Interaction Bench: leaderboard
Interaction Bench (v0.1 preview) tests whether a coding agent can find and repair a usability defect in a real open-source application, such as a map that will not pan or a form that gives no feedback. Each of the 44 tasks, across 25 applications, comes from a merged pull request; the agent gets the user's goal, the source code and a runnable offline copy of the app, but not the diagnosis or the fix. A separate grading agent replays the interaction on the patched app and checks reviewed behavioural requirements, and a task passes only when every case passes. Unlike SWE-bench, which checks a patch against the project's unit tests, the pass condition is the observed behaviour of the running interface. Each model runs in a harness the maintainers picked for it, at high reasoning effort. This board is the share of the 44 tasks passed.
Metric: Tasks passed, 44 interaction-repair tasks (%). Source: ibench.ai. Saturation forecast: Around January 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (High) | 34.1 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (High) | 27.3 | #5 (Claude Opus 5.5) |
| 3 | GLM-5.3 Flash (High) | 9.1 | #119 (GLM-5.3 Flash) |
| 4 | Qwen 3.8 Max (High) | 6.8 | #55 (Qwen 3.8 Max) |
| 5 | Kimi K3 (High) | 4.5 | #53 (Kimi K3) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #10 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=interaction-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.