GPT-6.2 Astra hits 78.3% on Agents' Last Exam at $3/task, crushing Fable's $60
haider1 · x · 2026-11-18
On Agents' Last Exam v1.1, GPT-6.2 Astra leads with 78.3% at just $3 per task, versus Claude Fable 5.2 at 72.8% ($60/task) and Claude Opus 5.5 at 71.6% ($17.50/task) — a dramatic cost-performance gap. The author also cites the claim that GPT-6.2 Astra is OpenAI's first publicly released model to reach the High threshold for AI self-improvement, arguing the path to RSI is becoming clearer. The attached OpenAI chart shows Astra sitting at the top-left of the score-vs-cost curve.
Related event: GPT-6.2 Astra Tops Agents' Last Exam at Fraction of Cost(4 posts)→
More from AGI Musings
- Building in the roaring 20s: like being dropped into the industrial revolution — tekbog · 2026-11-18
- GPT ran a real microscope for 72 hours, screening 3x more drugs and rewriting its own scheduler — anshulkundaje · 2026-11-18
- Dario Amodei warns open-source models could lead to "a very dangerous place" — mark_k · 2026-11-18
- OpenAI chief scientist's risk warning sparked a fresh case for serious AI regulation — trevposts · 2026-11-18
- Professor slams AI-detector reliance: review writing by reading it, not proxies — ipeirotis · 2026-11-18
- Thomas Nagel wins 2026 Berggruen Prize; his consciousness essay shapes AI debates — begusgasper · 2026-11-18