full method.
How do AI models rank on TSK-1?
ModelInterfaceHow complete and polished the finished app feels.TaskTask completion: how fully the app does what you asked, word for word.MemoryHow reliably the app keeps what people add to your workspace.AdaptHow cleanly the app changes when you ask.TSK ScoreHow well the model finished our tasks, out of 100. Each of the four columns gives up to 4 points (Leading 4, Strong 3, Emerging 2, Limited 1), so 16 points make 100. Under each model: the days we tested it, and how many apps it built in the side-by-side tests on this page.
- GPTOpenAI GPTDetailed requests and fast deliveryRanked 1.Interface: Strong.Task: Leading.Memory: Strong.Adapt: Leading.TSK Score 88 of 100, 14 of 16 points. Tested on 19 days · 29 builds side by side.
- ClaudeAnthropic ClaudePolished apps that keep improvingRanked 2.Interface: Strong.Task: Strong.Memory: Emerging.Adapt: Leading.TSK Score 75 of 100, 12 of 16 points. Tested on 12 days · 20 builds side by side.
- DeepSeekDeepSeekPolished apps with rich workspace dataRanked 2.Interface: Leading.Task: Strong.Memory: Strong.Adapt: Emerging.TSK Score 75 of 100, 12 of 16 points. Tested on 10 days · 17 builds side by side.
- KimiMoonshot KimiRich apps that need patienceRanked 2.Interface: Strong.Task: Strong.Memory: Strong.Adapt: Strong.TSK Score 75 of 100, 12 of 16 points. Tested on 4 days · 1 build side by side.
- GrokxAI GrokFollow-up edits that landRanked 2.Interface: Strong.Task: Strong.Memory: Strong.Adapt: Strong.TSK Score 75 of 100, 12 of 16 points. Tested on 4 days · 17 builds side by side.
Show every model
- GLMZhipu GLMGood judgment when details are unclearRanked 6.Interface: Emerging.Task: Strong.Memory: Emerging.Adapt: Strong.TSK Score 63 of 100, 10 of 16 points. Tested on 6 days · 2 builds side by side.
- QwenAlibaba QwenLong, detailed formsRanked 7.Interface: Emerging.Task: Strong.Memory: Emerging.Adapt: Limited.TSK Score 50 of 100, 8 of 16 points. Tested on 1 day · no side-by-side results published yet.
- GeminiGoogle GeminiA fast-improving Google modelRanked 8.Interface: Limited.Task: Emerging.Memory: Emerging.Adapt: Emerging.TSK Score 44 of 100, 7 of 16 points. Tested on 2 days · no side-by-side results published yet.
- MiniMaxMiniMaxNot scored in this testNot ranked.Interface: Not scored.Task: Not scored.Memory: Not scored.Adapt: Not scored.TSK Score not yet available. Tested on 2 days · no side-by-side results published yet.
- MistralMistralNot scored in this testNot ranked.Interface: Not scored.Task: Not scored.Memory: Not scored.Adapt: Not scored.TSK Score not yet available. Tested on 1 day · 1 build side by side.
- NemotronNVIDIA NemotronNo longer in the Taskade model menuScore pending
| Rank | Model | Interface | Task | Memory | Adapt | TSK Score | Points | Builds in side-by-side tests | Days tested | Newest version tested |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT | Strong | Leading | Strong | Leading | 88 out of 100 | 14 of 16 points | 29 | 19 | GPT-6.1 Sol |
| 2 | Claude | Strong | Strong | Emerging | Leading | 75 out of 100 | 12 of 16 points | 20 | 12 | Claude Sonnet 5 |
| 2 | DeepSeek | Leading | Strong | Strong | Emerging | 75 out of 100 | 12 of 16 points | 17 | 10 | DeepSeek V4.1 Flash |
| 2 | Kimi | Strong | Strong | Strong | Strong | 75 out of 100 | 12 of 16 points | 1 | 4 | Kimi K3 |
| 2 | Grok | Strong | Strong | Strong | Strong | 75 out of 100 | 12 of 16 points | 17 | 4 | Grok 4.7 |
| 6 | GLM | Emerging | Strong | Emerging | Strong | 63 out of 100 | 10 of 16 points | 2 | 6 | GLM 5 |
| 7 | Qwen | Emerging | Strong | Emerging | Limited | 50 out of 100 | 8 of 16 points | 0 | 1 | Qwen 3.7 Plus |
| 8 | Gemini | Limited | Emerging | Emerging | Emerging | 44 out of 100 | 7 of 16 points | 0 | 2 | Gemini 3.6 Flash |
| Not ranked | MiniMax | Not scored | Not scored | Not scored | Not scored | Not scored | Not scored | 0 | 2 | MiniMax M3 |
| Not ranked | Mistral | Not scored | Not scored | Not scored | Not scored | Not scored | Not scored | 1 | 1 | Mistral Large 4 |
- Leading
- Strong
- Emerging
- Limited
Treat this ranking as a guide. The tests ran from Jul 30, 2026 to Oct 23, 2026, and the test days differ by model. Most results come from one build.
Start building
Describe the app you need. TSK-1 coordinates the rest.
What did each test find?
Benchmark updates, newest first.
Terra against Sol, three apps
GPT-5.6 Terra and GPT-6.1 Sol built the client sign-up form, a newer family intake form and a Dota 2 match tracker. One grader scored each app without knowing which model built it. Terra finished each build in about 5 to 7 minutes, about twice as fast as Sol. Sol built the more complete intake form and match tracker. On the sign-up form, both Terra builds scored each applicant with an AI step, and the Sol build left the scoring to staff.
Five builds of each app
GPT-5.6 Luna built the reading list, the family intake form with its follow-up question, and the Dota 2 match tracker five times each. One grader scored every app without knowing which model built it. Every build saved its data, and every follow-up question saved and showed on the teacher page. Most builds took about 4 to 6 minutes. Only one intake form saved the date and time of each submission that the request asked for.
