Cognition: GPT-6 Astra nears Fable 5 coding quality on FrontierCode at 64% lower cost
rohanpaul_ai · x · 2026-09-05
Cognition says GPT-6 Astra is coming to Devin, scoring within 0.4 points of Fable 5 on FrontierCode 1.1 at 64% lower rollout cost, plus a new SOTA on its internal testing benchmark. FrontierCode grades whether an agent can produce PRs maintainers would actually merge: 150 tasks built with maintainers from 36 OSS repos, weighted rubrics across correctness, regression safety, tests, scope and style, with zero for any blocking failure. Context: METR found half of passing SWE-bench Verified patches wouldn't be merged.
Related event: GPT-6 Astra Lands on Devin, Nearing Fable 5 at 64% Lower Cost(4 posts)→
More from coding & agent
- Meta models hit 43% share on opencode, up from 3.5% two weeks ago — zephyr_z9 · 2026-09-05
- VLDB Talk: Deep Research Agents Need Co-Design with Retrievers, DR Tulu and AgentIR Show — CShorten30 · 2026-09-05
- VLDB talk: Deep Research agents need co-designed retrievers, DR Tulu and AgentIR show — CShorten30 · 2026-09-05
- LangChain revamps MCP support: moves into main package on FastMCP v4, adds stateless spec — LangChain · 2026-09-05
- Watch GPT Sol (ultra) Sculpt a Banana in Blender, Start to Finish (32x Speed) — D3VAUX · 2026-09-05
- Ling-3.0-flash-VL visual agent demos: screenshot-to-site, GUI operation, video tutoring — FellMentKE · 2026-09-05