GPT-6 Astra launches: 99.9% on ARC-AGI-3, $10/$50 pricing, and two zero-days found in testing
赛博禅心 · wechat · 2026-09-04
OpenAI released GPT-6 Astra overnight, with president Greg Brockman declaring "welcome to the AGI era." Key points from this Chinese rundown:
- Benchmarks & availability: 99.9% on ARC-AGI-3, 100% on ExploitBench, 97.6% on FrontierMath Tier 4; rolling out to ChatGPT Plus/Pro/Business/Enterprise over coming days, plus API and AWS. Priced like Fable: $10/M input, $50/M output.
- Capabilities: PCB layout in KiCad, FreeCAD rendering, form-filling/CRM/calendar tasks; 1.9x faster than Sol on Mind2Web with the new Codex; Codex gains experimental note-taking for exhausted context windows. Cognition, Jane Street and Harvey report major quality gains and up to 20% token savings.
- Math & safety: helped push short prime gaps from 246 to 186 and improve a large-prime-gaps term untouched for 80+ years. Hit the cybersecurity Critical threshold under the Preparedness Framework; scored 39.0% vs Sol's 5.5% on ExploitBench, discovering and exploiting two unknown zero-days (now being disclosed). Best-aligned OpenAI model yet, though its more implicit reasoning is harder to monitor—a regression OpenAI says it takes seriously.
Related event: OpenAI Launches GPT-6 Astra, Declaring 'Welcome to the AGI Era'(157 posts)→
More from Models
- OpenAI's Luna quietly tops OpenRouter charts, first closed model to do so in ages — sherwinwu · 2026-09-04
- Early Astra access floods timelines as users build 3D worlds and games — cedric_chee · 2026-09-04
- Perplexity says Astra excels at computer use, coming to Comet browser — AravSrinivas · 2026-09-04
- UK AISI: Astra's time horizon hits 30.9 minutes, nearly 9x GPT 5.6 Sol's 3.6 — Wonderful_Buffalo_32 · 2026-09-04
- OpenAI officially unveils GPT-6 Astra: anything you can do on a computer, it can do — minchoi · 2026-09-04
- Astra's official ARC-AGI 3 score: 62.7%, double that of Opus 5 — aqpstory · 2026-09-04