Glonce

Cognition ships SWE-2, nearing GPT-6 Astra at a quarter of the cost

modelsThe Glonce editorsHacker News🔥82Glonce rating82Impact75Novelty65Relevance100Surprise55Players60Credibility80
Cognition ships SWE-2, nearing GPT-6 Astra at a quarter of the cost

Cognition has released SWE-2, which it calls its most advanced coding model yet and describes as pushing the Pareto frontier of capability and cost. On the FrontierCode 1.1 Main benchmark, SWE-2 scores 50.0%, within one point of Fable 5.1 while costing 64% less to run. Across FrontierCode 1.1 Main and a second benchmark, DeepSWE 1.1, Cognition says SWE-2 beats its own predecessor SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of the cost. SWE-2 is available immediately in Devin Desktop and CLI, Cognition's coding-agent products, and is rolling out to Devin Web and Fusion as well.

SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter model that had already undergone extensive reinforcement learning for agentic coding, building on the training infrastructure and recipe from Cognition's earlier SWE-1.7. Cognition says this is the first time it has scaled RL to a multi-trillion-parameter regime, and that its post-training still adds 5 to 6 points on many benchmarks on top of what Kimi K3 already had, shifting K3's entire cost-performance frontier rather than moving a single point on it. The model ships in three reasoning-effort levels, medium, high and max: medium acts quickly for cost-efficient performance on 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.

The technical center of the post is a new RL method for training all three effort levels at once. Cognition contrasts this with how Kimi K3, the base model SWE-2 is built on, was itself post-trained: a separate expert for each domain-and-effort-level combination, consolidated by multi-teacher on-policy distillation with a problem-specific and training-step-specific token budget. Cognition instead uses one reward function across a single RL run, rewarding successful rollouts minus a cost penalty, and tunes that penalty separately for each effort level to match the local slope of that level's own cost-performance curve. Cognition argues, with a geometric argument about the frontier's iso-reward lines, that this specific choice is what guarantees training pushes the whole frontier upward rather than trading cost for score along the curve that already exists; set the penalty too high, its diagram shows, and a high-effort model gets rewarded for drifting toward medium-effort behavior, since cost falls faster than the solve rate does, without the frontier itself improving.

The behavioral upshot, per Cognition, is a model that spends less time exploring before it acts. SWE-1.7 had drawn user feedback for tending to over-explore and overthink simple tasks even as it boosted performance; Cognition says SWE-2's biggest efficiency gains come from focused exploration, since higher intelligence lets the model judge which parts of the codebase actually matter for a given task and start implementing sooner. On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average, and it reaches its first real edit after a median of 18 steps, versus 48 for SWE-1.7. In internal testing, Cognition also reports SWE-2 writing more reliable end-to-end tests that catch regressions and edge cases; being more resourceful when a direct path is blocked (in one case, reconstructing needed data from Slack channel history after a required MCP integration turned out to be unavailable); and showing more verification discipline, re-deriving conclusions rather than re-asserting them, checking a user's hypotheses instead of just agreeing, and running artifacts for evidence rather than trusting prose at face value.

The post also credits infrastructure work behind the training run. A length-weighted reward baseline, used since SWE-1.6, significantly stabilizes training, Cognition says; because the theoretically optimal baseline turned out to correlate strongly with rollout length, Cognition approximates it with length alone, a proxy that costs nothing extra to compute and that, in its ablations, proved more stable and better-performing, in part by keeping the inference-training KL divergence low. On the serving side, a 'prefill delayer' that batches nearby requests in the GPU scheduler improved both tokens-per-minute per GPU and tokens-per-second per request by 10 to 20%, at the cost of a longer time to first token; a new speculative-decoding draft model, trained with SpecForge, achieves 15% longer accept lengths and is kept in sync with the changing policy through online training; and NVFP4/FP8 quantization-aware training, plus simplifying the MLA attention layers to FP8 throughout (SWE-1.7 had mixed FP8 and BF16 precision there), gives SWE-2 lower inference-training KL divergence than SWE-1.7 at similar throughput, despite Kimi K3 having almost three times the parameter count of SWE-1.7's own base model. Cognition also says it tripled its RL environments, added instruction-following overlays, and built a flywheel that feeds SWE-2's own earlier checkpoints back in to keep hardening its verifiers.

