AI Signal 142
Cognition’s SWE-2 model reaches 50.0% FrontierCode score while cutting cost 64%
↗
SWE-2, a post-trained model from Kimi K33, scores 50.0% on FrontierCode 1.1 Main, matches Fable 5.1 and GPT-5.6 Sol, and is 64% cheaper
The model pushes the Pareto frontier of capability and cost, offering performance near the top of the leaderboard at a substantially lower price. For engineering teams, this means they can obtain comparable code-generation quality with reduced compute spend, lowering the barrier to using large language models in daily workflows. By scaling reinforcement learning to the multi-trillion-parameter regime and training all reasoning-effort levels in a single run, SWE-2 demonstrates a new way to advance the cost, performance curve without needing separate models for each effort level.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
SWE-2 achieves 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while being 64% cheaper than that model.
On DeepSWE 1.1 it scores 73.0%, outperforming SWE-1.7 and Grok 4.6 and matching GPT-5.6 Sol and Fable 5.1 at a lower cost.
Behaviorally, SWE-2 medium makes its first real edit after a median of 18 steps, compared with 48 steps for SWE-1.7, and uses 58% fewer turns and 81% less cost on FrontierCode 1.1 Main.
THE READ
What the cluster adds up to.
Cognition has released SWE-2, a post-trained coding model built from the Kimi K33 base. The model is positioned as the company’s most advanced coding model yet, pushing the Pareto frontier of capability and cost. It attains 50.0% on FrontierCode 1.1 Main, which is within one point of Fable 5.1 while being 64% cheaper. On DeepSWE 1.1 it reaches 73.0%, outperforming SWE-1.7 and Grok 4.6 and matching GPT-5.6 Sol and Fable 5.1 at a lower cost. These results place SWE-2 near the top of the public leaderboards for code-generation tasks.
SWE-2 is the first model to scale reinforcement learning to the multi-trillion-parameter regime. It builds on the SWE-1.72 training infrastructure and recipe, adding an RL algorithm that trains all reasoning-effort levels in a single run. The approach applies a linear cost penalty per effort level, tuned to the local slope of the base model’s Pareto frontier. This method is derived from first principles to advance the entire cost, performance frontier while preserving its shape. Reward baselines derived from length-weighted averages stabilize training, and improved rollout serving raises decoding throughput.
On FrontierCode 1.1 Main, SWE-2 medium completes tasks in 58% fewer steps and at 81% lower cost than SWE-1.7. The median number of steps before the first real edit drops from 48 for SWE-1.7 to 18 for SWE-2 medium. Higher intelligence enables focused exploration, allowing the model to begin implementation sooner. Additional behaviors observed include better end-to-end test coverage, greater resourcefulness when the obvious path is blocked, and a tendency to verify hypotheses rather than simply agree. These changes reduce detours and redundant reads, leading to more complete solutions.
Despite its advances, SWE-2 still trails the absolute top scores on some benchmarks. On FrontierCode 1.1 Main, GPT-6 Astra leads with 53.3% while SWE-2 is at 50.0%. On Terminal-Bench 4, SWE-2 scores 27.3%, below Fable 5.1’s 55.8%, GPT-5.6 Sol’s 37.3% and GPT-6 Astra’s 57.9%. The model’s cost advantage remains, but on tasks requiring the highest absolute performance it does not yet match the leading frontier models. Users needing the top-tier scores may still need to rely on larger, more expensive models for those specific workloads.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER