HealBench: leaderboard

Repository-level runtime error healing: 265 real crashes from 18 Python repositories, where a coding agent inspects the live state of the crashed program and submits healing code so execution can continue. Scored by correct rate, the share of crashes whose continued execution passes the target test, under the mini-SWE-agent harness; higher is better (HealBench, 2026).

Metric: Correct rate (%): share of 265 crashed executions whose continued execution passes the designated target test after the agent's healing, mini-SWE-agent harness without HealGuard, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 3 models tracked. Updated 2 Oct 2026.

Top models

#ModelScoreOverall rank
1GLM-5.225.28#152
2GPT-5.6 Terra24.53#68
3DeepSeek V4 Flash16.23#265

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=healbench · How It Works · Data refreshed daily, snapshot 2026-11-17.