AI model directory
Every AI model in Taskade, old and new
See the models you can pick now, the ones that came before, and how each one builds apps in our tests.
Best models by job
The models you can pick in Taskade now, by what they do best. The results come from TSK-1, and the specs from the model makers.
| Job | Models | Why |
|---|---|---|
| Building appsWorks on 3 of 3 requests in the TSK-1 Build arena. | Works on 3 of 3 requests in the TSK-1 Build arena. | |
| Follow-up code changesWorks on 2 of 2 changes in the TSK-1 Code arena. | Works on 2 of 2 changes in the TSK-1 Code arena. | |
| Long documentsReads the most at once: a context window of 1.05M tokens, per the model maker. A token is about three quarters of a word. | Reads the most at once: a context window of 1.05M tokens, per the model maker. A token is about three quarters of a word. | |
| Open modelsThe newest open-weight models: their makers let anyone download and run them. | The newest open-weight models: their makers let anyone download and run them. | |
| The newest releaseReleased Sep 29, 2026, per the model maker. | Released Sep 29, 2026, per the model maker. |
Published specs from the model maker, via models.dev (MIT).
| Date | OpenAI | Anthropic | DeepSeek | Moonshot | xAI | Zhipu | Alibaba | MiniMax | NVIDIA | Mistral | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Jul 30, 2026 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 |
| Jul 31, 2026 | 1 | 2 | 0 | 1 | 0 | 1 | 0 | 0 | 1 | 0 | 0 |
| Aug 1, 2026 | 2 | 3 | 1 | 1 | 0 | 2 | 0 | 1 | 1 | 0 | 0 |
| Aug 2, 2026 | 2 | 3 | 2 | 1 | 0 | 2 | 0 | 1 | 1 | 0 | 0 |
| Aug 3, 2026 | 3 | 4 | 3 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 4, 2026 | 4 | 4 | 3 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 5, 2026 | 4 | 4 | 4 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 6, 2026 | 4 | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 7, 2026 | 5 | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 8, 2026 | 6 | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 19, 2026 | 7 | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 20, 2026 | 7 | 5 | 5 | 1 | 0 | 2 | 0 | 2 | 1 | 0 | 0 |
| Aug 25, 2026 | 8 | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 1 | 0 |
| Aug 26, 2026 | 8 | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 1 | 0 |
| Aug 27, 2026 | 9 | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 1 | 0 |
| Aug 28, 2026 | 10 | 6 | 6 | 2 | 1 | 3 | 1 | 2 | 2 | 1 | 0 |
| Aug 29, 2026 | 11 | 6 | 6 | 2 | 1 | 4 | 1 | 2 | 2 | 1 | 0 |
| Sep 18, 2026 | 11 | 6 | 6 | 2 | 1 | 4 | 1 | 2 | 2 | 1 | 0 |
| Sep 19, 2026 | 12 | 7 | 7 | 3 | 2 | 5 | 1 | 2 | 2 | 1 | 0 |
| Sep 22, 2026 | 13 | 8 | 8 | 3 | 2 | 5 | 1 | 2 | 2 | 1 | 0 |
| Sep 30, 2026 | 14 | 9 | 9 | 4 | 2 | 5 | 1 | 2 | 2 | 1 | 0 |
| Oct 8, 2026 | 15 | 10 | 10 | 4 | 3 | 6 | 1 | 2 | 2 | 1 | 1 |
| Oct 10, 2026 | 16 | 10 | 10 | 4 | 3 | 6 | 1 | 2 | 2 | 1 | 1 |
| Oct 18, 2026 | 16 | 11 | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 | 1 |
| Oct 20, 2026 | 17 | 12 | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 | 1 |
| Oct 21, 2026 | 18 | 12 | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 | 1 |
| Oct 23, 2026 | 19 | 12 | 10 | 4 | 4 | 6 | 1 | 2 | 2 | 1 | 1 |
TSK-1 over time
Every lab, tested on the same request
Days each lab was tested in TSK-1, our app-building test, Jul 30 to Oct 23, 2026. See every result →
Then Claude, DeepSeek, Kimi and Grok at 75.
