Cognition's SWE-2 lands within a point of Fable 5.1

Cognition released SWE-2 on September 10, and the interesting number is not the benchmark — it is the gap between the benchmark and the bill. SWE-2 scores 50.0% on Cognition's FrontierCode 1.1 Main against Fable 5.1's 50.9%, a difference of nine tenths of a point, at 64% lower cost. If you are running a coding agent on frontier pricing out of habit, this is the release that makes you re-run the math.

What actually happened

Per Cognition's writeup, SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter open model that had already been through heavy RL for agentic coding. Cognition's own reinforcement learning sits on top of that base. Published scores: 50.0% on FrontierCode 1.1 Main, 73.0 on DeepSWE 1.1, 92.8 on Terminal-Bench 2.1, and 27.3 on Terminal-Bench 4.

The mechanism worth reading is effort levels. Cognition trained medium, high, and max in a single RL run rather than shipping three separate models. Medium favors speed and costs 81% less on average than SWE-1.7 on the same FrontierCode tasks; high and max spend more compute on hard problems. One training run, three points on the cost curve.

Cognition has not published a per-token price. Every cost claim is relative: 64% below Fable 5.1 at a matched score, 81% below its own predecessor. SWE-2 shipped first in Devin Desktop and the Devin CLI, with Devin Web and Fusion following.

Why a cheaper coding agent matters for your business

Most teams pick one model and route everything through it. That is a rounding error on a hobby project and a real line item at ten engineers. Effort levels turn model choice into a per-task decision, which is where it belonged all along: a dependency bump and a payment-flow refactor should not cost the same.

Two caveats we would hold onto. First, FrontierCode is Cognition's own benchmark. A vendor grading itself against a competitor is useful signal and not an independent result — measure SWE-2 on your repo, on the kind of ticket you actually close, before you switch anything.

Second, no published token price means no contract. "64% cheaper than a competitor" is a ratio that moves when the competitor moves, and it is only reachable through Devin's surfaces today. That is a product decision, not a model you can drop behind your own router. Price the lock-in along with the savings.

The base model deserves a note too. SWE-2 sits on open weights from Kimi K3, a Chinese open release. Your coding agent now has a supply chain, and it runs through a lab you did not pick. Know that before someone in procurement asks.

Sources: Cognition: SWE-2.