Frontier-Bench v0.1Coding

Anthropic internal evaluation using mini-SWE-agent with a GKE backend. Scores are the mean reward across five attempts per task. Safety-classifier refusals from Opus 5 and Fable 5 fall back to Opus 4.8.

Ranking Sources

#ModelScore

1

Claude Opus 5Anthropic

43.3%

2

GPT-5.6 SolOpenAI

34.4%

3

Claude Fable 5Anthropic

33.7%

4

Claude Opus 4.8Anthropic

21.1%

4 models total

Back to benchmarks