GPT-6 Astra
Points and comments are a snapshot, not live.
OpenAI announces GPT-6 Astra, claiming record intelligence and alignment across benchmarks.
GPT-6 Astra is OpenAI's latest model, said to saturate FrontierMath Tier 4 with 98%, ARC-AGI-3 with 99.9%, and ExploitBench with 100%. It improves computer use speed by 1.9x on Mind2Web compared to GPT-5.6 Sol. Pricing is $10 per million input tokens and $50 per million output tokens. Availability rolls out to limited organizations, then to ChatGPT Plus, Pro, Business, Enterprise, and via API, Azure, and AWS Bedrock. Alignment improvements include 0% task boundary violations versus 48% for Sol. Cyber capabilities meet Critical threshold under Preparedness Framework, with safeguards against misuse.
What commenters are saying
Discussion centers on skepticism about benchmark claims, especially ARC-AGI-3. Commenters note OpenAI used a custom Responses API harness that boosted Sol's score from ~8% to ~38% on the public set, questioning the 99.9% figure. Some defend the harness as realistic; ARC confirms the score on their leaderboard. Pricing at 2.5x Sol draws criticism, though token efficiency gains may offset cost. A few users recount high Codex token consumption, while others find caps hard to hit with efficient workflows.