Cognition launches SWE-2: frontier-level coding model at up to 70% lower cost
gethackteam · x · 2026-09-14
Cognition officially introduced SWE-2, its model closest to the frontier: it matches recent frontier models on leading evals at up to 70% lower cost. The team scaled RL to multiple trillions of parameters with a refined recipe pushing the Pareto curve on both capability and cost. The reposter adds that if a model post-trained on Kimi K3 can beat Fable/Astra, things will get interesting.
More from coding & agent
- LangChain Publishes Practical Guide to Building Custom Agent Harnesses — hwchase17 · 2026-09-14
- Multi-Agent Coordination in Prod Dev: Mixing Model Families Beats a Single Model — CellistOk6782 · 2026-09-14
- Dev shares AI code review in Clairvoyance: teaching computer players when to expand vs build outposts — draginol · 2026-09-14
- PicoLM runs a 1B-parameter LLM on a $10 board with 256MB RAM — pure C, zero dependencies — tom_doerr · 2026-09-14
- Tip: Make Codex Verify Its Own Work to Boost Visual Task Quality — reach_vb · 2026-09-14
- Driftproof Audited on Launch Day: Forged Passes, Runaway Spend and 6.2% Noise Rate Found — maverick_man1111 · 2026-09-14