Cognition Launches SWE-2, Claims It Is the Closest Self-Developed Coding Model to the Frontier
Cognition Releases Coding Model SWE-2
Coding agency Cognition has launched SWE-2, claiming it is the closest self-developed coding model to the frontier, performing comparably to recent frontier models on mainstream evaluations, with costs reduced by up to 70%. The official blog provides a detailed comparison: FrontierCode 1.1 scored 50.0%, just 1 percentage point behind Fable 5.1's 50.9%, with a price approximately 64% lower.
The company claims to have achieved reinforcement learning at a trillion-parameter scale for the first time, adding algorithms that can cover all inference intensities in a single training session on the infrastructure of SWE-1.7, thereby advancing both capability and cost on the Pareto curve. SWE-2 continues training from the Kimi K3 model (approximately 2.8 trillion parameters) that has already undergone process agent coding reinforcement learning, and it is said to be able to raise scores by another 5 to 6 points, shifting the entire cost-performance curve.
Comparison table: SWE-2 scored 73.0% on DeepSWE 1.1, higher than Kimi K3's 68.5%, Grok 4.6's 67.5%, and Fable 5.1's 67.4%, and is close to GPT-5.6 Sol's 72.7% and GPT-6 Astra's 74.1%; Terminal-Bench 2.1 scored 92.8%; Terminal-Bench 4 scored 27.3%, significantly lower than Fable 5.1's 55.8% and Astra's 57.9%. The company states that SWE-2 outperforms both SWE-1.7 and Grok 4.6 in scores and costs on FrontierCode and DeepSWE, and approaches Astra at about one-quarter of the cost.
Cognition's product is the software engineer agent Devin. A recent investor letter stated that Devin's internal code production share has risen from about 13% to over 90%. SWE-2 is a model layer update, still targeting workloads for agents writing code, rather than general chat. The evaluations are based on publicly referenced benchmarks, with the drop in Terminal-Bench 4 indicating that long-range terminal tasks have yet to align with the strongest frontiers.
In market mechanics, this represents a price war in coding inference. The buyers are engineering teams looking to reduce token bills; the sellers are application labs using open-source large bases combined with proprietary reinforcement learning to offer cheaper scores. Funding is shifting from calling closed-source frontier interfaces to replaceable coding-specific weights. Beneficiaries are Devin clients who have embedded agents into their repositories; those under pressure are general models selling scores close to FrontierCode at high prices. The event is driven by model releases.
On a supplementary note, the tweet stating "up to 70% lower" and the blog post stating "64% lower" are not the same statement. The ownership of the K3 base and commercial terms depend on licensing, which is not clearly stated in the release post.
Source: Public Information
ABAB AI Insight
Cognition does not train a world model from scratch, but continues reinforcement learning on the 2.8 trillion Kimi K3, buying scores at a lower cost. This is the standard path for application labs: the base already has agent coding capabilities, and RL digs out another 5 to 6 points, competing on the Pareto curve rather than as the top poster. FrontierCode closely follows Fable, while Terminal-Bench 4 still lags Astra by nearly 30 points, indicating that the cheaper benchmarks are not applicable to all long-range terminal tasks.
The capital path is "rent the base, self-train RL, sell agent hours." Devin has already written off 90% of internal code, and the model price drop directly translates into gross profit. a16z has framed the story as engineers transitioning from bricklaying to leading agents. Money flows from inference bills into Cognition, then into reinforcement learning across multiple data centers. The motivation is to divert coding workloads with pricing while Astra and Fable shout the AGI slogan.
Comparing DeepSeek using open-source bases to cut prices, and Windsurf separately reinforcing retrieval models: both are breaking down the premium of general models. The industry phase has shifted from who reaches a trillion first to who can do reinforcement learning cheaply at a trillion. Control lies in the evaluation combinations; those who choose FrontierCode appear close to the frontier, while those who choose Terminal-Bench 4 see the gap.
Structural changes indicate a transfer of pricing power. The mechanism is: once the willingness to pay for coding tasks separates from general dialogue, specialized post-training can penetrate the closed-source premium; as long as agents are already written into production repositories, models can be swapped, but workflows cannot. When scores are close, cost becomes the sole decision variable.
ABAB News · Cognitive Law
Second place close to the frontier becomes the first choice through pricing
Reinforcement learning on a trillion base buys scores, not origins
Once agents are written into repositories, models can be replaced with cheaper weights