Cognition's SWE-2 matches Fable 5.1 on code at 64% lower cost
It's built on Moonshot's open Kimi K3 and RL-tuned for cheap runs — but it lives only inside Devin, with no weights and no standalone API.
Copy markdown
The pitch: near-frontier code, 64% cheaper
SWE-2 scores 50.0% on FrontierCode 1.1, a hair behind Fable 5.1's 50.9%, while Cognition claims 64% lower cost. If you live in Devin, that's frontier-grade agentic coding without the frontier bill.
Built on open Kimi K3, RL-tuned for cost
It's a post-train of Moonshot's 2.8T-parameter Kimi K3 with RL that penalizes wasted steps. Three effort levels ship together; medium alone cuts turns 58% and cost 81% versus SWE-1.7, so you dial the speed/cost tradeoff per task.
The catch: Devin-only, no weights, no API
SWE-2 runs only inside Devin Desktop and CLI — no open weights, no standalone API to drop into your own stack. It's free through Oct 10 for paid Devin tiers, then metered.
Strong on real repos, weak on hard terminals
Beyond FrontierCode: 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, but just 27.3% on the tougher Terminal-Bench 4 — expect it to shine on everyday PRs and stumble on gnarly shell work.
Elsewhere: Cohere's open translator beats DeepL
Cohere Labs shipped North Small Translate 1.0, a 218B MoE scoring 83.6 on WMT26 across 50 languages and topping DeepL and Google Translate. Weights are on Hugging Face (noncommercial) plus an API — a self-hostable translation layer for your app.