Cognition ships SWE-2 Flash, swyx says 3x faster at same evals
The claim is distillation with no quality cost, and so far nobody outside Cognition has tested it on a real codebase.
Cognition released SWE-2 Flash, which it describes as a fast and efficient coding model. The company says that on leading evals it outperforms all American open-weight models, including ones 3x larger. Rather than publish the weights, Cognition says it is building a program to share them securely with partners and researchers, citing a balance between transparency and cyber risks.
On leading evals, it outperforms all American open-weight models, even ones 3x larger.
The number that got attention came from swyx, who says the model is 3x faster and 5x cheaper while scoring the same on FrontierCode 2. His read is that this is distillation without tradeoffs, meaning a smaller model trained from a larger one that normally gives up some quality in exchange for speed and price.
3x faster 5x cheaper but evals SAME ON FRONTIERCODE 2
What is actually established
Two things, both from the people with an interest in the result. Cognition's eval claim is Cognition's own, and the comparison class is specific, American open-weight models, not frontier closed models in general. swyx's speed and cost numbers are his framing of the release, not an independent run, and he does not describe a codebase he tested it on.
swyx also makes a prediction worth separating from the measurement. He calls Thinking Machines' Inkling one of the most underrated base models in the world and expects to see a lot more models like this one, especially built on US pretrained models. That is a bet about where the next wave of fast coding models comes from, not evidence about SWE-2 Flash.
For working engineers
A same-score-but-cheaper claim is the easiest kind to make and the hardest kind to keep. Eval parity is measured on fixed tasks, and the gap between a benchmark number and a model that holds up across a week of agent runs on your repo is where most speed-versus-quality claims go to die. The useful signal will be someone with no stake in the release pointing the model at a large codebase and reporting what broke.
The weights policy also matters for anyone who wanted to verify the open-weight comparison. Cognition is not publishing them, so the claim that SWE-2 Flash beats models 3x its size cannot be independently reproduced by anyone outside the partner and researcher program. Nobody has posted a third-party result yet.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.