Cognition launches SWE-2, matching Fable 5.1 on FrontierCode at 64% lower cost
simonguozirui · x · 2026-09-11
Cognition announced SWE-2, its first attempt at scaling intelligence/cost, matching Fable 5.1 on FrontierCode at 64% lower cost. The key was a penalty that preserves the shape of the test-time scaling curve during training. SWE-2 is also Cognition's first model with effort levels: a single RL run pushes the whole Pareto curve — medium effort gets both cheaper and smarter, while max effort learns to spend more tokens and turns for top scores.
More from coding & agent
- User burns $200 of Codex credits in one agent turn — 4,700 of 5,000 credits, task unfinished — RileyRalmuto · 2026-09-11
- Real token value of LLM subscriptions measured: SuperGrok Heavy hits 40x ROI, Cursor Ultra lowest — thesaraharminta · 2026-09-11
- Treating Claude like a new teammate: one Slack message, QA bug report in 96 minutes — max_gladysh · 2026-09-11
- Replit ships Routines with budgets for automated agent tasks — amasad · 2026-09-11
- Replit Agent in a robot builds and publishes websites autonomously via MCP — amasad · 2026-09-11
- OmniGet: open-source app downloads and transcribes content from 1,800+ sites — tom_doerr · 2026-09-11