Every's Vibe Check: GPT-6 Astra impresses at writing and coding, but Fable still builds better
every · x · 2026-09-04
Every published a deep Vibe Check of OpenAI's GPT-6 Astra:
- Model-written draft: The review's first draft came from a single prompt; CEO Dan Shipper mistook it for human writing — "RIP my job."
- Claimed specs: 99.9% on ARC-AGI-3, 98% on FrontierMath Tier 4, with SOTA computer use/software engineering and template-following documents; OpenAI says it helped solve long-standing open math problems.
- Verdict: Impressive writing, software operation and visual design, but with bad habits; Anthropic's Fable still has better instincts for building products.
A live Fable 5.1 vs GPT-6 Astra camp for members follows on September 4.
Related event: Every's Hands-On GPT-6 Astra Test: Best Writing Model, Still Behind Fable(13 posts)→
More from Models
- UK AISI assessment: Astra reasons in a more compressed style than GPT 5.6 Sol — maksym_andr · 2026-09-04
- OpenAI Mathematician Elliot Glazer: Only o3 and Sol Were Step Changes in Autonomous Math — danintheory · 2026-09-04
- Tester: AI-text detector Pangram shows zero false positives, but adversarial rewriting evades it — alex_peys · 2026-09-04
- OpenAI researcher teases upcoming model: 'best in class' at computer use, research, coding — athyuttamre · 2026-09-04
- FrontierMath is 'dead' after 664 days as frontier models conquer the math benchmark — willdepue · 2026-09-04
- Matt Shumer's GPT-6 Astra review: Manager Loop sustains a week-long Manhattan build in Unreal — mattshumer_ · 2026-09-04