Three builds, not one
Claude Sonnet 5 built the reading list, the family intake form and an Our Families page three times each, and GPT-5.6 Luna built the Dota 2 match tracker five times. Two graders scored each app without knowing which model built it. Every build we could check saved its data, but repeats changed the results: none of the Sonnet 5 intake forms saved the time each form was sent, and the five match trackers ranged from complete to having gaps.
A reading list joins the test set
A reading list app joined the test set. Claude Sonnet 5, Grok 4.7 and Grok Build 0.1 built it and rebuilt the family intake form, and each follow-up question landed. Each model then added an Our Families page to an intake app that another model had built. One build per model and app, so read it as an early signal.
GPT-6 Luna on the match tracker
GPT-6 Luna built the Dota 2 match tracker with every detail the request named, a coach agent and a weekly recap, then added a heroes page in a follow-up change and put it live. One build, so read it as an early signal.
Show all 27 updatesShow fewer
More models on two real apps
A wide field of models and settings rebuilt the family intake form and a Dota 2 match tracker, and then took one follow-up change each. Grok 4.7 and GPT-5.6 Sol built both apps well, several builds showed sample families or matches as if they were real, and a few forms could not save at all.
A phone form with photos
A real family intake form joined the test set: parents fill it in on a phone with two to four photos, and the teacher sees every family on one page. GPT-6.1 Sol and Claude Opus 5.5 shared the top result, and every model that finished also added a follow-up question cleanly.
One match tracker, model by model
Several models built the same match tracker. GPT-5.6 Luna was the fastest complete build, DeepSeek V4.1 Flash ran a coaching automation on the first match logged, and two models limited their match forms to a short list of heroes.
Real-world apps and follow-up edits
Three real-world apps joined the set: a sales CRM, a fleet inspection, and an AI governance office. Every model that took the full set of follow-up edits passed it.
Seventeen models, one sign-up form
Seventeen models built the same client sign-up form from the same request, word for word.
Same request, four setups
We ran the same request on identical setups to see how much results change by chance. They differed noticeably, so we now only claim a difference between models when it is larger than that gap, and we grade on what the app does.
Same result after a Taskade update
We re-ran earlier tests after a large Taskade update. The apps came out the same, with fewer failed build steps along the way.
Sixteen builds, five models, identical prompts
Five models built and edited the same apps from identical prompts. GPT-5.6 Terra was the quality pick. Most builds now keep a checklist of the request inside the app.
Cleaner follow-up edits
Follow-up edits now start from a map of the app, and they landed cleaner. GPT-5.6 Luna edited apps other models had built without breaking them, and a missing page now explains itself.
Two new model families, three new apps
Qwen and Grok entered the benchmark, and three real-world apps joined the test set: a sales CRM, a field inspection audit, and an AI governance office. Every model still gets the same words.
A dedicated follow-up-edit test
We added a dedicated follow-up-edit pass and stricter instruction checks. On its first run, none of the 31 edit attempts finished - a strong first build is not enough on its own; an app must also change cleanly and stay within the brief.
Faster builds, verified results
A complete app finished in under five minutes without follow-up. We also rechecked every published result.
Accuracy came with clear explanations
The strongest results reproduced a long brief exactly and explained their design choices clearly.
A long form, built in full
An efficient build completed a long form, while the leading result also saved every submitted field correctly.
Long briefs became a harder test
Several builds followed the brief word for word, and one handled a 60-field form without cutting requirements.
Working software became the baseline
We made a running app the entry requirement. Attractive results that failed to open no longer received a score.
The first completely clean test
The first completely clean test showed that careful model selection can improve quality.
Brief fidelity and saved data
The strongest builds matched every requested field, stored submitted data correctly, and remained useful from form to follow-up.
Prompt length changed the outcome
Shorter instructions improved brief matching and edits, but sometimes reduced visual quality. The benchmark now balances all four dimensions.
Nine models, one consistent standard
Nine models built the same app. Several produced polished designs, but only working results qualified for comparison.
Self-checking improved reliability
The most dependable results detected and repaired problems before finishing, while weaker builds reported success too early.
Restraint mattered
The strongest designs followed the brief without adding unwanted gates or steps. Doing only what was asked became part of quality.
How did each model do on the same request?
The family intake form, five builds by one model
The request
Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.
The reading list, five builds by one model
The request
A simple reading list: books with a title, an author and a status (to read, reading or done), on one page that lists them and filters them by status.
The family intake form, three builds by one model
The request
Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.
Works, with gaps
The reading list, three builds by one model
The request
A simple reading list: books with a title, an author and a status (to read, reading or done), on one page that lists them and filters them by status.
The family intake form, three more builds
The request
Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.
Works, with gaps
A reading list app
The request
A simple reading list: books with a title, an author and a status (to read, reading or done), on one page that lists them and filters them by status.
The family intake form, more models
The request
Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.
Works
Works, with gaps
- GPT-5.6 Luna2 builds (best works, worst has gaps)
- GPT-5.6 Terra3 builds (best works, worst has gaps)
- Grok 4.31 build
- Claude Haiku 4.51 build
- Claude Sonnet 4.51 build
- Grok Build 0.11 build
- GLM 4.71 build
- DeepSeek V4 Pro, high thinking1 build
- Mistral Large 41 build
- DeepSeek V3.22 builds (same result)
- Grok 4.202 builds (same result)
A Dota 2 match tracker with a heroes page
The request
One player logs each match with hero, result, kills, deaths, assists, duration and notes, a dashboard shows win rate and streaks in a premium esports look, and then we ask for a heroes page.
Works
- Grok 4.71 build
- GPT-5.6 Sol1 build
- DeepSeek V4 Pro, max thinking1 build
- DeepSeek V4 Pro, high thinking1 build
- DeepSeek V4 Pro 08132 builds (same result)
- DeepSeek V4 Pro1 build
- Grok 4.51 build
- Grok Build 0.11 build
- GPT-5.6 Luna2 builds (same result)
- Claude Sonnet 52 builds (same result)
- GPT-5.6 Terra2 builds (same result)
A family intake form on a phone
The request
Parents fill in a form on their phone with two to four photos, the teacher sees every family on one page, and then we ask for one more question on the form.