Top score
The GPT family leads
GPT family: 88 out of 100 on our app-building test, across every version we tested. See OpenAI models →
Newest
Just released
The newest models in Taskade, by the date the lab released them.
- OpenAI8814 of 16 points29 builds side by side
- Anthropic7512 of 16 points20 builds side by side
- DeepSeek7512 of 16 points17 builds side by side
- Moonshot7512 of 16 points1 build side by side
- xAI7512 of 16 points17 builds side by side
- Zhipu6310 of 16 points2 builds side by side
- Alibaba508 of 16 pointsNo side-by-side results published yet
- Google447 of 16 pointsNo side-by-side results published yet
- MiniMaxNot scored
- NVIDIAScore pending
- MistralNot scored
Every lab
Scores by lab
Out of 100 on our app-building test. Open a lab to see its models.
All models at a glance
Each row opens the model page. A model that left the model menu keeps its test results.
| Model | Result | |||
|---|---|---|---|---|
| GPT-5.6 LunaAvailableReleased 1.05M context | Available | 1.05M | Works 3/3From 12 builds | |
| GPT-5.6 SolAvailableReleased 1.05M context | Available | 1.05M | Works 2/2From 2 builds | |
| GPT-5.6 TerraAvailableReleased 1.05M context | Available | 1.05M | Partial 1/2From 5 builds | |
| GPT-6 LunaAvailableReleased 1.05M context | Available | 1.05M | Tested | |
| GPT-6.1 SolAvailableReleased 1.05M context | Available | 1.05M | Works 1/1From 3 builds | |
| GPT OSS 120BOwn keyReleased 131K context | Own key | 131K | Preview | |
| GPT OSS 20BOwn keyReleased 131K context | Own key | 131K | Preview | |
| Claude Haiku 4.5Own keyReleased 200K context | Own key | 200K | GapsFrom 2 builds | |
| Claude Opus 5Own keyReleased 1M context | Own key | 1M | Works 1/1From 1 build | |
| Claude Opus 5.5Own keyReleased 1M context | Own key | 1M | Works 1/1From 1 build | |
| Claude Sonnet 5Own keyReleased 1M context | Own key | 1M | Partial 2/3From 8 builds | |
| Claude Sonnet 5.5Own keyReleased 1M context | Own key | 1M | Not tested | |
| Claude Sonnet 4.5RetiredReleased 200K context | Retired | 200K | GapsFrom 2 builds | |
| Claude Opus 4.8Not in TaskadeReleased 1M context | Not in Taskade | 1M | Preview | |
| Claude Opus 4.6Not in TaskadeReleased 1M context | Not in Taskade | 1M | Preview | |
| Claude Sonnet 4.6Not in TaskadeReleased 1M context | Not in Taskade | 1M | Preview | |
| DeepSeek V4 ProAvailableReleased 1M context | Available | 1M | Works 2/2From 2 builds | |
| DeepSeek V4.1 FlashAvailableReleased 1M context | Available | 1M | Partial 1/2From 2 builds | |
| DeepSeek V4 FlashReplacedReleased 1M context | Replaced | 1M | Tested | |
| DeepSeek V3.2Not in TaskadeReleased 128K context | Not in Taskade | 128K | GapsFrom 2 builds | |
| Kimi K2.6AvailableReleased 262K context | Available | 262K | Tested | |
| Kimi K2.7 CodeAvailableReleased 262K context | Available | 262K | Tested | |
| Kimi K3AvailableReleased 1.05M context | Available | 1.05M | No appFrom 1 build | |
| Grok 4.6AvailableReleased 500K context | Available | 500K | Works 1/1From 1 build | |
| Grok 4.7AvailableReleased 500K context | Available | 500K | Works 3/3From 3 builds | |
| Grok Build 0.1AvailableReleased 256K context | Available | 256K | Partial 2/3From 3 builds | |
| Grok 4.5RetiredReleased 500K context | Retired | 500K | Works 2/2From 2 builds | |
| Grok 4.3RetiredReleased 1M context | Retired | 1M | GapsFrom 2 builds | |
| Grok 4.20Not in TaskadeReleased 1M context | Not in Taskade | 1M | GapsFrom 4 builds | |
| GLM 5.2AvailableReleased 1M context | Available | 1M | Tested | |
| GLM 5.3AvailableReleased 1M context | Available | 1M | Tested | |
