Cognition releases SWE-2 with lower FrontierCode costs

• SWE-2 reaches 50.0% on FrontierCode 1.1 Main, within 0.9 points of Fable 5.1 at 64% lower cost, according to cognition's FrontierCode claim; cognition's launch announcement frames the broader claim as up to 70% lower cost on leading evals.
• The model is a post-train of Kimi K3, and cognition's model description says its RL recipe still found substantial headroom on an already agentic-coding-trained base.
• Selectable medium, high, and max effort modes come from one RL run, cognition's effort-level description says, while cognition's efficiency data reports medium beat SWE-1.7 with 58% fewer turns and 81% lower average cost on FrontierCode.
• Terminal-Bench 4 remains a conspicuous gap: the comparison table shared by TheRundownAI puts SWE-2 at 27.3%, far behind Fable 5.1's 55.8% and GPT-6 Astra's 57.9%.
• SWE-2 is live in Devin Desktop and CLI, with a month free for Pro, Max, and Teams subscribers per cognition's availability post and a launch summary.

Cognition's technical post publishes the reward function, a median first-edit count of 18 steps versus 48 for SWE-1.7, and the serving stack behind the rollout. The FrontierCode 1.1 revision also relaxed 75 of more than 1,000 grading criteria, while a later desk audit argues that a patch-only evaluator cannot fully answer its stated mergeability question.

• SWE-2 is a Kimi K3 post-train, as cognition's model description states; Cognition's launch post identifies Kimi K3 as a 2.8T-parameter base model.
• Medium, high, and max are the exposed reasoning efforts. cognition's effort-level post says medium targets a cheaper, smarter point, while max is trained to spend more tokens and turns for the highest scores.
• Devin Desktop and CLI have SWE-2 now, according to cognition's availability post; Cognition says Devin Web and Fusion are rolling out separately in its launch post.
• Pro, Max, and Teams subscribers receive a month of free access, cognition's offer says, a detail also repeated in a launch summary.
• The free-use boundary is agent placement. A billing reply says local-agent SWE-2 usage does not deduct billing, while Desktop sessions that route to a cloud agent do incur charges; a follow-up describes Desktop and CLI as free for the month.

First-party

• FrontierCode 1.1 Main, SWE-1.7 → SWE-2: 42.0% → 50.0%, +8.0 points, in TheRundownAI's comparison table.
• DeepSWE 1.1, SWE-1.7 → SWE-2: 37.7% → 73.0%, +35.3 points, in TheRundownAI's comparison table.
• Terminal-Bench 2.1, SWE-1.7 → SWE-2: 81.5% → 92.8%, +11.3 points, in TheRundownAI's comparison table.
• Terminal-Bench 4, SWE-1.7 → SWE-2: 7.6% → 27.3%, +19.7 points, in TheRundownAI's comparison table.

Third-party evaluators

Cognition's launch post supplies the comparison table, rather than an independently run old-to-new evaluation. No third-party evaluator result was linked in the announcement.

Customer-reported

No customer-run score appears in the launch material. cognition's internal-use post says Cognition's own team uses SWE-2 for feature work, debugging, and novel-math visualizations.

Outside the company, imjaredz restated the claim that SWE-2 matches or beats Fable and Astra on many coding benchmarks at 60% less cost. When asked whether that was real, imjaredz's reply confirmed the claim without providing an independent run.

Terminal-Bench 4 is the clear relative weakness. TheRundownAI's comparison table places SWE-2 at 27.3%, 28.5 points below Fable 5.1 and 30.6 below GPT-6 Astra, even though it improves sharply over SWE-1.7 on that test.

Several Hacker News commenters argued that Terminal-Bench 2.1 is close to solved and treated Terminal-Bench 4 as the more informative separator. Cognition's FrontierCode 1.1 update also changed the release's headline eval before these results, refining internet-use rules and relaxing 75 overly strict grading criteria.

Cognition trained effort levels together rather than producing separate models. The training post defines the RL reward as R = S - λₑC, where S is binary rollout success, C combines inference cost in dollars with rollout time, and each effort level gets a cost penalty tuned to its local Pareto-frontier slope.

• Effort behavior: cognition's efficiency data says medium makes its first real edit after a median 18 steps, versus 48 for SWE-1.7. High and max plan, explore, and verify more on complex work, per the training post.
• Gradient baseline: Cognition uses a length-weighted reward baseline, Σ(RᵢLᵢ)/ΣLᵢ, as a no-extra-backward-pass proxy for gradient-norm weighting; it reports lower inference-training KL during RL in the technical write-up.
• Rollout serving: A prefill delayer improved both tokens per minute per GPU and tokens per second per request by 10% to 20%. An online speculative-decoding draft model produced 15% longer accepted sequences, according to the training post.
• Numerics: The serving stack uses NVFP4 and FP8 kernels with quantization-aware training, including FP8 for MLA K, Q, V, and score computation. Cognition reports similar throughput to SWE-1.7 despite a base model nearly three times larger.
• Data and verifiers: Cognition tripled its RL environments, added instruction-following overlays, and used earlier SWE-2 checkpoints to find verifier false positives and negatives, then harden the evaluators.

Cognition also ran a vendor-designed propaganda and censorship evaluation: 145 politically sensitive questions in English, Simplified Chinese, and Traditional Chinese, judged by GPT-5.6 Luna against reference material. SWE-2 passed 98.0% overall in the training post, which defines a pass as a substantive answer that does not adopt the official PRC position.

On a separate ten-task coding evaluation, Cognition varied customer identity, language, and secure-implementation instructions. Its reported result found no statistically significant framing effect on judged vulnerability for SWE-2 or the five comparison models.

Claim: FrontierCode measures whether a maintainer would merge an agent-generated pull request. Cited by: Cognition's FrontierCode leaderboard, which combines unit tests, maintainer-authored rubrics, and verifiers, and zeroes runs that consult solution-bearing sources.

Counter: A GitHub desk audit says a patch alone cannot measure mergeability. Its two-coder review of 98 closed-unmerged pull requests found that 52 of 59 confidently coded maintainer closures were decided for reasons outside the benchmark's six rubric axes.

Evidence so far: The audit credits maintainer-authored rubrics and the benchmark's private data as improvements, but notes its sample came from one AI-assisted contributor population and that harder FrontierCode tasks may make diff correctness more binding. Cognition's revision process has audited more than 1,000 criteria, yet the benchmark's task set and reruns remain nonpublic.

Cognition also introduced Devin Voice, a voice interface powered by GPT-Live and SWE-2, in cognition's Devin Voice post. juberti's explanation says the real-time model can continue a conversation while delegation runs asynchronously and then streams new context back into the exchange.

Devin Voice launch demo