A match tracker
The request
One player logs each game with hero, result and notes, and a dashboard shows win rate and streaks.
Most results come from one build. Where a model built the app more than once, we show its best and worst build. Read the results as a guide, not a final ranking.
Which model versions did we test?
Open a version to see what it built.
How do two models compare?
Preview runs
Early runs from our newest tests that the ranking does not count yet. Each one used different model settings than Taskade uses, and some were design tests run by hand. Each run says whether it finished an app and what it built. They get no points and no rank, so read them as notes, not results.
217 preview runs, newest Nov 10, 2026
Claude Fable 5.236 preview runs, newest Nov 10, 2026Preview, not ranked
Anthropic · Not in Taskade
| Asked for | Default | Low | Medium | High | Extra-high | Max |
|---|---|---|---|---|---|---|
| A Dota 2 match tracker from a one- or two-word request | Ran out of test time | Finished an app | Finished an app | Ran out of test time | Ran out of test time | Ran out of test time |
| An app from the one-word request "prometheus" | Ran out of test time | Finished an app | Finished an app | Ran out of test time | Ran out of test time | Ran out of test time |
| A staffing agency candidate placement intake | Ran out of test time | Ran out of test time | Ran out of test time | Ran out of test time | The test ended early | The test ended early |
| A warmer feel for the family intake form | Finished an app | Finished an app | Finished an app | Finished an app | Ran out of test time | Ran out of test time |
| A history page added to the family intake form | Finished an app | Finished an app | Finished an app | Finished an app | Ran out of test time | Ran out of test time |
| An "Our Families" page added to the family intake form | Ran out of test time | Finished an app | Finished an app | Ran out of test time | The test ended early | The test ended early |
Finished an appRan out of test timeThe test ended early
· Default thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a Dota 2 match tracker from a one- or two-word request.
Data projects, a coach agent and a weekly automation, but no app pages.
It thought about as long as the run at high. The test time ran out before any app page was written.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Three pages with a match log form, match history, per-hero stats, a win-rate chart, a coach and a weekly automation, in light and dark looks.
On a phone the win-rate chart ran past the edge of the screen, and a logged match could not be edited or deleted.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Three pages with a win and loss form strip, match history with filters and search, a hero pool, and a weekly recap that the dashboard reads, in light and dark looks.
A logged match could not be edited or deleted.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a Dota 2 match tracker from a one- or two-word request.
Three pages that save, change and delete matches, twelve hero cards and a lane-map design, in light and dark looks.
The test time ran out before its closing summary and before the app went live. On a phone some match columns were cut off.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a Dota 2 match tracker from a one- or two-word request.
The test time ran out before any app page was written.
· Max thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a Dota 2 match tracker from a one- or two-word request.
The test time ran out before any app page was written.
· Default thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an app from the one-word request "prometheus".
Data projects with sample rows, an agent and a weekly automation, but no app pages.
The test time ran out before any app page was written.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
An on-call alert board in a night-shift look: alerts that move from open to resolved, a runbook list that each alert links to, and a weekly page that an automation fills.
It read the one word as an alerting tool for an on-call team. Alerts had no assignee and could not be deleted.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A "Night Watch" alert board with filters, search, acknowledge, resolve and fix notes, and a runbook page with search, add and edit.
Alerts had no assignee and could not be reopened, edited or deleted.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an app from the one-word request "prometheus".
Data projects, an agent and an automation, but no app pages.
The test time ran out before any app page was written.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an app from the one-word request "prometheus".
The test time ran out before any app page was written.
· Max thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an app from the one-word request "prometheus".
The test time ran out before any app page was written.
· Default thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
No model in this test wrote an app page for this request within the test time.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
No model in this test wrote an app page for this request within the test time.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
No model in this test wrote an app page for this request within the test time.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
No model in this test wrote an app page for this request within the test time.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: a staffing agency candidate placement intake.
No app was built, so this run says nothing about the model.
· Max thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: a staffing agency candidate placement intake.
No app was built, so this run says nothing about the model.
· Default thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A warmer palette on the family intake form, with every feature kept.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
Peach cream and warm brown-black backgrounds, with every feature kept.
The main button stayed green, so the shift was modest.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A warmer palette on the family intake form, with every feature kept.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A warmer palette on the family intake form, with every feature kept.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a warmer feel for the family intake form.
A warmer palette, and a fix for a print style that was hard to read in the dark look.
The test time ran out before its closing summary.
· Max thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a warmer feel for the family intake form.
The test time ran out before the change was made.
· Default thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A history page that lists every family form by day, newest first, linked from the menu.
It also changed two files that the request did not name, with no visible effect.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A timeline page of every family form in the look of the app, linked from the menu.
One line said the list was in arrival order while it showed newest first.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A timeline page of every family form in the look of the app, linked from the menu.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A timeline page of every family form in the look of the app, linked from the menu.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a history page added to the family intake form.
A timeline page of every family form, linked from the menu.
The test time ran out before its closing summary.
· Max thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a history page added to the family intake form.
The test time ran out before the change was made.
· Default thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
The test time ran out before the change was made.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in, with a family card grid, a profile view, a consent filter and a "View Our Families" button.
It hid private fields on the page, not in the saved data, and the new button sent a parent who had just submitted to a sign-in screen.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in, with a family card grid and a profile view.
It hid private fields on the page, not in the saved data, and it removed the Teacher link from the header for everyone but staff.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
The test time ran out before the change was made.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: an "Our Families" page added to the family intake form.
No app was built, so this run says nothing about the model.
· Max thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: an "Our Families" page added to the family intake form.
No app was built, so this run says nothing about the model.
