Cognition SWE-2: Near-Frontier Coding AI at 64% Lower Cost Than Fable 5.1

Cognition, the company behind the Devin coding agent, released SWE-2 on September 12, 2026. The model is post-trained with reinforcement learning on top of Kimi K3, Moonshot AI's 2.8 trillion-parameter open base model. It scores 50.0% on Cognition's own FrontierCode 1.1 Main benchmark, landing within 1 point of Fable 5.1 at 64% lower cost. It also cuts the average number of task steps from 127 to 53 compared to SWE-1.7. The catch: no open weights, no standalone API, Devin only.

What happened

Detail	Value
Release date	September 12, 2026
Base model	Kimi K3 (Moonshot AI, 2.8T parameters)
FrontierCode 1.1 Main score	50.0%
Cost vs Fable 5.1	64% lower
Average steps (medium) vs SWE-1.7	53 vs 127
Cost reduction at medium effort vs SWE-1.7	81% lower
Effort levels trained	3, in one RL run
Free access window for paid tiers	Until October 10, 2026

SWE-2 is Cognition’s most capable coding model yet. It follows SWE-1.7, which was itself post-trained from Kimi K2.7. This time Cognition scaled its reinforcement learning process to a base model nearly three times as large. According to Cognition, RL on top of K3 adds 5 to 6 points across most benchmarks, showing there is still meaningful headroom even at the multi-trillion-parameter scale.

The model is the first from Cognition to offer selectable reasoning-effort levels: medium, high, and max. All three are trained simultaneously in a single RL run, which is a notable efficiency over training separate models for each cost tier.

How does SWE-2 compare to rival models?

Benchmark	SWE-2	Kimi K3	Grok 4.6	Fable 5.1	GPT-5.6 Sol	GPT-6 Astra	SWE-1.7
FrontierCode 1.1 Main	50.0%	44.2%	48.0%	50.9%	47.5%	53.3%	42.0%
DeepSWE 1.1	73.0%	68.5%	67.5%	67.4%	72.7%	74.1%	37.7%
Terminal-Bench 2.1	92.8%	88.3%	88.4%	91.4%	88.8%	89.9%	81.5%
Terminal-Bench 4	27.3%	21.5%	20.3%	55.8%	37.3%	57.9%	7.6%

SWE-2 leads the field on Terminal-Bench 2.1 and beats its K3 base on every row. One clear weak spot stands out: Terminal-Bench 4, where SWE-2 scores 27.3% compared to 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra. That is a roughly 30-point gap. Cognition also says SWE-2 comes within a few points of GPT-6 Astra at about a quarter of the cost, though that claim is worth reading carefully since FrontierCode is Cognition’s own benchmark and rival numbers come from Cognition’s own evaluations.

What changed in the training recipe

The RL reward function is defined as R = S minus lambda times C, where S is binary success (did the task complete?) and C is a mix of inference cost in USD and rollout time. Cognition argues that only a linear cost penalty keeps the RL objective cleanly tied to average cost and solve rate. Each effort level’s lambda (the cost weight) is set to the local slope of the base model’s Pareto curve, so the reward surface is tangent to the frontier. In plain terms: the training naturally pushes all three effort levels forward together, rather than trading off against each other.

A few other technical changes are worth noting for teams evaluating AI coding tooling:

• A length-weighted reward baseline, in use since SWE-1.6, weights the group baseline by tokens (sum of R times L divided by sum of L). Cognition says this kept training more stable with no extra compute.
• A prefill delayer batches nearby requests to raise throughput by 10 to 20% per GPU.
• DSpark speculative decoding (a technique where a smaller draft model predicts tokens for the main model to accept or reject) accelerates rollouts. The draft model was retrained via SpecForge for 15% longer accept lengths and then trained online alongside the main policy.
• NVFP4 and FP8 quantization-aware training keeps memory down and reduces the gap between training and inference behavior.

Efficiency gains over SWE-1.7

SWE-1.7 had a tendency to over-explore on simple tasks, spending many turns before making a first edit. SWE-2 medium makes its first real code edit after a median of 18 steps, compared to 48 for SWE-1.7. Mean steps per run drop to 53 (medium), 80 (high), and 98 (max) from 127 for SWE-1.7. On FrontierCode 1.1 Main, SWE-2 medium actually scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less.

Cognition also reports three behavioral improvements: stronger end-to-end test coverage, better resourcefulness when a tool is blocked or unavailable, and what they call verification discipline, meaning when challenged on a conclusion the model re-derives the answer rather than simply re-asserting it.

Where you can actually use it

SWE-2 has no open weights and no public API. It runs exclusively inside Devin: currently available in Devin Desktop and CLI, with Devin Web and Fusion still rolling out. Paid Devin subscribers get access free until October 10, 2026. If you were hoping to self-host or call it from your own stack, that is not an option.

Our take

The efficiency story here is the real headline, not the raw benchmark score. Matching a frontier model at 64% lower cost, while cutting task turns by more than half, is what matters for teams running agents on real work at volume. Cost and turn count are the two numbers that blow up production budgets.

That said, the Terminal-Bench 4 result (27.3% versus Fable’s 55.8%) is a significant gap that Cognition does not explain away. For anything involving sustained, complex terminal workflows, SWE-2 is not the top pick. And the lock-in is real: no weights, no API, Devin-only. For teams thinking about AI integration into their development workflows, that is a meaningful constraint worth factoring in before building any process around it.

The training methodology is genuinely interesting, particularly the Pareto-informed cost penalties and the single-run multi-effort approach. If these techniques carry over to open-weight post-training, they could matter a lot for the broader ecosystem. For now, they are Cognition’s proprietary advantage, explored further in the discussion of deterministic versus agentic process models.

Watch Terminal-Bench 4 as SWE-2 matures. If that gap closes, the cost argument becomes very hard to argue against.

Source: Marktechpost