Cognition's SWE-2 Lands Within a Point of Fable 5.1 at 64% Lower Cost

Cognition shipped SWE-2 on September 10, scoring 50.0% on its own FrontierCode 1.1 Main benchmark against 50.9% for Anthropic's Fable 5.1 and 53.3% for GPT-6 Astra, at what the company says is 64% lower per-task cost than Fable and roughly a quarter the cost of Astra. The interesting number isn't the headline score; it's the gap between what the model does and where you can run it.

SWE-2 is post-trained from Moonshot's Kimi K3, a 2.8-trillion-parameter mixture-of-experts base, roughly 3x the parameters of SWE-1.7's foundation. Cognition's RL adds 5 to 6 points of headroom on top of K3 across most benchmarks. On DeepSWE 1.1 the model posts 73.0% versus K3's 68.5% and Fable's 67.4%; on Terminal-Bench 2.1 it hits 92.8%, ahead of both Fable (91.4%) and Astra (89.9%).

The training story is where a portable idea lives. Three reasoning-effort tiers (medium, high, max) were trained in a single RL run using a reward of the form R = S − λₑC, where S is a binary success signal, C mixes inference dollars and rollout time, and λₑ is a per-tier cost coefficient. Appendix B of the launch post argues the linear cost penalty is the only functional form that makes the objective depend solely on average cost and solve rate. TechTimes notes the result is portable: any lab training multi-effort models can apply it directly. The length-weighted group baseline (b̂ = Σ(RᵢLᵢ)/ΣLᵢ), first used in SWE-1.6, carries over.

What that buys, mechanically: 53 mean steps per run for SWE-2 medium against 127 for SWE-1.7, an 81% cut in average cost per task and 58% fewer turns on FrontierCode 1.1 Main. Median steps to first real code edit drop from 48 to 18. Cognition calls the behavior "focused exploration." A reward function that docks points for spend the model doesn't need tends to produce agents that stop rummaging.

SWE-2 runs only inside Devin (Desktop and CLI at launch, Web and Fusion rolling out). No weights, no standalone API. Devin Pro is $20/month; unlimited SWE-2 access is bundled into Pro, Max, and Teams through October 10, 2026. Cognition, backed by a16z and Accel, was last valued at $48 billion per OfficeChai, and TechTimes reports it reached roughly $492 million in annualized revenue by May 2026, a 13-fold increase over twelve months. The distribution follows an architectural efficiency wave that DeepSeek's causal encoder-decoder work made unavoidable.

The strategic read is old. When Bloomberg terminals bundled data into a product rather than a feed, the moat became the seat, not the pipe. Cognition is betting the same on Devin.

Sources

• https://cognition.com/blog/swe-2
• https://www.marktechpost.com/2026/09/12/cognition-releases-swe-2-a-kimi-k3-post-trained-coding-model-that-matches-fable-5-1-on-frontiercode-at-64-lower-cost/
• https://officechai.com/ai/cognition-releases-swe-2-says-it-performs-close-to-frontier-at-70-lower-cost/
• https://www.mindstudio.ai/blog/cognition-swe-2-coding-model
• https://www.techtimes.com/articles/327315/20260911/cognition-swe-2-beats-frontier-coding-ai-64-lower-cost-using-single-run-rl-training.htm