GPT-6 Astra vs Claude: all 27 head-to-heads from OpenAI's own tables, read honestly

GPT-6 Astra launched September 3, 2026. This scorecard takes every benchmark in OpenAI's nine launch comparison tables where a Claude number is printed and puts Astra against the best Claude score on that row: 21 wins, 5 losses, 1 tie across 27 head-to-heads, with the full table grouped by category (Astra, best Claude number, which Claude model, win/loss/tie) and the five losses marked in red (Humanity's Last Exam, the Artificial Analysis Intelligence and Coding Agent indexes, FrontierCode Main and Extended). Then the ARC-AGI-3 asterisk (62.7% at $26,098 on the standard harness vs 99.9% at $19,000 on OpenAI's provider adapter, with both effort tables from ARC Prize), price per token vs OpenAI's cost-per-task claims ($10/$50 list, 2x Sol, same as Fable 5.1), the safety card in plain English with system-card section numbers (first Critical cyber tier, two zero-days found in testing, harder-to-monitor reasoning, and the alignment wins), who gets it and when, per OpenAI's own rollout (a limited set of organizations at launch, Plus/Pro/Business/Enterprise over the coming days with Astra Pro for Pro and above, Enterprise off by default until an admin enables it, API gpt-6-astra on Bedrock with ZDR, and the Codex notes-across-context-windows flag as OpenAI describes it), with the advice written for once Astra reaches your account. Independent scorecard from Hyperautomation Labs; every number is OpenAI's or ARC Prize's, footnotes included.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — three practical AI emails a week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.