Gemini 4 Carbon hits 40.0% on Devin’s coding test, up from 26.3%

Cognition, the company behind Devin and Windsurf, says Gemini 4 Carbon is now live in Devin Desktop and Devin CLI. On its own FrontierCode 2.0 benchmark the model scores 40.0%, against 26.3% for Gemini 4 Argon.

A 14-point jump over Argon, at lower cost

FrontierCode 2.0 is Cognition’s benchmark for real-world engineering tasks. It grades models on quality and on whether the result is mergeable. According to the announcement, the score is a weighted aggregate of rubric items, and any solution that fails a blocking criterion gets 0.

Cognition puts Carbon ahead of GPT-6.1 Sol, which it lists at 39.0%. The gap to that model is one point, so the two are close. The gap to Gemini 4 Argon is 13.7 points.

The company also makes a cost claim. At low reasoning effort, Carbon already beats the best result Argon reached at any setting. Carbon’s top score costs less than two-thirds as much per task as Argon’s. Cognition defines cost as the mean spend per rollout at list API pricing, including public discounts. The post shows the figures in a score-versus-cost chart, and the text does not give dollar amounts.

Access is through Devin’s own tools: Devin Desktop and Devin CLI, both linked from the announcement. The post does not mention plan restrictions, regional limits or a price for using Carbon inside Devin, so those remain unclear.

Where 40.0% sits against the rest of the field

The score looks different next to Cognition’s earlier results. Two days ago, Cognition’s leaderboard write-up put Claude Fable 5.2 on top at 49.1%. Claude Opus 5.5 was listed at 45.5% and GPT-6 Astra at 42.8%, according to our report. By those figures Carbon, at 40.0%, is a strong second-tier result and not a new leader. The announcement itself does not compare it with Anthropic’s models.

The Argon comparison is the cleaner one. Argon reached general availability on October 13, and its 26.3% put it in the lower half of that earlier list, below Grok 4.7 at 28.6%. Carbon’s 40.0% would have placed it behind only the top three entries on that list.

The pricing context matters too. Google lists Carbon in the Gemini API at $2.00 per million input tokens and $10.00 per million output tokens, as we reported when it opened to all developers. Those introductory rates run through November 30, 2026, and the changelog does not state the price afterward. Our earlier report listed Fable 5.2 at $8 per million input tokens and $40 per million output tokens. Cognition’s per-task cost figures depend on token use as well as list prices, so rates alone do not settle which model is cheaper in practice.

One more detail from the Carbon coverage: Google described the model at its October 28 unveiling as strong at cybersecurity tasks, and it first reached a limited group of defenders. Cognition’s post says nothing about security and tests only engineering quality.

What the benchmark claims do not establish

Several caveats apply. Cognition builds and runs FrontierCode 2.0 and sells a coding agent, so the figures are the company’s own and have not been independently reproduced. The announcement does not say how many tasks the benchmark contains or how many runs each score averages. It also does not say whether the scores reflect the bare model or Devin’s scaffolding around it, and it does not give the reasoning-effort settings behind each number.

The post also lists GPT-6.1 Sol at 39.0% without saying at which reasoning effort or cost that result was reached, so the one-point lead over it should be read cautiously. A margin that small could easily change with a different run or configuration.

For developers already working in Devin, the practical change is a new model option that Cognition says scores well above Argon and costs less to run per task. Whether it beats the alternatives on a specific codebase is not something a single aggregate score can show. Cognition points readers to its FrontierCode leaderboard for the full score and cost charts.