FrontierCode 1.1 leaderboard tests whether maintainers would merge AI code: Opus 5.5 leads at 52%, MiMo-V2.6-Pro hits 45% at $0.18 per task

Why it matters

【Key points】

On October 12, Artificial Analysis launched a leaderboard for FrontierCode 1.1 (a private benchmark Cognition built with open-source maintainers). Each task asks an agent to change a real open-source repository, and grading goes beyond passing tests: it also checks for scope creep, oversized diffs and unhelpful tests, i.e. whether a maintainer would merge the change. Failing any blocking check, or being flagged for looking up the solution, scores zero. Each of the 100 main tasks is run three times in one harness.

Of 22 launch models, Claude Opus 5.5 (max) leads at 52.2% with a 58% blocker pass rate, followed by GPT-6.1 Sol 49.1%, Claude Fable 5.1 49.0%, GPT-6 Astra 47.0% and Claude Sonnet 5.5 45.8%. The top open-weights model is Xiaomi's MiMo-V2.6-Pro at 45.4%, ahead of DeepSeek V4.1 Flash 44.0%, Kimi K3 43.6% and Qwen3.8 Max 43.2%. No model solved nine of the tasks.

Cost per task spans two orders of magnitude: GPT-6 Luna scores 42% for $0.09, MiMo-V2.6-Pro 45% for $0.18, GPT-6.1 Sol 49% for $0.81, while Opus 5.5 costs $10.15 and Fable 5.1 $12. GPT-6 models emit about 27k to 48k output tokens per task, versus about 157k for Opus 5.5 and 236k for Sonnet 5.5.

【Verifiable】

The leaderboard, launch article and methodology are on the Artificial Analysis site (figures as of October 12). It is a standalone board, not part of its Intelligence Index.

【Our read】

The striking part is how crowded the top is: nine models from six labs sit between 43% and 52%. Chinese open-weights models trail the leader by about 7 points at roughly one-fiftieth of Opus's cost per task; for teams running coding agents at scale, that matters more than the points.

Limits: the tasks are private, so outsiders can't reproduce them, and the launch set omits the just-released Muse Wonder and Gemini 4 Argon, so the board will move soon. Leading models still fail more than 40% of the time, a clear gap between writing code and writing code a maintainer would merge.