| GLM 5Not in TaskadeReleased 205K context | Not in Taskade | 205K | Works 1/1From 1 build | |
| GLM 4.7Not in TaskadeReleased 205K context | Not in Taskade | 205K | GapsFrom 1 build | |
| GLM 5.3 FlashNot in TaskadeReleased 1M context | Not in Taskade | 1M | Preview | |
| Qwen 3.7 MaxAvailableReleased 1M context | Available | 1M | Tested | |
| Qwen 3.7 PlusAvailableReleased 1M context | Available | 1M | Tested | |
| Qwen 3.8 2.4TNot in TaskadeReleased 262K context | Not in Taskade | 262K | Preview | |
| Qwen 3.8 27BNot in TaskadeReleased 262K context | Not in Taskade | 262K | Preview | |
| Gemini 3.6 FlashOwn keyReleased 1.05M context | Own key | 1.05M | Tested | |
| Gemini 3.1 ProOwn keyReleased 1.05M context | Own key | 1.05M | Not tested | |
| Gemma 4 26B A4BNot in TaskadeReleased 262K context | Not in Taskade | 262K | Preview | |
| MiniMax M3AvailableReleased 1M context | Available | 1M | Tested | |
| Nemotron 3 SuperRetiredReleased 262K context | Retired | 262K | Tested | |
| Mistral Large 4Not in TaskadeReleased 1.05M context | Not in Taskade | 1.05M | GapsFrom 1 build |
Context is how much a model can read at once, counted in tokens. A token is about three quarters of a word.
Published specs from the model maker, via models.dev (MIT).
Bring your own key
Pick a model with your own API key in a custom agent. Each provider page lists the models its key can pick.
FAQ
Which AI models can I use in Taskade?
Taskade provides frontier AI models from top AI labs, including OpenAI GPT and open-weight models such as DeepSeek and Qwen. Users on paid plans can select their preferred model for AI agents and app generation. On Pro and up, you can add Anthropic Claude with your own API key.
Which AI model is best for building apps in Taskade?
In our TSK-1 Build test, GPT-5.6 Luna and Grok 4.7 each work on 3 of 3 requests, the best result among the models you can pick now. TSK-Auto picks a model for each request, so you can leave it on. The TSK-1 Benchmark page shows every result.
Why does this page list models I cannot pick?
We keep a page for every model Taskade served or tested. A model that left the model menu keeps its test results, and a model we tested but never offered says so on its page, so you can compare it with the models you can pick.
What is the difference between TSK-Auto and a TSK speed?
TSK-Auto is the default. It picks a model for each request. A TSK speed sets how fast or how careful the work is, and Taskade picks the model for that speed. Custom Engine keeps one exact model.
Can I run a model on my own API key?
Yes. On Pro and up, you can bring a key from Google Vertex AI, OpenAI, Anthropic or Google Gemini API. On Business and up, you can bring a key from OpenRouter, Fireworks, xAI, Groq, Mistral, Together AI, DeepSeek, Z.ai, Moonshot AI, Alibaba Qwen, Cerebras, Azure OpenAI or Cloudflare Workers AI. A key serves custom agents, Ask Agent, Taskade EVE and automation AI steps on the models its provider page lists.
Does Taskade charge for a model on my own key?
Your provider bills the model use to your own account. Agent steps on your key are free up to a monthly number that your plan includes. Above it, each agent step uses a small Taskade step fee, and a daily cap applies. Each Taskade EVE step and each automation AI step on your key also uses the step fee.
How does Taskade test models?
Every model builds an app in Taskade Genesis from the same request, word for word. We open each app, enter data the way a customer would, make sure that the app keeps the data, and ask for a follow-up change. The TSK-1 Benchmark page shows every result.