What Cognition released
Cognition released SWE-2 on 10 September, a coding model that succeeds SWE-1.7 and runs inside its Devin agent. In its launch post, the company said SWE-2 scores 50.0% on FrontierCode 1.1 Main, within one point of Anthropic’s Claude Fable 5.1, at 64% lower cost.
Every benchmark figure in this story comes from Cognition’s own evaluation. No independent result for SWE-2 had been published at the time of writing, and early coverage such as OfficeChai’s report repeats the company’s numbers rather than testing them.
SWE-2 is not trained from scratch. Cognition says it is post-trained from Kimi K3, a 2.8-trillion-parameter model that OfficeChai describes as an open model from Moonshot AI. Cognition applied reinforcement learning at that multi-trillion-parameter scale, and says a single training run produces three effort levels, medium, high and max, instead of separate models for quick and thorough work.
The numbers, as Cognition reports them
On FrontierCode 1.1 Main, SWE-2 scores 50.0% against 42.0% for SWE-1.7. The jump is larger on DeepSWE 1.1, where Cognition reports 73.0% against 37.7%. On Terminal-Bench 2.1 the model reaches 92.8%, up from 81.5% and the highest figure in the company’s comparison table.
Cognition says SWE-2 “beats SWE-1.7 and Grok 4.6 on both score and cost”, and that it matches OpenAI’s GPT-5.6 Sol and Claude Fable 5 and 5.1 for a fraction of their price. It puts SWE-2 at about a quarter of the cost of GPT-6 Astra, which still leads it on FrontierCode, 53.3% to 50.0%, and on DeepSWE, 74.1% to 73.0%.
The same table shows where SWE-2 falls behind. On Terminal-Bench 4 it scores 27.3%, less than half of Claude Fable 5.1’s 55.8% and GPT-6 Astra’s 57.9%. The post does not explain the gap.
Cognition says it used published results where they existed and its own harness for each model otherwise, reported each model’s best score across reasoning-effort settings, and calculated cost from list prices including public discounts. Those choices are common in launch posts, and they are one reason vendor tables rarely match independent runs.
Fewer steps before the first edit
The company’s main argument is efficiency. It says SWE-2 at medium effort makes its first real code edit after a median of 18 steps, against 48 for SWE-1.7, and averages 53 steps per FrontierCode run against 127. Cognition puts the medium setting’s average cost 81% below its predecessor’s.
For an agent, the number of steps is most of the bill. A model that reads less of a repository before acting is cheaper to run; whether it is also more careless is a question these benchmarks do not settle.
Because the base model is Kimi K3, Cognition also published tests of whether SWE-2 repeats propaganda or censors answers. It reports a 98.0% pass rate overall, with 99.8% in English and 95.2% in Simplified Chinese.
Where it runs, and what it costs
SWE-2 is available from launch in Devin Desktop and the Devin CLI, and Cognition says it is rolling out to Devin Web and Devin Fusion. The launch post gives no per-token price, no plan pricing and no separate API, so its cost comparisons cannot yet be checked against a price a customer would pay.
The release follows Cognition’s $2bn funding round at a $48bn valuation, led by Andreessen Horowitz and Accel, announced the day before.
What to watch
The first test is replication. A third-party run on FrontierCode or Terminal-Bench would show whether the gains over SWE-1.7 hold outside Cognition’s own harness, and whether the Terminal-Bench 4 gap is as wide as the company’s table suggests. The second is price: once SWE-2 reaches Devin Web, the claim of 64% lower cost will have a number to be measured against.
