SWE-2 Matches Fable 5.1 on FrontierCode at 64% Lower Cost

Cognition has released SWE-2, a coding model post-trained via reinforcement learning on Moonshot AI's Kimi K3 — a 2.8-trillion-parameter open base model. The result scores 50.0% on FrontierCode 1.1 Main, landing within one point of Fable 5.1 while costing 64% less. It is also Cognition's first model with selectable reasoning-effort levels, all trained in a single RL run.

Architecture and RL training recipe

SWE-2 follows the same post-training lineage as SWE-1.7, which was built on Kimi K2.7. Scaling to K3 brings roughly a 3× parameter increase; Cognition reports RL still finds 5 to 6 points of headroom on top of the base model across most benchmarks — consistent with the pattern that architectural specificity and training recipe choices outperform raw GPU scaling when the base model is already strong.

The core algorithmic contribution is the reward formulation: R = S − λC, where S is binary task success and C is a weighted mix of inference cost in USD and rollout time. Cognition proves that a strictly linear penalty is required for the RL objective to reduce cleanly to a function of average cost and solve rate. Each effort level's λ is set to the local slope of the base model's Pareto frontier, making the iso-reward line tangent to that frontier — so reward can only increase by pushing the frontier upward, not by drifting to a cheaper or more expensive point within the existing curve. All three effort levels are trained in a single RL run.

A second detail, carried forward from SWE-1.6, is a length-weighted reward baseline. Gradient magnitude correlates with rollout length, so the group baseline is computed as sum(R × L) / sum(L) rather than a plain mean. Cognition reports this kept inference-to-training KL divergence lower and stabilized training without additional compute.

On the infrastructure side, a prefill delayer batches nearby requests, raising throughput by 10–20% in TPM per GPU and TPS per request. DSpark speculative decoding accelerates rollouts, with the draft model retrained via SpecForge for 15% longer accept lengths and then trained online alongside the policy. NVFP4 and FP8 kernels with quantization-aware training hold train-inference mismatch below SWE-1.7 levels.

Benchmark results

SWE-2 leads Terminal-Bench 2.1 outright and beats its K3 base across every row. The conspicuous gap is Terminal-Bench 4, where SWE-2's 27.3% trails Fable 5.1 by 28.5 points and GPT-6 Astra by 30.6 points. FrontierCode is Cognition's own benchmark, and all competitor scores come from Cognition's evaluation harness — a relevant consideration when interpreting the one-point gap over Fable 5.1.

Behavioral efficiency

SWE-1.7 over-explored on straightforward tasks. SWE-2 medium addresses this: on FrontierCode 1.1 Main it scores higher than SWE-1.7 while reducing turns by 58% and cost by 81%. Mean steps per run drop from 127 for SWE-1.7 to 53 for SWE-2 medium, 80 for high, and 98 for max. The median steps-to-first-real-edit falls from 48 to 18 at the medium effort level. This kind of pipeline efficiency gain matters in production agentic workflows where turn count directly translates to latency and cost.

AI Mastery analysis

The Terminal-Bench 4 deficit deserves the most scrutiny. At 27.3%, SWE-2 is much closer to its K3 base (21.5%) than to Fable 5.1 (55.8%) or GPT-6 Astra (57.9%) — a roughly 30-point gap that suggests the RL recipe, despite its sophistication, does not yet generalize to the hardest long-horizon terminal tasks. That is a meaningful constraint for teams considering SWE-2 for open-ended infrastructure automation rather than well-scoped coding tickets.

The deployment constraint is equally consequential: no open weights, no standalone API. SWE-2 runs exclusively inside Devin — Desktop and CLI today, with Web and Fusion rolling out. Teams cannot self-host, fine-tune further, or integrate the model directly into their own agent stacks. Given ongoing discussion about infrastructure isolation as the operative AI security boundary, the closed deployment model reduces certain attack surfaces while eliminating the optionality that open-weight alternatives provide.

The length-weighted baseline (sum(R × L) / sum(L)) is an underappreciated systems detail. Standard group baselines in GRPO-style training treat a one-token rollout and a 10,000-token rollout identically when computing the advantage signal, inflating gradient variance on long sequences. Weighting by token count normalizes this without requiring a separate value network — a compute-free stabilization Cognition has now used across two model generations.

The slope-matched λ formulation is conceptually clean, but its practical effect depends on how accurately the base model's Pareto curve is characterized before RL begins. If the curve is sparsely sampled, the tangent point for each effort level's λ is misspecified, and the guarantee that iso-reward lines don't encourage drift within the existing frontier breaks down. Cognition does not detail how densely it samples K3's Pareto curve prior to setting λ values.

The free-trial window for paid Devin subscribers (Pro, Max, and Teams) closes October 10, 2026, after which standard pricing applies. Cognition's approach — closing capability gaps with post-training rather than proprietary pretraining — illustrates that the marginal value of base model scale is increasingly being captured at the RL layer, a dynamic with broad implications for how coding agent vendors differentiate.

Primary source

Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost — MarkTechPost