Cognition Launches SWE-2, a Coding Model Built on Kimi K3 That Undercuts GPT-6 Astra on Price

Cognition released SWE-2 on September 10, 2026, its newest coding model and the successor to SWE-1.7. The model is built on top of Moonshot AI’s 2.8-trillion-parameter Kimi K3 base and is aimed squarely at the cost-performance tradeoff in agentic software engineering, rather than chasing raw benchmark scores alone.

Details

• Base model and training: SWE-2 is trained on Kimi K3, with Cognition describing it as the first time RL training has been scaled to a multi-trillion-parameter regime using a single, unified effort-level training run
• Performance gain over the base: SWE-2 adds roughly 5-6 points over the untuned Kimi K3 base across many benchmarks
• Benchmark results: on FrontierCode 1.1 Main, SWE-2 scores 50.0% versus Kimi K3’s 44.2%, Fable 5.1’s 50.9%, and GPT-6 Astra’s 53.3%; on DeepSWE 1.1 it hits 73.0%; on Terminal-Bench 2.1 it leads the comparison set at 92.8%
• Pricing: Cognition says SWE-2 is 64% cheaper than Fable 5.1 while matching it on FrontierCode 1.1 Main, and costs about one-quarter of GPT-6 Astra’s price
• Efficiency: at medium effort level, SWE-2 uses 58% fewer turns than SWE-1.7 while cutting cost by 81%, thanks to a linear cost-penalty tuned per effort level to the local Pareto curve slope
• Behavior improvements: Cognition highlights better test coverage, more resourcefulness when stuck, and stronger verification discipline before marking work done
• Inference stack: SWE-2 runs on NVFP4/FP8 kernels with quantization-aware training for higher throughput
• Availability: live starting September 10 in Devin Desktop and CLI, with rollout to Devin Web and the newly announced Fusion harness following

What happened next

SWE-2 slots in as Devin’s default cost-efficient model option, sitting alongside frontier models like Fable 5.1 and GPT-6 Astra rather than replacing them. Cognition positions it as the model to reach for when a task doesn’t need top-tier reasoning but still benefits from an agentic coding-tuned model — a role it’s designed to fill inside Fusion, the dual-agent harness Cognition introduced the following day, where SWE-2 is the recommended “sidekick” model pairing with a frontier lead model. Full benchmark methodology and pricing details are on Cognition’s blog at cognition.com/blog/swe-2.