Cognition scores Gemini 4 Carbon at 40% on FrontierCode 2.0 for $4.04 a task, near the top, while Gemini 4 Argon gets 26.3% at $6.41

Background

Most coding benchmarks only check whether code passes tests. FrontierCode, built by Cognition with open-source maintainers, asks something closer to real work: would the maintainer merge this PR? Version 2.0, released Oct 27, expanded to 132 tasks, added security patches and performance work, and raised the bar for mergeability; Claude Opus 5.5 led at 45.5%, so the benchmark is far from saturated. After Google launched Gemini 4 Carbon on Oct 28, the open question for developers was how it handles real codebases, and Cognition has now published its own numbers.

How the benchmark works

The FrontierCode score is a weighted aggregate of rubric items, and any solution that fails a blocking criterion scores 0. Cost is the mean spend per rollout at list API pricing, including public discounts. Internet use is restricted: models may read docs and search error messages like an engineer, but runs that consult solution-bearing sources such as the original PR are scored zero. Cognition's post does not say which reasoning effort or harness Carbon used; the leaderboard lets you expand each model by harness and effort.

Benchmarks

According to Cognition's post, Gemini 4 Carbon scores 40% at $4.04 per task, near the top and behind Claude Opus 5.5, GPT-6 Astra and Claude Sonnet 5.5. At the 2.0 launch Cognition reported 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra; the post does not give Sonnet 5.5's exact score. Gemini 4 Argon scores 26.3% at $6.41 per task, ahead of Muse Wonder 1.0 (24.3%) but behind Claude Haiku 5.5 and Grok 4.7 (28.6%). So Google's newer model gains about 13.7 points on this test while costing about $2.37 less per task. These are Cognition's runs, not Google-reported results.

What developers should do

If you are choosing a base model for a coding agent, Carbon is now the strongest FrontierCode option after the top Anthropic and OpenAI models, and it is cheaper per task than Argon, so it is worth comparing against Opus 5.5 and GPT-6 Astra on your own codebase. The FrontierCode leaderboard plots score against cost, which helps find the best model and reasoning effort for a given budget. A 40% score still means most solutions fall short of the merge bar, so human review and tests remain necessary.