Claude Fable 5.14 preview runs, newest Nov 10, 2026Preview, not ranked
Anthropic · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Three pages with a scoreboard match list, a heroes page with twelve hero cards and counters, and a coach notes timeline, in light and dark looks.
The header ran past the edge of a phone screen, and a logged match could not be edited.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
Eight data projects and five agents, but no app pages and no working automations.
No model in this test wrote an app page for this request within the test time.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A clear warm shift: terracotta buttons, peach cream paper and a warm brown dark look, with every feature kept.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
Part of the change: a sign-in step in front of the families page.
The test time ran out partway through the change, before the card grid appeared.
Claude Opus 5.54 preview runs, newest Nov 10, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Three pages with green and red team roles, a win-rate ring, a hero pool and a weekly recap automation, in light and dark looks.
The dashboard listed only the eight newest matches, and a logged match could not be edited or deleted.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
Eight data projects, five agents and seven automations, but no app pages.
It got the furthest on this request. The test time ran out before any app page was written.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A clear warm shift: terracotta buttons, honey cream paper and a warm brown dark look, with layout and type kept.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in, with a family card grid and a profile view.
It hid private fields on the page, not in the saved data, and it removed the helper chat and the Teacher link without asking.
GPT-5.6 Terra11 preview runs, newest Nov 10, 2026Preview, not ranked
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Two pages with a match form that saves K/D/A and role, and a match history.
On a laptop the stats row ran over the match history, and it saved no rank points or gold per minute.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A personal goals tracker in a warm observatory look, with a theme switch and weekly summaries.
Its link to the word was only the name and a flame icon. On a phone the home page ran past the edge of the screen.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a staffing agency candidate placement intake.
An apply page with all 21 questions word for word, nine staff tabs, a 100-point match score form, five agents and seven automations.
Its apply form saved empty answers, and every automation only sent an email.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A terracotta and cream palette, with every feature kept.
The print style kept the old green.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A timeline history page with chips for each family, linked from the menu.
It had no search or grouping by date.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page with family profiles behind a sign-in.
Profiles printed placeholder lines for blank answers, and it removed the helper chat without asking.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a nonprofit intake and case management app.
An intake form, a case workspace, and automations that kept saving answers after the chat was summarized.
It asked to confirm who can open the app, even though that choice was already saved.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a clinic patient intake form.
A clinic form and a review page.
Its review page was open to anyone, and it offered a staff sign-in only as a suggestion.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
One page in ember orange on warm paper, with a large headline.
It built one page only.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A one-question-at-a-time form with a progress bar, all 32 questions word for word, and AI scoring with the full formula.
A public link pointed to the staff dashboard, and its closing summary did not say that the page was open.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Two pages, three sample matches, a private coach, a public guide and a weekly automation.
On a phone its match form ran past the edge of the screen.
GPT-6 Luna42 preview runs, newest Nov 10, 2026Preview, not ranked
| Asked for | Low | Medium | High | Extra-high |
|---|---|---|---|---|
| A Dota 2 match tracker from a one- or two-word request | Finished an appFinished an appFinished an app | Finished an appFinished an appFinished an app | Finished an appFinished an appFinished an app | Finished an appRan out of test timeRan out of test time |
| An app from the one-word request "prometheus" | Finished an app | Finished an app | Finished an app | Ran out of test time |
| A staffing agency candidate placement intake | Finished an app | Ran out of test time | Ran out of test time | The test ended early |
| A warmer feel for the family intake form | Finished an app | Finished an app | Finished an app | Finished an app |
| A history page added to the family intake form | Finished an app | Finished an app | Finished an app | Finished an app |
| An "Our Families" page added to the family intake form | Finished an appFinished an appFinished an app | Ran out of test timeRan out of test timeRan out of test time | Ran out of test timeRan out of test timeRan out of test time | Ran out of test timeRan out of test timeRan out of test time |
Finished an appRan out of test timeThe test ended early
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match log and search.
It saved hero, result and notes but no K/D/A, rank points or gold per minute, and a match could not be edited or deleted.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with summary tiles, a match form and a match history.
It saved no gold per minute or match length, had no chart, filter or search, and a match could not be edited or deleted.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One short page that records hero, result, role and notes.
It saved no date, match length or stats, and had no chart, filter, search, edit or delete.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One calm page in light and dark looks with a match form and a match list.
It saved hero, result, date, K/D/A, mode and notes but no role or rank points, and a match could not be edited or deleted.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with summary tiles, a match form and a match list.
A test match sent from the form did not save, and it had no chart, filter or search.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match form and a match list.
It saved no role, rank points or gold per minute, and had no chart, filter or search. The test ended before the app went live.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match form, a match history and a coach chat.
It saved rank point changes but no K/D/A or role, and the coach button covered part of the match history on a laptop.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match form and a match list.
Hero and role were free text, and it had no chart, filter or search.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match form and a match list.
It saved no K/D/A, rank points or gold per minute, and a match could not be edited or deleted.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
One warm page in light and dark looks with a match form and a strip of recent results.
It saved rank point changes and game mode but no K/D/A or role, and the coach button covered the match form heading on a laptop.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match form and a match list.
The test time ran out before its closing summary and before the app went live.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a Dota 2 match tracker from a one- or two-word request.
One page with a match form, a match list and a weekly review automation.
The test time ran out before its closing summary and before the app went live.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A one-page monitoring board in a graphite look with three sample signals and a theme switch.
It read the word as service monitoring. The page was read-only, so nobody could record or update a signal in the app.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A one-page monitoring board.
The page was read-only and sent people to the raw project to make a change.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A mint and graphite monitoring page with search, two filters and a date range.
It said plainly that it was not connected to a live server. The page was read-only, so nobody could record a measurement.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an app from the one-word request "prometheus".
A calm service check log where people add services and record checks.
It said plainly that the app does not check services by itself. The test time ran out before its closing summary and before the app went live.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a staffing agency candidate placement intake.
