Cognition Unveils FrontierCode 2.0 Coding Benchmark; Claude Opus 5.5 Leads at 46%

On October 28, AI coding company Cognition released FrontierCode 2.0, the latest update to its private coding benchmark. The benchmark comprises 132 tasks sourced from real open-source repositories, scored based on whether a maintainer would merge the agent's proposed change. Any attempt that fails a blocking check receives zero, and the tasks remain private to avoid contamination.

According to results shared by Artificial Analysis, Claude Opus 5.5 (max) leads with 46%, followed by GPT-6 Astra (max), GPT-6.1 Terra (max), and Claude Sonnet 5.5 (max), each at 42%. DeepSeek V4.1 Flash 1017 (max) is the highest-scoring open-weights model at 28%, 18 points behind the leader.

Each task was run three times in Stirrup, Cognition's agentic harness. FrontierCode 2.0 is designed to evaluate AI agents' ability to make code changes in real-world software development settings.