Cognition's SWE-2 Gets Within a Point of the Frontier for 64% Less

Cognition released SWE-2 on Thursday, a coding model the company says lands within one point of Fable 5.1 on its own FrontierCode benchmark while costing 64% less to run. The more interesting detail is where the model came from: SWE-2 is post-trained from Kimi K3, the 2.8-trillion-parameter open-weight model from Moonshot AI, which makes it a frontier-adjacent product built on a Chinese base. The launch post reached the front page of Hacker News with 243 points.

Key takeaways

• SWE-2 scores 50.0% on FrontierCode 1.1 Main against Fable 5.1's 50.9%, at roughly a third of the price of the frontier models it trails.
• Cognition post-trained the model from Moonshot AI's open-weight Kimi K3 and says its reinforcement learning added 5–6 points across many benchmarks.
• SWE-2 medium beats the older SWE-1.7 while using 58% fewer turns and costing 81% less per task.

How SWE-2 ranks against the frontier

On FrontierCode 1.1 Main, Cognition reports 50.0% for SWE-2, ahead of Kimi K3 at 44.2%, Grok 4.6 at 48.0% and GPT-5.6 Sol at 47.5%, and behind only Fable 5.1 at 50.9% and GPT-6 Astra at 53.3%. On DeepSWE 1.1 the model hits 73.0%, roughly a point off Astra. Terminal-Bench 2.1 is the clean win, where SWE-2 tops the table at 92.8%.

The picture is less flattering on Terminal-Bench 4, a harder suite, where SWE-2 manages 27.3% against 57.9% for Astra and 55.8% for Fable 5.1. That gap is the honest read on what a cheaper model still gives up: on the hardest long-horizon agentic tasks, the frontier remains well ahead.

Training every effort level in one run

The technical claim behind the pricing is a change to how Cognition runs reinforcement learning. Instead of training separate low, medium and high reasoning modes, the team trains all effort levels inside a single RL run using a cost-penalized reward — success minus a per-effort-level coefficient multiplied by rollout cost, where that coefficient is tuned to the local slope of the base model's cost-performance curve.

Cognition argues the penalty has to be linear in cost, since only a linear term gives the same answer whether it is applied before or after averaging. Setting the coefficient to match the frontier's tangent, rather than treating it as a hyperparameter to sweep, is what keeps the optimizer from quietly collapsing a high-effort mode into a cheaper medium-effort one.

Serving changes carried the rest. The team batched nearby prefill requests to lift throughput 10–20%, trained a fresh speculative decoding draft model that accepted 15% longer sequences, and moved to NVFP4 and FP8 kernels with quantization-aware training. Despite a base model with nearly three times the parameters of its predecessor, Cognition reports lower train-inference divergence than SWE-1.7.

Fewer detours, earlier edits

Users complained that SWE-1.7 over-explored simple tasks, reading far more of a codebase than a small fix warranted. SWE-2 makes its first real edit after a median of 18 steps on FrontierCode 1.1 Main, compared with 48 for SWE-1.7. Cognition frames this as focused exploration rather than laziness: the model is better at judging which parts of a repository matter.

The company also says SWE-2 re-derives its conclusions when challenged instead of folding, and runs artifacts to gather evidence rather than trusting prose. SWE-2 is live now in Devin Desktop and CLI, with rollout to Devin Web and Fusion underway.

What it means for coding agent pricing

Cognition's bet, consistent with the pitch behind its $2 billion raise at a $48 billion valuation, is that most engineering work does not need the most expensive model in the market. Taking a strong open-weight base and spending the compute on task-specific RL is now a credible route to within a point or two of the frontier, and it is a route that does not require pretraining a multi-trillion-parameter model from scratch.

Cognition also published trustworthiness results, reporting that SWE-2 answered 98.0% of politically sensitive prompts without adopting official Chinese state positions, including 95.2% when asked in Simplified Chinese. That evaluation matters precisely because the base model is Moonshot AI's, and it is the kind of number buyers will now expect from anyone shipping a derivative of an open Chinese model.

FAQ

Is SWE-2 open source?

No. SWE-2 is a proprietary model served through Cognition's Devin products. Its base model, Kimi K3 from Moonshot AI, is open weight, but Cognition has not released the post-trained weights.

How much cheaper is SWE-2 than comparable models?

Cognition reports SWE-2 costs 64% less than Fable 5.1 while scoring within one point of it on FrontierCode 1.1 Main, and roughly a quarter the cost of GPT-6 Astra. Against Cognition's own SWE-1.7, SWE-2 medium costs 81% less per task on the same benchmark.

Where can I use SWE-2 today?

SWE-2 is available immediately in Devin Desktop and the Devin CLI. Cognition says it is also rolling the model out across Devin Web and Fusion.