Key facts

  • SWE-2 scores 50.0% on the FrontierCode 1.1 Main benchmark, within one point of Fable 5.1, while costing 64% less to run.
  • Across FrontierCode 1.1 Main and DeepSWE 1.1, SWE-2 beats SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of its cost.
  • SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter model that had already undergone extensive RL for agentic coding; Cognition's own RL still adds 5 to 6 points on many benchmarks on top of that.
  • The core training advance is an RL method that trains SWE-2's medium, high and max effort levels together in a single run, with a linear cost penalty tuned to each level's own Pareto-frontier slope.
  • On FrontierCode 1.1 Main, SWE-2 medium takes 58% fewer turns and costs 81% less than SWE-1.7 on average, reaching its first real edit after a median of 18 steps versus 48; it is live now in Devin Desktop and CLI, and rolling out to Devin Web and Fusion.

Why it matters

Coding-agent vendors have mostly competed by chasing a single benchmark score at whatever cost that takes. SWE-2's stated goal is different: Cognition says its new RL method optimizes the entire cost-performance curve at once, training the medium, high and max effort levels together in a single run instead of treating each as a separate optimization problem. That is why SWE-2 gets pitched several ways depending on the rival: it beats SWE-1.7 and Grok 4.6 outright on both score and cost, matches Fable 5/5.1 and GPT-5.6 Sol at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of the cost. If that cost-performance framing holds up under independent use, it changes what 'competitive' means for a coding model: not the top score, but the best point on the curve for a given budget.

Who it affects

Cognition's own customers feel this first: anyone running Devin Desktop, CLI, Web or Fusion gets SWE-2 as an upgrade to the coding agent they already use, with three effort levels to match cost against task difficulty. It also raises the bar for the other coding models the post names as comparison points, Grok, the GPT line and Fable among them, since Cognition is explicitly pricing SWE-2 against their scores. Teams that evaluate coding agents on cost per solved task, not just raw benchmark score, are the intended audience for the comparison Cognition is making.

How to use it

SWE-2 is live now in Devin Desktop and CLI, with Devin Web and Fusion rolling out. It ships as three effort levels, medium, high and max, so a task's difficulty rather than one fixed setting determines the cost paid: medium is built to act quickly on simple and intermediate work, while high and max spend more on planning, codebase exploration and verification for harder problems. The post gives no dollar price for SWE-2 or for any individual effort level; every cost figure it states is relative to other models, so actual pricing needs to be checked separately, wherever Cognition publishes it.

How solid is it

Every figure here is Cognition's own report of SWE-2's performance on benchmarks it names (FrontierCode 1.1 Main, DeepSWE 1.1), and none of it is verified independently in this text. The company name, Cognition, comes from the article's listing and URL rather than the body text, which never names the company directly, though it does name its product, Devin, and otherwise refers to itself only as 'we' or 'our'. The retrieved text is also cut short, ending mid-sentence in its 'Data Improvements' section, and the mathematical notation behind the cost-penalty derivation is missing its variable symbols throughout, so the formal proofs Cognition points to (Appendix B for the cost penalty, Appendix C for the reward baseline) cannot be checked from what is available here, though the plain-language geometric argument for the cost penalty is intact and internally consistent.

Risks and caveats

Cognition frames the causal link between its Pareto-informed cost penalty and SWE-2's behavioral gains as a belief ('we believe'), not a proven mechanism. The resourcefulness example it gives, reconstructing missing data from Slack history after an MCP integration turned out to be unavailable, is one internal anecdote, not a measured rate. And because every cost claim in the post is relative (a percentage, 'a fraction', 'a quarter') rather than an absolute price, a reader cannot tell from this post alone whether SWE-2 is cheap in absolute terms or only cheaper than expensive rivals.

“When challenged, SWE-2 re-derives conclusions rather than re-asserting. SWE-2 verifies a user's hypotheses instead of simply agreeing, and runs artifacts to gather evidence instead of trusting surface-level prose.”

— Cognition's SWE-2 announcement