Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra

	Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra(cognition.com)
	424 points by seelos 20 hours ago \| 176 comments
	tl;dr: Cognition released SWE-2, a coding model post-trained via RL from Kimi K3 (2.8T params) that hits 50% on FrontierCode 1.1 Main—matching Fable 5.1 at 64% lower cost and approaching GPT-6 Astra at a quarter of the price. Key technical contributions include a principled linear cost penalty tuned to the Pareto frontier's slope (allowing all reasoning-effort levels to be trained in a single RL run), a length-weighted reward baseline that stabilizes training, and improved rollout infrastructure using NVFP4/FP8 kernels with an online-trained draft model for speculative decoding.
	HN Discussion: ↓Benchmark gap between TB2.1 and TB4 suggests benchmaxxing rather than real generalization ↓Cognition has a history of overhyped demos, so claims should be taken skeptically ↓Closed-weight models are unwanted; open alternatives like DeepSeek are more compelling ↓Lived experience with Devin/SWE products has been poor, casting doubt on claims ↑RL post-training from Kimi K3 to reach Fable 5 capabilities is an interesting demonstration