TEBench - Breaking Test Identification: leaderboard
Identification F1 (%) for tests that the code change makes fail on TEBench (2026), a project-level test-evolution benchmark of 314 commits from 10 Java projects in which an agent in its default configuration updates the test suite after a production-code change. Higher is better.
Metric: Identification F1 (%) for tests the change breaks: the tests the agent modified, deleted (method level) or added (file level) against the developer's own test changes, on TEBench (314 commits from 10 Java projects): given the project at the commit with its production changes but not its test changes, the agent edits the test suite in an isolated worktree with default agent settings; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 60.7 | #266 |
| 2 | GLM-5 | 60.6 | #269 |
| 3 | Qwen3.5 (TEBench checkpoint unspecified) | 60.5 | |
| 4 | GPT-5.3 Codex | 60.3 | #124 |
| 5 | Claude Sonnet 4.6 | 58.8 | #173 |
| 6 | DeepSeek V3.2 | 58.6 | #419 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=tebench-breaking-test-identification · How It Works · Data refreshed daily, snapshot 2026-11-21.