A calm candidate portal with all 21 questions word for word, a recruiter page and a match score form.
It built one of the five agents and one of the seven automations asked for, and most staff views were read-only tables. On a phone the recruiter sign-in button was cut off.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
An apply page with all 21 questions word for word, eleven staff pages, five named agents and hold rules for checks and schedules.
The home page showed only a sign-in box, and the public header showed every staff tab to candidates. The test time ran out before its closing summary.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
The test time ran out before its pages were written.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: a staffing agency candidate placement intake.
No app was built, so this run says nothing about the model.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
Warmer backgrounds in light and dark looks, with every feature kept.
The main button stayed sage green, so the change was modest.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A warmer palette, with every feature kept.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A clay color for buttons and accents, with every feature kept.
In the dark look the button read salmon pink.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A terracotta color for buttons and accents, with every feature kept.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A plain history page that lists each family by name and time.
It showed nothing more than name and time, and one number on each row had no label.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A history page with each family name, nickname and time.
It did not show who sent the form or what they chose.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A timeline history page with chips for guardian, favorite and photo choice.
It had no search or grouping by date.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A history page with search that lists each family name, nickname and time.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in, with cover photo tiles and a family profile dialog.
It collected no sibling details, and it hid private fields on the page, not in the saved data.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in.
It collected no sibling details, and it left the Teacher page open to any signed-in family.
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in.
It collected no sibling details, and it removed the helper chat from the form without asking.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" directory behind a sign-in that an automation fills with the families who agreed to share, without contact details.
Siblings were one text field with no photos. The test time ran out before its closing summary.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in.
It left the Teacher page open to any signed-in family. The test time ran out before its closing summary and before the change went live.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page that checked the class sign-in in the browser only.
It removed the helper chat without asking. The test time ran out before the access step and its closing summary.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in.
The profile showed no siblings or family activities, and it hid the helper chat on the form. The test time ran out before its closing summary.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in.
Siblings were one text field with no photos, and it hid the helper chat for signed-out parents without asking. The test time ran out before its closing summary.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
The test time ran out before the change was made.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
The test time ran out before the new page was linked into the app.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
The test time ran out right after its first change.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
The test time ran out partway through the change, before anything linked to the new page.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
An eligibility intake with staff review, scoring and a dashboard that saved every assessment field.
The first question card counted down and picked the suggested answers on its own.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a property maintenance and work orders app.
A resident form with twenty questions, work-order views, and automations. The fullest app from that day of tests.
The build took a long time, and a later edit ended before it could send a closing message.
DeepSeek V4.1 Flash 101718 preview runs, newest Nov 10, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Pages for matches, heroes and recaps, with a win-rate gauge.
A match could not be edited, and one tile counted the heroes wrong.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
A tracker with a match table, a rank points chart, heroes and weekly recaps.
On a phone the match table and the menu ran past the edge of the screen.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A service operations console with a status page.
Its link to the word was the name more than the content, with no alert rules.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: a staffing agency candidate placement intake.
All 21 questions and all nine tabs, a matching rubric with hold and follow-up rules, and a dark look.
It built two of the five agents and four of the seven automations, some charts were empty, and candidate data had no staff sign-in. The test time ran out before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A clay button and a soft warm wash on the paper, with every feature kept.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A history page with stat tiles and an order switch, linked from the menu.
One line said the list was in arrival order while it showed newest first.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page with a sign-in card that links back to the form.
The Teacher page stayed open to any signed-in family. The test time ran out before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page with family profiles.
Family details were open to anyone with the link, and one form submit stopped working. The test time ran out before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: one more question on the family intake form.
Added the optional teacher question, saved it and showed it in a new teacher panel, with the photos still working.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form, organizer cards with status counts, a facesheet preview, a private helper agent and a Monday automation.
It added the send time to each card.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the family intake form.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A cream paper form with numbered cards and pastel doodles, and an organizer with filter counts, family cards and a facesheet preview.
The cards did not show when each form was sent.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the family intake form.
An organizer with family cards, a filter and a facesheet view.
Family details were open to anyone with the link. The test time ran out before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the client sign-up form with scoring.
A review desk with score bands, a page per applicant with every answer and an accept, waitlist or decline choice, a scoring automation and an agent.
The test time ran out before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the client sign-up form with scoring.
An editorial form, an overview, a scoring guide page and an agent.
Its scoring automation never ran. The test time ran out before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a ready-made app from the public app list.
Two ready-made apps with a clear layout, and both saved and showed the records they were asked for.
One app was not listed, because another featured app already used the same name.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a construction change-order tracker.
A change-order tracker that saved records, and an assistant that answered from those records.
The chat could not be summarized, and the app still finished.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A ten-step form that saves progress, a staff dashboard with counts, a chart and a decision for each applicant, a scoring guide page and AI scoring.
It grouped the questions out of their original order and shortened the saved field names.
GLM 5.3 Flash11 preview runs, newest Nov 10, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
A plain match log with a hero page.
No link led to the hero page, the list showed only five matches, and heroes came from a fixed list of eight.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
An incident board with severity levels, services and checks, with search and a status filter.
Nobody could create or update an incident in the app.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a staffing agency candidate placement intake.
Twelve pages with all 21 questions word for word and a match rubric with hard rules.
Two pages crashed, the dashboard counts read zero, and it built three of the seven automations.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A terracotta and cream palette, with every feature kept.
A few text colors stayed hard to read.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A history page grouped by month, with the menu label changed to Timeline.
One line said the list was in arrival order while it showed newest first.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page with small family cards.
It skipped both access steps, so family details were open to anyone with the link.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form with cream cards and a dusty pink button, and an organizer with photos, status, send time, filters and search.
Its agent tool did not work.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A form, a dashboard and an automation.
The chat button in the app opened an agent page that did not exist.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form that saved every answer and four photos.
