Cognition's FrontierCode 2.0 grades coding agents on whether maintainers would merge

Cognition released FrontierCode 2.0, a harder mergeability-graded coding benchmark where Claude Opus 5.5 leads at 45.5%.

Cognition, the company behind Devin, released FrontierCode 2.0 on October 27, 2026. The coding benchmark grades agent output on whether open-source maintainers would actually merge it, using deliberately informal prompts. If you're choosing a model for an agent that works from loose issue descriptions, this is closer to your real question than benchmarks built on detailed specs.

Cognition says tasks come from 55% more repositories than 1.0 and now include security patches and performance work. Blocking criteria rose 43%, and thoroughness checks, such as fixing a bug in a sibling code path as well as the one reported, roughly tripled. In Cognition's table, Claude Opus 5.5 leads at 45.5%, GPT-6 Astra scores 42.8%, and Grok 4.7 and Gemini 4 Argon fall to 28.6% and 26.3%. On cost, Cognition's X post puts GPT-6.1 Sol at 39% for $1.03 per task, against $12.8 for Opus 5.5.

A practical finding: higher reasoning effort can lower scores. Cognition reports that models write more code and touch more files at max effort, so it's worth testing effort settings on your own repos rather than defaulting to max.

The caveats are real. The tasks are private, and the benchmark comes from a vendor that sells a coding agent. The closest thing to outside checking is Artificial Analysis, which runs the 132 tasks three times each in its own harness. It also has Opus 5.5 on top at 46%, with GPT-6 Astra, GPT-6.1 Terra and Claude Sonnet 5.5 at 42%. explainx.ai notes that 2.0 scores aren't comparable with 1.1 scores, and that it could not verify an "Obsidian 3" result circulating on X.

Sources