Cognition released SWE-2 in Devin, saying it costs 64 percent less

Cognition released SWE-2 today, the follow-up to its SWE-1.7 coding model, and put it in Devin Desktop and the CLI with a rollout to Devin Web and Fusion in progress. Cognition says SWE-2 was post-trained on Kimi K3, which it describes as a 2.8-trillion-parameter base model that had already been reinforcement-learned for agentic coding.

The numbers Cognition is claiming

On FrontierCode 1.1 Main, Cognition says SWE-2 scores 50.0 percent, against 50.9 percent for Fable 5.1 and 53.3 percent for GPT-6 Astra, with SWE-1.7 at 42.0 percent on the same benchmark. On DeepSWE 1.1 the company reports 73.0 percent for SWE-2 and 74.1 percent for GPT-6 Astra, with SWE-1.7 at 37.7 percent. On Terminal-Bench 2.1 Cognition puts SWE-2 at 92.8 percent, up from 81.5 percent for SWE-1.7. Every one of these is a vendor benchmark from Cognition's own release post, so treat the comparisons as marketing until an independent harness replicates them.

The cost claim is where the news is. Cognition says SWE-2 scores "within one point" of Fable 5.1 on FrontierCode while costing "64 percent" less to run, and that SWE-2 medium beats SWE-1.7 on the same benchmark while taking "58 percent fewer turns" and costing "81 percent less on average". The model exposes three effort levels: medium is described as quicker action for simple and intermediate tasks, and high and max as more planning and exploration for complex ones.

What actually changes for a Devin user

If your team already runs Devin on SWE-1.7 for day-to-day coding, SWE-2 is a direct replacement that Cognition claims halves the turn count and cuts the cost by around four fifths on the same jobs. If you are paying a Fable 5.1 or an Astra price for top performance and only occasionally need it, medium and high on SWE-2 are presented as most of the score at a smaller fraction of the cost, with max available for the harder work. Cognition does not give public pricing in the post, so the concrete number will come from your account, not from the blog.

The Fusion product, which routes work between models in a Devin session, is still rolling out with SWE-2 as one of the options, and Devin Web is being updated. Devin Desktop and the CLI have it available immediately.

The claim to wait on until an outside test is published is the Fable 5.1 comparison. A model priced 64 percent below a frontier competitor while scoring within a point of it on the vendor's own benchmark is exactly the kind of claim that real workloads often contradict within a week, and Terminal-Bench 4 is where SWE-2 falls to 27.3 percent in Cognition's own table, so its results on the hardest tasks do not yet match the leading models. Read the release post for the caveats.

Source

SWE-2 on the Cognition blog, 10 September 2026. Hacker News discussion.