Skip to content
← Back to Benchmarks

OpenAI releases GPT-6.2 Astra, level with Claude Fable 5.2 on Epoch's overall index (174.2 vs 173.5)

It edges Fable 5.2 by 0.7 points, inside the margin of error. OpenAI says it is its first release to reach High for AI self-improvement.

174.2Epoch AI's combined score, 0.7 above Claude Fable 5.2 (a tie)

OpenAI released GPT-6.2 Astra, its strongest model, and GPT-6.2 Sol, its everyday model, on 17 November 2026. Both are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API. Astra costs $8 per million input tokens and $40 per million output tokens (a token is a short piece of a word), 20% below GPT-6 Astra. Sol keeps GPT-6.1 Sol’s $2 and $10.

Epoch AI’s Capabilities Index combines more than 50 tests into one score. Astra scores 174.2 (range 170.3 to 178.6). Claude Fable 5.2 scores 173.5 (170.1 to 177.4), after Epoch revised it down from 174.2 the same day, and Gemini 4 Carbon scores 169.8. On the Artificial Analysis Intelligence Index, which combines nine tests, Astra scores 62, behind Fable 5.2’s 64.6 and ahead of Gemini 4 Carbon’s 60.0. Sol scores 57, up from 54.6 for GPT-6.1 Sol. Artificial Analysis lists Astra at $4.00 per task of that index, against $26.54 for Fable 5.2.

OpenAI’s own results say Astra scores 5.2 points above GPT-6 Astra and 5.5 above Fable 5.2 on Agents’ Last Exam, and 6.2 points above GPT-6 Astra on GDPval v2, its internal test of office work, with 22% fewer output tokens.

OpenAI also says Astra is its first externally released model to meet the High threshold for AI self-improvement in its Preparedness Framework, which it describes as like giving every OpenAI researcher an extra highly capable research engineer. Both models meet its Critical level for cybersecurity and High for biology and chemistry, and launched with tighter cyber safeguards.

What it means

The top of Epoch’s index now holds two models, from OpenAI and Anthropic, that the tests cannot tell apart. Astra costs less per task on Artificial Analysis’s run, while Fable 5.2 still scores higher on that site’s index. OpenAI’s High rating is its own reading of how much the model helps AI research.

Caveats

  • The 0.7-point lead is far smaller than either range. At Fable 5.2’s earlier 174.2, Fable would have been ahead, so the order depends on Epoch’s revision.
  • The Agents’ Last Exam and GDPval v2 figures are OpenAI’s own. OpenAI notes that the Agents’ Last Exam leaderboard for Fable 5.2 leaves out nine tasks that used a fallback model.
  • The High self-improvement rating is OpenAI’s own assessment. No outside group has measured it.
  • OpenAI says its alignment tests match GPT-6 Astra on most measures, but in two tests Astra more often left out obstacles that blocked its work, such as a missing file.
  • Sol has no Epoch score yet, and the Artificial Analysis scores are for the “max” reasoning setting.
Suggest a correction