Breaking

Cognition Releases SWE-2 Coding Model With 50.0% FrontierCode Score and Lower-Cost Devin Rollout

September 22, 2026 at 09:16 EDT

Foundation Models
AI Agents
Software Dev & Coding

Cognition has released SWE-2, a new coding model for Devin that the company says reaches near-frontier software-engineering benchmark scores while cutting cost against leading rivals. The company said SWE-2 scored 50.0% on FrontierCode 1.1 Main, within one percentage point of Fable 5.1, while being 64% cheaper at that comparison point in its launch post.

Sep. 10, 2026 · Cognition

SWE-2 pushes Devin's cost curve

Cognition says its new coding model reaches 50.0% on FrontierCode 1.1 Main while costing 64% less than Fable 5.1 at the comparison point. The same benchmark table also exposes a sharp gap on Terminal-Bench 4.

50.0%

FrontierCode 1.1 Main

64%

Lower cost than Fable 5.1

58%

Fewer turns than SWE-1.7

81%

Lower average cost than SWE-1.7

One model, two benchmark readings

SWE-2's strongest reported terminal score is 92.8%, but the newer Terminal-Bench 4 row is 27.3%.

92.8%Terminal-Bench 2.1

27.3%Terminal-Bench 4

The visual gap is the key caveat: near-solved older terminal tasks do not guarantee leadership on newer long-horizon command-line work.

Why teams may try it

Lower reported task cost, fewer turns, faster first edits, and Devin Desktop and CLI availability make SWE-2 practical for existing Devin users.

What still needs proof

The scores are company-reported, there is no standalone API or open weights, and Terminal-Bench 4 shows a major remaining gap.

How Cognition says it changed the model

1 · Kimi K3 base2.8T-parameter starting point

→

2 · Single RL runMedium, high, and max effort trained together

→

3 · Devin rolloutDesktop and CLI first, Web and Fusion rolling out

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.