Cognition SWE-2 posts 92.8 on the Terminal-Bench 2.1 benchmark, 27.3 on Terminal-Bench 4

SWE-2 is post-trained from Kimi K3, a 2.8T-parameter mixture-of-experts model with 104B active parameters per token that had already undergone extensive RL for agentic coding. Cognition says its own RL pass adds 5 to 6 points on many benchmarks and shifts K3's entire cost-performance frontier. This is the first time Cognition has scaled RL into the multi-trillion-parameter regime, building on the SWE-1.7 training infrastructure and recipe.

The framing Cognition chose is cost, not raw score: 50.0% on FrontierCode 1.1 Main is within one point of Fable 5.1 while being 64% cheaper, and within 3.3 points of GPT-6 Astra at a quarter of the cost. Those are ratios against competitors' costs, not a rate card. Cognition publishes no per-token API for SWE-2, so no one outside the company can reconstruct a dollar figure per task from a published price.

The usable efficiency claim is a trajectory claim. Mean steps per run on FrontierCode 1.1 Main drop from 127 for SWE-1.7 to 53 for SWE-2 medium, 80 for high and 98 for max, measured over all 100 tasks with three runs per task per model. SWE-2 medium makes its first real edit after a median of 18 steps, against 48 for SWE-1.7. Those steps are grouped by the tools each one calls: explore (read / grep / ls), plan / todo, write / edit code, build (make / lint), run tests, git add / commit, final message.

Cognition attributes the largest efficiency gains to focused exploration: the model judges which parts of the codebase actually matter for a task and begins implementation sooner. The company had previously described SWE-1.7 as exceedingly careful, and reports user feedback that it over-explored and overthought simple tasks. For anyone budgeting an agent loop, the step count is the concrete planning number here, since the cost claim is expressed only as a percentage against SWE-1.7 and competitor models.

Cognition also reports behavioral differences between effort levels: medium steps into action quicker for simple and intermediate tasks, while high and max plan more, explore more of the codebase and manage uncertainty through more complex verification on harder tasks. From internal testing it claims better end-to-end test writing, more willingness to find another route when the obvious path is blocked, and re-deriving conclusions when challenged instead of re-asserting them. In one cited case an MCP integration the model needed was unavailable, so it reconstructed the data from Slack channel history it already had access to. The post reports no measurement for any of these behaviors.

The 27.3% on Terminal-Bench 4 is the soft spot, and it is a wide one: Fable 5.1 scores 55.8% and GPT-6 Astra 57.9% on the same suite. Long-horizon agentic work is where the distance to the frontier still sits. Every number in the table is Cognition's own, pending independent replication, and FrontierCode is a Cognition-built benchmark.

Commenters on Hacker News went straight at that delta. One read the spread between the two terminal suites as a generalization test, putting it crudely as how benchmaxxed the model is, and said Terminal-Bench 4 had been released a couple of weeks earlier; another asked why other benchmarks were not shown. One commenter pushed back that the comparison assumes equal difficulty: Terminal-Bench 2.1 is saturated and Terminal-Bench 4 is not, and by that commenter's figures Sol at xhigh effort scores 90% on the old suite against 37% on the new, so the same argument would indict Sol.

Not as weights, and not as an API. The weights are proprietary and unpublished, so there is nothing to download and no quantization ladder to wait for. Distribution is Devin Desktop and CLI today, with Devin Web and Fusion rolling out. On Hacker News, one commenter said the Devin requirement alone would stop them trying it; a commenter who says they work at Cognition, though not on SWE-2, replied that SWE-2 is free for such users for the next month and that almost all usage should be supported via the Devin CLI.

That still leaves harness builders with a closed surface. If your scaffold is your own loop over an OpenAI-compatible endpoint, the sources describe no path to point it at SWE-2. Two commenters said they would rather use an open-weight model they can serve themselves than add another closed provider, both naming DeepSeek v4.1 Flash, launched the same day, as the alternative they would reach for.

The reward is R = S − λ_e C, where S is binary rollout success, C mixes inference cost in USD with rollout time, and λ_e is tuned per effort level to the local slope of the base model's Pareto curve at that point. Cognition argues linearity is forced: only a linear penalty gives the same result whether applied before or after averaging cost. Set λ too high for the high-effort level and the model is rewarded for an unhelpful update where high-effort starts behaving like medium, raising reward without improving the frontier. Cognition contrasts this with Kimi K3's approach of training a separate expert for each combination of domain and effort level, then consolidating the experts through multi-teacher on-policy distillation.

Two supporting pieces matter to anyone doing RL on agent traces. The length-weighted group baseline, in use since SWE-1.6, weights each rollout's reward by its length instead of taking the plain group mean; Cognition reports it reduces gradient variance at no extra cost, stabilizes training and keeps the inference-training KL low. On serving, the stack uses NVFP4 and FP8 kernels with quantization-aware training, with FP8 carrying K, Q, V and score computations in the MLA layers; a draft model retrained with SpecForge gives 15% longer accept lengths, and a prefill delayer lifts TPM per GPU and tokens per second per request by 10 to 20%, with time-to-first-token taking the hit. Cognition reports lower train-inference mismatch than SWE-1.7 at similar throughput despite a base model with almost 3x the parameters. On data, it tripled the number of RL environments, added instruction-following overlays and built a flywheel powered by earlier SWE-2 checkpoints that iteratively hardens its verifiers.

Three things are unresolved. There is no independent replication of any score in the table, and the headline FrontierCode suite is Cognition's own. There is no published price per million tokens, so the 64% and quarter-of-Astra claims cannot be checked against a bill; the free month on the Devin CLI is the only way to sample real cost, and the sources do not say what the terms are after it. The 27.3% on Terminal-Bench 4 is the reported ceiling on long-horizon work right now, whatever the older suite says, and for a harness the number worth planning against is 53 mean steps at medium effort rather than any of the percentages.