Gemini 4 Carbon scores 40% on Cognition's FrontierCode 2.0 at $4.04 per task

The Cognition team, which builds the AI engineer Devin, has posted on X the first results for Gemini 4 Carbon and Gemini 4 Argon on its own benchmark, FrontierCode 2.0. According to the authors, the test measures a model's ability to go beyond the spec, that is, to do more than the task explicitly states. Gemini 4 Carbon scored 40% and costs $4.04 per task. Cognition calls this a near-SOTA result, but acknowledges that the model trails Claude Opus 5.5, GPT-6 Astra and one more Claude model.

What we know about Gemini 4 Carbon and Argon

The post says little, so the numbers are worth listing carefully. Gemini 4 Carbon scored 40% on FrontierCode 2.0. The average cost of solving one task was $4.04. Ranked above it are Claude Opus 5.5, GPT-6 Astra and another model from the Claude family. Its name is cut off in the text of the post, so we cannot say exactly where Carbon sits in the table. All we know is that it is behind three models.

For the second model, Gemini 4 Argon, the text of the post gives neither a score nor a cost. Cognition says both models were run through the benchmark, but it provides figures only for Carbon. The full picture is probably in an attached image or table, which we do not have access to. For that reason, Carbon and Argon cannot be compared with each other based on this one post.

Why test a model on going beyond the spec

Most popular coding tests check whether a model can complete a clearly defined task: fix a bug, pass pre-written tests. In real development, a task is almost always described incompletely. A good engineer spots missing edge cases, aligns interfaces with the rest of the code and does not break things the ticket never mentions. According to Cognition, this is the ability FrontierCode 2.0 tries to measure.

For the company, this is a practical question. Devin works on tasks from real repositories, where the spec is rarely exhaustive. A benchmark that distinguishes a model that follows the text to the letter from one that understands the intent helps decide which neural network to build an agent on. It is important to remember here that Cognition itself published the methodology and the task set, not an independent lab. The benchmark is tailored to its own scenarios, so the result should not be carried over to any other stack without testing.

How to read "40% and $4.04" together

The 40% figure alone says little: it depends on how hard the task set is. If FrontierCode 2.0 is deliberately difficult, 40% may be a strong result, and the phrase "near SOTA" hints at exactly that. It is more useful to look at quality together with cost. As a rough estimate, $4.04 per attempt and 40% successful attempts add up to about $10 per solved task. This is our editorial estimate, not a Cognition claim. It assumes the cost is roughly the same across all tasks, whereas in practice failed attempts are often more expensive.

For teams choosing a model for agentic coding, this has several practical implications:

• Gemini 4 Carbon is worth including in a comparison alongside Claude Opus 5.5 and GPT-6 Astra, especially if run cost matters.
• The post gives no figures for Claude Opus 5.5 and GPT-6 Astra, so the gap in points and in dollars cannot be assessed from the text.
• To avoid depending on someone else's benchmark, run the models on 20–30 of your own tasks from a real backlog and count the cost per solved task, not per attempt.

The takeaway is modest but useful. With Gemini 4 Carbon, Google joins the group of competitive models for agentic coding, but Anthropic and OpenAI still lead on this test. Final conclusions can be drawn once the full table is published, including the Gemini 4 Argon result and the cost of the other models.