Cognition's new coding model undercuts its rival on price

Cognition has released SWE-2, its new coding model. According to the company's blog post, SWE-2 scores 50.0% on FrontierCode 1.1 Main, landing within one point of Fable 5.1 while being 64% cheaper.

SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter model that had already gone through extensive reinforcement learning (RL) for agentic coding. Cognition says its own RL stage still added 5 to 6 points on many benchmarks. The company also states it scaled RL to the multi-trillion-parameter regime for the first time.

Cost and behaviour

The comparisons Cognition reports: on 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 the cost.

On efficiency, the company reports that SWE-2 medium outscores SWE-1.7 while taking 58% fewer turns and costing 81% less on average. SWE-2 medium makes its first real edit after a median of 18 steps, compared with 48 for SWE-1.7. Cognition notes user feedback that SWE-1.7 over-explored on simple tasks, and says most of the new efficiency comes from focused exploration. The high and max effort levels plan more and run more verification on complex tasks.

All effort levels in one run

The main methodological change is an RL algorithm that trains every reasoning-effort level in a single run. Cognition uses a cost-penalised reward of the form R = S - λ_e C, where S marks whether a rollout succeeded, C is the rollout cost (a mix of inference cost in USD and rollout time), and λ_e is tuned to the slope of the base model's Pareto curve at that effort level. The company argues a linear penalty is the only form that keeps the objective consistent before and after averaging.

SWE-2 is available from today in Devin Desktop and CLI, with rollout underway on Devin Web and Fusion. Full details are in Cognition's announcement.