Its teacher page sat behind a sign-in and showed a blank page, and its doodles did not draw. It spent many steps repeating one fix.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
Kimi K3.1 Pro12 preview runs, newest Nov 10, 2026Preview, not ranked
Moonshot · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
A tracker with win rate, rank points, K/D/A and gold per minute, a rank points chart, filters and a weekly automation.
Its sample matches counted in the stats, and on a phone the theme switch was cut off.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
A tracker with a match form and a K/D/A chart.
It saved no rank points, and a match could not be edited or deleted.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A monitoring app with services, incidents and alert rules, and a performance chart.
The chart used fixed sample values, and on a phone it ran past the edge of the screen.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a staffing agency candidate placement intake.
Eleven pages with all 21 questions word for word, nine staff tabs, a 100-point match score, five agents and seven automations.
Staff pages sat behind a sign-in, and the consent question could only be answered yes.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a warmer feel for the family intake form.
A warmer palette, with every feature kept.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a history page added to the family intake form.
A history page and menu links.
The page showed an error on every load, and its closing said the app worked.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in that an automation fills with the families who agreed to share.
The "View Our Families" button showed only for families who agreed to share.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an "Our Families" page added to the family intake form.
An "Our Families" page behind a sign-in.
Any invited family could open the Teacher page with contacts and teacher notes.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: one more question on the family intake form.
Added the optional teacher question in three files, saved it and showed it on the teacher page, with the photos still working.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form and a teacher page with family cards, eight filters, a status choice and a facesheet preview.
The facesheet showed the child photo even when the family said no to photo use.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A cream paper form with pencil doodles and numbered sections, and a staff page with family cards, a status filter, search and a full facesheet view.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A form with clear type and a scoring automation.
The staff page showed an error and no results.
Show all preview modelsShow fewer
GPT-OSS 120B2 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the client sign-up form with scoring.
It wrote a plan and stopped without building.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
GPT-OSS 20B1 preview run, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
DeepSeek V4.1 Flash5 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: one more question on the family intake form.
Added the optional teacher question, saved it and showed it on the teacher page, with the photos still working.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A cream ruled form and an organizer with photos, a status filter and a facesheet preview.
Family details were open to anyone with the link.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A cream grid-paper form with pencil hearts, and an organizer with family cards, photos, a status filter and a facesheet page for each family.
The build stopped before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
Polished cards and filter chips over a form and a submissions page.
A later change to its agent broke the scoring, and the build stopped before its closing summary.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the family intake form.
A form that saved every answer and four photos, and a teacher page with family cards, status filters and a preview. The only build of its test that saved when each form was sent.
The teacher page opened with no sign-in, so parents could see every family. It spent many steps repeating one fix.
DeepSeek V4 Pro 08133 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a staffing agency candidate placement intake.
A twenty-one-question recruiter form, five assistants, and automations. The assistants answered from the saved records.
The first reply stopped before it wrote anything. A shorter follow-up let the build continue.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
Our test setup cut its first reply short, and the run lost the request before it built the app.
DeepSeek V4 Flash 07311 preview run, newest Nov 6, 2026Preview, not ranked
DeepSeek · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
GLM 5.38 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A cream form with small doodles that saved the answers and photos.
The organizer page did not work.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the family intake form.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A plain form, a staff dashboard with counts and an average, a scoring automation and an agent.
It kept the question wording closest to the request in its test. Its agent could not state the score cutoff.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a ready-made app from the public app list.
Two ready-made apps that shipped on the public list after one follow-up each.
Each app needed one small fix, a color that was hard to read or a date on a card.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a grant application and review workflow.
A grant form and a review workspace. A follow-up repaired a program menu that would not save.
A few of its form replies were unreadable until a later pass repaired them.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form that saved a family with one caregiver.
A family with a second caregiver did not save, and a retry saved that family twice. Its doodles overlapped the page heading.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
GLM 5.24 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A seven-step phone form with large buttons, and an organizer.
Family details were open to anyone with the link.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A home page, the 32-question form, a staff table with filters, an automation and two agents.
It used a generic look, and on a phone the staff table cut off a column.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: an insurance broker client and renewal intake.
A broker intake on live data with a navy shell, though it saved only part of the requested answers.
It kept checking the app log after it was told to stop, until a later message ended that loop.
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the client sign-up form with scoring.
Two projects with the questions and a scoring guide, and an assistant.
It built no pages, and its automation was left empty and off.
GLM 4.7 Flash1 preview run, newest Nov 6, 2026Preview, not ranked
Zhipu · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
Kimi K35 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A phone form with numbered pastel sections, role chips and photo tiles.
The organizer page crashed on load.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A warm six-step form, a staff dashboard with search and counts, a scoring automation and a staff agent.
On a phone the result badge ran past the edge of the screen.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a ready-made app from the public app list.
Two featured apps with tidy pages, and both kept private records behind sign-in.
One app hid chat behind sign-in until a follow-up opened the public chat the request asked for.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a childcare enrollment form.
An enrollment form and a preschool workspace that kept its records linked, in under thirteen minutes.
It tried to read one field the form does not have.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form that saved every answer and four photos, with the closest fit to the requested look in its test. The fastest build of its test.
After sign-in its teacher page crashed, while its closing summary said the app worked.
Kimi K2.7 Code2 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the client sign-up form with scoring.
A home page, a six-page form that kept the original questions, and a staff dashboard.
Every entry sat at pending review while every counter read zero, and its scoring automation failed on its first run.
Kimi K2.62 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
Clean cream family cards with a filter and a facesheet preview.
It set no sign-in for the organizer, so family details were open to anyone with the link.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the client sign-up form with scoring.
The test time ran out before any app page was written.
Qwen 3.8 2.4T3 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form with cream and white cards, and an admin page with photos, a filter, a status choice and a facesheet preview.
Its closing summary matched what it built. The cards did not show when each form was sent.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A form, an admin dashboard, an automation and an agent.
