Cognition Releases SWE-1.7: Near-Frontier Coding Model at Low Cost
Melvin Vivas · X post · 2026-07-09 · Open on X
Topics: Industry Trends & Job Market, LLM Fundamentals, Fine-tuning & Model Customization · Level: intermediate
Summary
Shares Cognition's announcement of SWE-1.7, their most capable model so far. Cognition says it scores within a few points of the strongest frontier models at a fraction of the cost and runs at 1000 tokens/s. They also say that scaling RL with a refined recipe keeps producing gains.
Key points
- SWE-1.7 is Cognition's most capable model to date.
- Scores within a few points of top frontier models at a fraction of the cost.
- Served at about 1000 tokens per second.
- Cognition says RL is 'not hitting its limit': after refining their recipe, gains continue as they scale.
Resources mentioned
- Devin · tool · x.com · paid
A desktop app from the makers of the Devin coding agent for planning, delegating, reviewing and shipping work across fleets of local and cloud coding agents.
Also in: Devin's open-source vulnerability-fixes agent plugin (Melvin Vivas on X · notes), Gemini 4 Argon now available in Devin (Melvin Vivas on X · notes), Devin Defaults to Opus 5.5 in Normal and Fusion Modes (Cloud) (Melvin Vivas on X · notes), Devin: Preview Websites in Its Built-in Cloud Browser (Melvin Vivas on X · notes) and 57 more - SWE-1.7 · tool · cognition.com · paid
Cognition's coding model, said to be near frontier quality at low cost and 1000 tok/s.
More in Industry Trends & Job Market
- Kimi K3 release: 2.8T params, 1M context, native multimodal
- AI News Roundup: Grok 4.5, GPT-5.6 Launch, GPT-Live Voice Models
- Grok 4.5 First Impressions: Opus-Level Output vs Fable
- Meta's Muse Image: An Agentic Image Generation Model
- News: Alibaba Bans Anthropic's Claude After 'Distillation Attack' Claim
- AI News Roundup: Claude Fable 5, Scientist AI, ZCode, NVIDIA RL