Cognition Releases SWE-2 for Devin, Claiming Lower-Cost Coding Performance

Cognition has released SWE-2, a new coding model for its Devin agent, with availability now in Devin Desktop and the command-line interface. The company says the model narrows the gap with leading coding systems while cutting the cost and detours of an agent’s work.

Devin Web and Fusion are next: Cognition says SWE-2 is rolling out to those products. The release is aimed at a familiar problem with coding agents: a model can spend a long time reading files, planning, and revisiting the same repository before it makes a useful change.

A cheaper result, by Cognition’s measure

Cognition reports a 50.0% score for SWE-2 on FrontierCode 1.1 Main, compared with 50.9% for Fable 5.1 and 53.3% for GPT-6 Astra. It says SWE-2 was 64% cheaper than Fable 5.1 on that evaluation and cost about one-quarter as much as GPT-6 Astra.

Those numbers describe Cognition’s reported benchmark results, not a settled ranking across every coding workload. The company’s own table also shows a mixed picture: SWE-2 scored 92.8% on Terminal-Bench 2.1, but 27.3% on Terminal-Bench 4. A model can be competitive on one kind of agent task without matching the same rivals everywhere.

50.0%FrontierCode 1.1 Main score

Cognition reports SWE-2 at 50.0% on its FrontierCode 1.1 Main evaluation.

64% lowerReported cost versus Fable 5.1

Cognition says SWE-2’s mean rollout cost was 64% lower than Fable 5.1’s on this comparison.

27.3%Terminal-Bench 4 score

Cognition’s published results show a 27.3% score for SWE-2 on Terminal-Bench 4.

Training the agent to move sooner

Cognition says SWE-2 was post-trained from Kimi K3, a 2.8-trillion-parameter model, then tuned with a reinforcement-learning approach that trains multiple reasoning-effort levels in one run. The training objective applies different cost penalties to different effort levels, seeking better solve rates without freely extending each task’s trajectory.

The intended behavior differs by setting. Cognition says medium effort should start work faster on simple and intermediate tasks. High and maximum effort are designed to spend more time planning, inspecting the codebase, and verifying uncertain or complex work.

What Cognition says changed in medium effort

• Average steps per FrontierCode task fell from 127 with SWE-1.7 to 53 with SWE-2 medium.
• The median first substantive edit arrived after 18 steps, down from 48 for SWE-1.7.
• Cognition says SWE-2 medium scored higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average in that test.

The practical test is accepted work

Fewer exploratory steps could make an agent cheaper to run and easier to supervise. But faster action also puts more weight on whether the model chooses the right files, writes a complete change, and verifies it before stopping. Cognition says SWE-2 shows stronger test coverage and verification discipline in its internal testing; those behavioral claims will need to prove useful in ordinary Devin work.

For now, the release offers a concrete choice rather than a universal performance verdict: use a lower-effort mode intended to act quickly, or spend more on longer planning and verification when the job demands it. The critical measure is not simply a benchmark point, but whether those settings help teams reach dependable accepted changes at a lower cost.