The admin page had no link from the form and no sign-in.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A form that showed the exact thank-you text and saved every answer with four photos.
Its teacher page showed only a sign-in prompt, and the family records stayed readable without a sign-in until the user turned on access. It built an automation, then deleted it.
Qwen 3.8 27B2 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
Nemotron 3 120B2 preview runs, newest Nov 6, 2026Preview, not ranked
NVIDIA · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
Gemma 4 26B2 preview runs, newest Nov 6, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the client sign-up form with scoring.
It wrote a closing summary with links to pages that it had not built.
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the client sign-up form with scoring.
It called a tool that does not exist, then wrote a closing summary about projects, assistants and automations that it had not made.
Muse Spark Open5 preview runs, newest Nov 6, 2026Preview, not ranked
Meta · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A plain form and an organizer with family cards and filter chips.
The app had no sign-in, so family details were open to anyone.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A plain one-page form and a thin dashboard.
19 of the 32 answers did not get a field of their own, and one page did not open on a second check.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a ready-made app from the public app list.
Two featured apps after several follow-ups. Both looked finished on a phone and a laptop.
The first reply said the app was ready while the staff pages still crashed. Later follow-ups filled in the records and fixed who could open each page.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a cleaning operations quality app.
Cleaning checklists and quality views. One color pair was hard to read.
A later follow-up changed no visible screen.
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the family intake form.
It wrote a short plan and stopped without building.
Inkling3 preview runs, newest Nov 6, 2026Preview, not ranked
Thinking Machines · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the client sign-up form with scoring.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the family intake form.
No app was built, so this run says nothing about the model.
· Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: the family intake form.
No app was built, so this run says nothing about the model.
Granite 4.0 H Micro1 preview run, newest Nov 6, 2026Preview, not ranked
IBM · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the client sign-up form with scoring.
It answered in prose and stopped without building.
PhoneLLM Mini Alpha 21 preview run, newest Nov 6, 2026Preview, not ranked
Daily · Not in Taskade
· Preview run, not ranked: tested with different model settings than Taskade uses.
No working app. Asked for: the client sign-up form with scoring.
It asked a question in prose and stopped without building.
GPT-6.1 Terra4 preview runs, newest Oct 27, 2026Preview, not ranked
OpenAI · Not in Taskade
| Asked for | Low | Medium | High | Extra-high |
|---|---|---|---|---|
| A legal client intake and eligibility form | The test ended early | Not run | Not run | Not run |
| A revenue lead qualification intake | Not run | Finished an app | Not run | Not run |
| A field service operations app | Not run | Not run | Finished an app | Not run |
| A classroom operations app | Not run | Not run | Not run | Finished an app |
Finished an appThe test ended earlyNot run
· Low thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
The test ended early. Asked for: a legal client intake and eligibility form.
No app was built, so this run says nothing about the model.
· Medium thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a revenue lead qualification intake.
A public intake form with a five-step rail, a progress bar and a dark look, plus nine pages and a staff area behind an invite sign-in.
It did not compute the total score or the lead class, so the dashboard counts stayed empty.
· High thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a field service operations app.
A sign-in card and an intake form with all 32 questions word for word.
It built one page, with no dashboard or dispatch view and no score.
· Extra-high thinking · Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a classroom operations app.
A calm evergreen design with nine pages over one list of demo rows.
It chose a read-only demo over the requested system without asking, and its closing said so plainly.
Claude Opus 4.84 preview runs, newest Oct 23, 2026Preview, not ranked
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A generic dashboard with five pages, editable score cutoffs and four labeled sample entries.
Its automation only sent an email alert, so staff scored each answer by hand. A restart in our test setup interrupted the run before its first question, so it set its own score cutoffs and said so.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
Three pages, eight sample matches and a public coach.
It picked a dark look that the request did not ask for, and its page ran wider than a phone screen.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A form with generic feature cards and a stat row, and a scoring automation that ran once.
It changed one question to allow more than one answer and said so. Its home page said submissions were private while the staff dashboard was open to anyone, and its closing summary did not mention it.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the family intake form.
A calm cream form in one narrow column with pencil doodles, the only build of its test in the calm one-column look.
The teacher page sat behind the app sign-in. It checked a real submission before it finished.
Claude Opus 4.65 preview runs, newest Oct 23, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a clinic patient intake form.
A plain centered form with no header, and a review page.
Its review page was open to anyone, and it offered a staff sign-in only as a suggestion.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Ran out of test time. Asked for: the client sign-up form with scoring.
A plain form under a white top bar.
The test time ran out while it checked the form, before its closing summary.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A status page with a status banner, active incidents, grouped services, a history page and an admin page.
The admin page was open to anyone with the link, and its closing summary said so. In the light look the status text was hard to read.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A dark first view with a glowing accent, a plain light view, and AI scoring.
It chose the score cutoffs without asking and said so. On a phone the dashboard dropped the applicant name column, and a chart from its plan never appeared.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A generic admin dashboard with stat cards and charts, and a clean eight-section form that kept each question word for word.
Its saved fields shortened the long questions, which the request said not to do, and its three sample scores came from the model, not from the scoring automation.
Claude Sonnet 52 preview runs, newest Oct 23, 2026Preview, not ranked
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: an app from the one-word request "prometheus".
A cool slate and cyan operations dashboard with six pages, clear in the light and the dark look.
One section crashed for about a minute, and it found and fixed the crash itself. Every incident date showed one day early, and its closing summary called the app fully checked.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A calm six-step form with no header, a staff page behind an admin sign-in, and a scoring automation that ran once on a test entry.
After it summarized its own chat, it thanked the user for pasting the questions again, though no new message had been sent.
Claude Sonnet 4.61 preview run, newest Oct 23, 2026Preview, not ranked
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
A generic tracker: a small title, four stat tiles and a list.
Its red accent sat close to the color for a loss, it picked a dark look that the request did not ask for, and its closing summary said that a digest emailed the workspace when none did.
