The Combination Is What Matters, Not the Score
SWE-2 does not train its own base. It takes Kimi K3 — Moonshot's July release, a 2.8-trillion-parameter sparse MoE with a 1-million-token context window, the largest open-weight model shipped at the time, already put through heavy RL for agentic coding — and continues with its own reinforcement learning on top. This site covered Kimi K3's open-weight release; this is the first notable downstream commercial product on that line. Cognition's reported increment is that its own RL "still finds substantial headroom," adding 5–6 points across several benchmarks. That magnitude is not dazzling in itself, but the question it answers matters: once a strong open-weight base is out, how much room is left in post-training? Enough to carry a product, it turns out. For teams with neither the ability nor the reason to train a base model, the reproducibility of this route is worth more study than the benchmark numbers — provided you have the RL engineering capability and a usable evaluation environment, neither of which is cheap.
Training Effort Levels Together With a Cost Penalty Is the Interesting Part
The more technical highlight is effort-level training. The usual approach finishes the model, then bolts on a router that decides whether a given task deserves more thinking. Cognition instead trains every effort level inside one RL run, attaching a linear cost penalty per level, with each penalty's magnitude set by the local slope of the base model's Pareto frontier at that point. In other words, "should I spend more compute here" becomes a decision internal to the model and already priced against cost, rather than a guess made by an external scheduler. The downstream effect is concrete: SWE-2 medium scores higher than the previous generation while using 58% fewer turns and costing 81% less on average, with the median step before its first real edit dropping from 48 to 18. Cognition attributes this to "focused exploration" — the prior generation's problem being precisely over-exploration, with user feedback that it overthought simple tasks.
Three Things to Settle Before Choosing It
First, 27.3% on Terminal-Bench 4 against GPT-6 Astra's 57.9%. That gap is in Cognition's own materials, and it says this model is not yet in the same class on longer tasks that more closely resemble real terminal work. Second, what "64% cheaper" refers to. Cognition published no per-token API price, no context window and no model card. So that 64% is a usage figure inside Devin, not a number you can divide against another vendor's per-token rate card. Third, every comparison figure is Cognition's own run, with no independent reproduction so far. That does not make the numbers untrustworthy; it means that until a third party runs them, they remain vendor-reported. Access, at least, has no barrier: it is in Devin Desktop and the CLI, free for Pro/Max/Teams subscribers for one month through October 10, with Devin Pro listed at $20 a month. Anyone who wants to verify it has a month to run their own real tasks through it — which beats reading any table.
via: Cognition, "Introducing SWE-2", Kimi K3 tech blog, MindStudio's benchmark and pricing write-up, CellCog on the 64% cost claim