TSK-Bench · Code arena
Which AI model codes the change you ask for, without breaking the app?
After the first build, we send one follow-up coding request in the same chat. The change must land without breaking what already worked. Each page ranks one kind of change.
Each result covers the first build and the follow-up change together. Where the test recorded it, the finding says what happened to the change.
In our tests, DeepSeek V4 Pro, DeepSeek V4 Pro 0813, DeepSeek V4 Pro, max thinking, GPT-5.6 Luna, GPT-5.6 Sol, Grok 4.5, and Grok 4.7 each got 2 tested changes working, the most of any model version. Each result comes from one to five builds, so read it as a direction, not a final rank.
Last test · 59 builds
How did each model do on each change?
Open a kind of change to see every model ranked on it, with the app it was made to and the request word for word.
| Change | Models | Works |
|---|---|---|
| Add a page to an app you built Add a new page to an app the model already built. The new page must work, and every page that worked before must still work. Made to: game tracker app. | 14 | 9 |
| Add a field to a form Add one more question to a form the model already built. The new answer must save, and every earlier answer must still save. Made to: intake form. | 23 | 12 |
No side-by-side results published yet: Change text and headings.
Which models did best across these changes?
Each model version on each tested change. Open a result to see the change and every model ranked on it. Each result comes from one to five builds, so read it as a direction.
| Model | New page | New field |
|---|---|---|
| Not tested | ||
| Not tested | ||
| Not tested | ||
| Not tested | ||
| Not tested | ||
| Not tested | ||
| Not tested | ||
| Not tested | ||
| Not tested |
WorksWorks, with gapsNo appNot tested
More arenas coming: Automate (AI steps in workflows), Delegate (agents), Research (web research), Answer (chat), and Read (vision).
The full TSK-Bench ranking →Download the data (JSON)
AI model directoryModel docsPick a model in TaskadeMeet TSK-1Plans and pricing
How we grade these apps: read the full method.
FAQ
Which AI model makes a follow-up change without breaking the app?
In our tests, DeepSeek V4 Pro, DeepSeek V4 Pro 0813, DeepSeek V4 Pro, max thinking, GPT-5.6 Luna, GPT-5.6 Sol, Grok 4.5, and Grok 4.7 each got 2 tested changes working, the most of any model version. Each result comes from one to five builds, so read it as a direction, not a final rank.
How does the Code arena test a model?
After the first build, we send one follow-up coding request in the same chat. The change must land without breaking what already worked. Each page ranks one kind of change. Each result covers the first build and the follow-up change together. Where the test recorded it, the finding says what happened to the change.
Which changes does the Code arena rank?
Add a page to an app you built and add a field to a form. Each has its own page that ranks every model on the same words.
Can I pick the model in Taskade?
Yes. Taskade runs frontier models from top AI labs, and TSK-1 Auto picks one for you by default. You can also pick a model yourself. The AI model directory lists every model, with its status, its plan and how to pick it.