GPT-5.6 Sol7 preview runs, newest Oct 23, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a clinic patient intake form.
A clinic form, a review page and an automation that only adds a comment.
Its review page was open to anyone, and it offered a staff sign-in only as a suggestion.
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
The most finished first view of its test: a large headline and four stat tiles.
The form called itself confidential while the dashboard with every answer was the public home page.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A light first view with ruled lines and a score rail, and AI scoring with the cutoff it asked the user for.
It added no sample entries, so the review queue opened empty. Its closing summary said no staff sign-in existed yet.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A clean four-step form that kept every question word for word, and AI scoring that was set up but not yet run.
The dashboard was public and linked from the top bar, and its closing summary said it had no sign-in.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A warm linen first view with a score rail, a matching dark look, and AI scoring.
Its closing summary said the staff dashboard was open to anyone with the link. On a phone its chat button covered a label.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
The AI scoring that the request asked for, with the cutoffs from the user’s answer.
It asked who could open the app before it built. Its closing summary asked the user to turn on access that was already on.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
A warm ivory, teal and copper review workspace with all 32 questions word for word and a scoring guide.
It pointed out an unclear question in the request instead of guessing.
GPT-5.6 Luna3 preview runs, newest Oct 23, 2026Preview, not ranked
· Preview run, not ranked: tested with different model settings than Taskade uses.
Finished an app. Asked for: a clinic patient intake form.
A clinic form, a review page and an automation that only adds a comment.
Its review page was open to anyone, and it offered a staff sign-in only as a suggestion.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: the client sign-up form with scoring.
One long page of 32 question cards.
Its scoring step got the answers with no questions and no formula, and its closing diagram showed a scoring guide that the automation did not have.
· Preview run, not ranked: a design test run by hand, with different model settings than Taskade uses.
Finished an app. Asked for: a Dota 2 match tracker from a one- or two-word request.
It asked two questions, then built a match log, a hero reference and a weekly recap automation.
Its chat button covered the match list on a laptop, and the recap page stayed empty until the first weekly run.
FAQ
Which AI model is best for building apps?
GPT has the highest TSK Score, 88 out of 100, but the best pick depends on the app. Some models follow a request more closely, some finish faster, some look more polished. In Taskade you do not have to choose: TSK-1 Auto handles the default, and you can pick a model yourself when you want more control.
What can Taskade Genesis build with TSK-1?
Taskade Genesis can turn one request into CRMs, dashboards, portals, forms, trackers, and internal tools. TSK-1 coordinates the AI model, workspace memory, agents, and automations so the result can store information, answer questions, run workflows, and keep improving.
What does the TSK-1 Benchmark measure?
Whether a model can turn one prompt into a complete, working app. We grade four qualities of a living system: Interface, Task, Memory, and Adaptation: how finished it feels, how closely it follows your request, whether it keeps your data, and how cleanly it handles follow-up changes.
How is TSK-1 different from Vibe Code Bench, WebDev Arena, and Design Arena?
Those benchmarks grade code, a web page, or crowd votes on a screenshot. TSK-1 grades a running app: we open it, enter data the way a customer would, check that the data was kept, run its automation, and ask for a follow-up change. A build that looks right but loses your data does not score.
Does TSK-1 benchmark app builders or AI models?
AI models. Every model builds inside the same app builder, Taskade Genesis, from the same request, word for word, so the builder is held constant. A repeat test in August showed that two identical runs can still differ, so we only call out a difference between models when it is larger than that gap.
Do you test the apps or just the code?
We test the apps. Every build gets opened and used. We fill in its form the way a customer would, submit it, then check that the answers landed in the right place. An app that looks beautiful and loses your data fails.
Which models can I use in Taskade?
Taskade runs frontier models from top AI labs, and TSK-1 Auto handles the default. Built in, with no key, every plan can also pick TSK-Instant. The AI model directory lists every model, with its status, its plan and how to pick it.
How often is this updated?
We run a new test whenever a notable model ships. Every published result stays on this page, including the ones that did not go well. The newest results are from Oct 23, 2026.
What is TSK-1?
TSK-1 is the intelligence behind Taskade Genesis. It brings AI models, workspace memory, agents, and automations together so one request can become a working app. The TSK-1 Benchmark shows how that process performs on real app builds.
Can I try the same app request?
Yes. Every model in a comparison receives the same request, word for word. Describe the same idea in Taskade Genesis, choose a model or let TSK-1 Auto decide, and see what it builds.
Are the benchmarked apps publicly available?
Not yet one by one. We are listing every finished benchmark build under a single official creator so you can open and clone it. Until then, each model page explains what that model produced, and the App Kits on this page are live systems you can open and clone today.
How many builds are behind each result?
Each result comes from a small set of hands-on app builds, and most results rest on one build. In October one model built the same match tracker five times, and the builds ranged from complete to having gaps. So a small difference between two single builds is not a result. The TSK Score turns the four ratings (Leading, Strong, Emerging, Limited) into a number out of 100 so models are easy to compare. It is not 100 separate checks. Models with the same result share a rank.
Finance dashboard
DASHBOARD · 3 projects · 1 agent · 2 flows
Equipment maintenance tracker
OPS · 3 projects · 1 agent · 2 flows
Client pipeline CRM
CRM · 2 projects · 1 agent · 2 flows
Invoice generator
TOOL · 3 projects · 1 agent · 4 flows
Project portal
PORTAL · 3 projects · 1 agent · 5 flows
Sales CRM
CRM · 3 projects · 1 agent · 2 flows
Sales pipeline dashboard
DASHBOARD · 3 projects · 1 agent
Project dashboard
TRACKER · 3 projects · 1 agent · 2 flows
Team capacity planner
OPS · 2 projects · 3 agents · 2 flows
Growth dashboard
DASHBOARD · 3 projects · 1 agent · 1 flow
AI productivity planner
TOOL · Tasks · calendar · insights