Cognition introduces SWE-2, a coding model for Devin

Cognition introduced SWE-2, a coding model for its Devin product, and claims 50.0% on FrontierCode 1.1 Main — within about a point of Anthropic’s Fable 5.1 — at 64% less cost.

A Devin-only model that claims near-Fable scores at a lower cost is a product launch, not a public weight drop. Keep the benches and the 64% cheaper line as Cognition’s.

Cognition published “Introducing SWE-2.” It calls this the company’s most advanced coding model yet. The post says it is available that day in Devin Desktop and CLI (command-line interface), with Devin Web and Fusion rolling out.

Cognition’s headline claim is 50.0% on FrontierCode 1.1 Main, a company coding test. It says that score is “within one point of Fable 5.1” (50.9% on the same table) and “64% cheaper.” The score sits 0.9 percentage points below Anthropic’s listed Fable 5.1 result. Score and cost cut are Cognition’s.

The same Cognition table: DeepSWE 1.1 at 73.0%, Terminal-Bench 2.1 at 92.8%, Terminal-Bench 4 at 27.3% — versus 55.8% for Fable 5.1 on that last row. Still Cognition’s numbers.

Cognition says SWE-2 was further trained from Kimi K3, a model with 2.8 trillion learned weights. That base had already had extra training that rewards finishing coding jobs — reinforcement learning (RL). The post says this is the first time the company scaled that recipe to a multi-trillion-weight model.

No public weights and no standalone per-token API are described on the page. Use is inside Devin.