Cognition ships SWE-2, a coding model post-trained from the open Kimi K3 base
Cognition released SWE-2 on September 10, a coding-agent model post-trained from Kimi K3, Moonshot AI’s open 2.8-trillion-parameter base. It approaches Fable 5.1 on coding benchmarks at a claimed 64% lower cost, but collapses on Terminal-Bench 4 — the signal that its gains do not generalize to the hardest test.
September 10, 2026. Cognition publishes SWE-2, the model now driving its Devin agent, with an argument that is not raw power: it reaches 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 (50.9%), at a claimed 64% lower cost. The model is closed — no API, no weights, only Devin Desktop, CLI, Web, and Fusion. Its base, however, is open: Kimi K3, the 2.8-trillion-parameter model from Moonshot AI. Post-training on an open base is becoming the most credible way to build a coding agent at market price.
The recipe: an open base, a reinforcement specialization
SWE-2 is not a from-scratch pretrain. It is Kimi K3, already heavily reinforcement-trained for agentic coding, reworked by Cognition with its own objective. The base is an MoE of 2.8 trillion parameters — 104 billion active per token, 896 experts with 16 activated, 93 layers (69 Kimi Delta Attention, 24 Gated MLA), a MoonViT-V2 vision encoder of 401 million parameters, a 160,000-token vocabulary, and a 1-million-token context window.
Cognition did not invent the base: it rented it at the price of an open model, then spent its budget where differentiation lives — post-training. The result, per the vendor, is a 5-to-6-point gain over Kimi K3 on many benchmarks, shifting the base model’s entire cost-performance frontier.
That architecture has a direct consequence for the ecosystem: Kimi K3 is available as open weights on Hugging Face under the Kimi K3 License. You cannot obtain SWE-2, but you can obtain its base. The gap between the two — what Cognition adds — is exactly the value the company sells.
Reinforcement that penalizes cost
Cognition’s methodological contribution fits in one reward function: R = S − λₑ · C, where S is 0 or 1 depending on rollout success, C is the rollout’s cost (a mix of inference in dollars and rollout time), and λₑ is a penalty coefficient tuned to the local slope of the base model’s Pareto frontier, per effort level.
The novelty is less the formula than what it enables: a single training run covers all three effort levels (medium, high, max), instead of three separate checkpoints. A high λ pushes toward shorter, cheaper rollouts; a low λ lets the model spend more compute chasing correctness. Cognition claims to have thereby “scaled RL to the multi-trillion-parameter regime for the first time.”
The whole thing is backed by engineering detail: a tripling of RL environments, NVFP4/FP8 kernels with quantization-aware training, DSpark speculative decoding with an online draft model, and a verifier hardened by a recursive flywheel — SWE-2 checkpoints are themselves used to stress the verifiers that grade rollouts.
The numbers, and the gap that unsettles
The benchmarks Cognition published show a model that beats its predecessor SWE-1.7 on all four tests and sits near the market frontier.
| Benchmark | SWE-2 | Kimi K3 | Grok 4.6 | Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 44.2% | 48.0% | 50.9% | 53.3% |
| DeepSWE 1.1 | 73.0% | 68.5% | 67.5% | 67.4% | 74.1% |
| Terminal-Bench 2.1 | 92.8% | 88.3% | 88.4% | 91.4% | 89.9% |
| Terminal-Bench 4 | 27.3% | 21.5% | 20.3% | 55.8% | 57.9% |
The Terminal-Bench 4 row breaks the narrative. On this newer, less-saturated revision, SWE-2 tops out at 27.3%, far behind Fable 5.1 (55.8%) and GPT-6 Astra (57.9%). Every model drops on the harder revision, but not by the same amount: the frontier models keep roughly half their score, while SWE-2, Kimi K3, and Grok 4.6 lose two-thirds. That is the classic signature of a gain that reflects fitting to known training distributions rather than real generalization.
The point that annoys engineers most: Terminal-Bench 4 was not in SWE-2’s training loop, unlike FrontierCode and Terminal-Bench 2.1-style tasks. Cognition publishes all four benchmarks side by side — including the one where its model slips — which is to its credit, but the conclusion stands.
The efficiency gains are real, and they show in the steps
What is solid is the efficiency gain over its own predecessor. On FrontierCode 1.1 Main, SWE-2 at medium effort beats SWE-1.7 with 58% fewer turns and 81% lower cost on average. The first real code edit arrives at a median step of 18, versus 48 for SWE-1.7.
The cause is focused exploration: a smarter model judges which parts of the codebase actually matter, and starts implementing sooner. Cognition also documents substantive behaviors — better test coverage, more resourcefulness within the user’s boundaries, and a verification discipline that re-derives conclusions instead of re-asserting.
That is the true nature of SWE-2’s value: not a new capability ceiling, but a better ratio of result to cost, for a given class of tasks.
The efficiency story also explains the three effort levels. SWE-2 exposes medium, high, and max, all trained in the same run, so a routine edit does not pay the price of a long-horizon refactor. Cognition’s headline “64% cheaper” is a comparative claim against Fable 5.1 at list pricing — not an absolute per-token figure — and the most honest reading is that the savings come from spending fewer turns, not from cheaper tokens. Run your own suite before trusting the number.
Closed on the surface, open underneath
SWE-2 illustrates a shift larger than Cognition’s announcement alone. For two years, the “open or closed?” question played out at the level of generalist models. In 2026, it plays out at the level of specialized agents: the base is open, the specialization is closed.
The pattern is now established. Kimi K3 is published as open weights by Moonshot AI; Cognition post-trains it into SWE-2 and sells it closed, with no API and no weights, only inside Devin. The value no longer sits in the pretraining — a 2.8-trillion-parameter pretrain has become a rentable input — but in the post-training recipe: the reward function, the RL environments, the hardened verifier.
For a lab that does not want to spend hundreds of millions on pretraining, that is excellent news. For a customer who wants to audit what runs on their code, it is more mixed: the base is inspectable, the agent is not. The trade-off is no longer “open versus closed,” it is “open base, closed agent” — and that is likely to be the dominant economics of coding agents for years to come.
Verdict
SWE-2 does not beat the frontier models on the hardest benchmark, and it does not have to: its argument is the cost-performance ratio. It holds on the saturated benchmarks, and loses on the one that matters for judging generalization.
If you already use Devin, the evaluation is direct: SWE-2 is a clear win over SWE-1.7 — fewer turns, cheaper, better everywhere. Run your own task suite before believing the 50.0% on FrontierCode.
If you are choosing a coding agent this quarter, do not decide on FrontierCode or Terminal-Bench 2.1 alone: check Terminal-Bench 4 — the only row in the table that still separates models — and your own repositories. And watch the underlying signal: post-training on an open base like Kimi K3 is becoming the standard shortcut to a competitive coding agent, closed on the surface, open underneath.