Cognition's FrontierCode 2.0 puts Claude Opus 5.5 on top at 45.5%
Claude Opus 5.5 leads Cognition's new coding benchmark, but models failed 19.2% of thoroughness items, more than core feature tasks.
Cognition has published the results of FrontierCode 2.0, its new coding benchmark, with Anthropic's Claude Opus 5.5 taking first place on the official leaderboard at 45.5%, according to aitimes.com.
Behind it, OpenAI's GPT-6 Astra scored 42.8%, Claude Sonnet 5.5 reached 41.0%, Gemini 4 Carbon recorded 40.0% and GPT-6.1 Sol finished at 39.0%. The benchmark comprises 132 tasks drawn from real open-source projects, including security vulnerability fixes and performance optimization work.
The results show models struggling more with thoroughness than with building features. The failure rate on essential thoroughness items was 19.2%, against 14.9% on core feature implementation, and even at the maximum reasoning setting the thoroughness failure rate stood at 15.3%.
Individual tasks produced sharp divergences. Claude Opus 5.5 failed all 25 attempts on the Mock Service Worker bug-fix task, across five reasoning settings run five times each. GPT-6 Astra succeeded 5 out of 5 times on the PostHog task at low reasoning level but only 1 out of 5 at maximum reasoning, where it removed a public program interface and broke compatibility.
An earlier version
Cognition, the company behind the Devin coding agent, released FrontierCode 2.0 on its official blog on October 27, 2026. In the previous version, FrontierCode 1.1, Claude Opus 5.5 scored 54.6% and GPT-6 Astra 53.3%.
Quick answers
What is FrontierCode 2.0?
A coding benchmark from Cognition made up of 132 tasks drawn from real open-source projects, including security vulnerability fixes and performance optimization.
Which model tops the FrontierCode 2.0 leaderboard?
Anthropic's Claude Opus 5.5 ranks first at 45.5%, ahead of GPT-6 Astra at 42.8%.
How did Claude Opus 5.5 score on FrontierCode 1.1?
It scored 54.6% in the earlier version of the benchmark, with GPT-6 Astra at 53.3%.