Cognition's SWE-2: near-frontier coding at 64% less, but Devin-only
Post-trained on Moonshot's 2.8T Kimi K3, it tops Terminal-Bench at 92.8% and edits far sooner. Plus Cohere's open 218B translator and OpenAI's Agents API.
Copy markdown
Near-frontier code, a third of the price
SWE-2 scores 50.0% on FrontierCode 1.1 Main — within a point of Anthropic's Fable 5.1 — while costing 64% less, and posts 92.8% on Terminal-Bench 2.1. For daily coding work, that's frontier-class autonomy without the frontier bill.
It stops overthinking
Cognition tuned SWE-2 for 'focused exploration': the medium tier makes its first real edit after a median of 18 steps, versus 48 for SWE-1.7. Less flailing means tighter agent loops and faster turnarounds on real tasks.
The catch: it's Devin-only
SWE-2 is proprietary, post-trained from Moonshot's 2.8T-parameter Kimi K3, and ships only inside Devin Desktop, CLI, Web, and Fusion — there's no standalone API. If you're not a Devin user, you can't drop it into your own harness yet.
Elsewhere: Cohere open-sources a 218B translation MoE
North-Small-Translate-1.0 (218B total, 25B active, 128 experts) scores 83.6 on WMT26 across 50+ languages, beating DeepL and Google Translate — and the weights are on Hugging Face under CC BY-NC 4.0. Self-host research-grade translation for free.
Elsewhere: OpenAI's Agents API hits public beta
OpenAI opened its Agents API to all developers — a managed Codex harness handling sessions, orchestration, context compaction, recovery, sandboxed execution, and MCP connections (US data residency for now). It's the 'build an agent without wiring the plumbing' path.