GPT-6.2 Astra
OpenAI's new flagship, 62 points and second overall, $8/$40 — a fifth cheaper than its predecessor
Benchmark scores
How it is measured →Ranked 2 · bars are scaled to the field maximum for that benchmark.
| Benchmarks | Score | Vs. best in field | Gap |
|---|---|---|---|
| AA Intelligence Index v5 Intelligence index | 62.0 | 95% | −3.0 pts |
| AA Intelligence Index v4.3.2 | — | — | not indexed |
| GPQA Diamond | — | — | not indexed |
| Terminal-Bench 4.0 Terminal-agent tasks | 64.1 | 95% | −3.1 pts |
| LiveBench Design | — | — | not indexed |
| Vals Index | — | — | not indexed |
Assessment
Released on 17 November 2026, replacing GPT-6 Astra, whose source row was marked deprecated and pulled the same day. AA Intelligence Index v5 tops out at 62.14 at max effort (xhigh also 62), second overall and three points behind Fable 5.2's 65; Terminal-Bench 4.0 (AA's own run) is 64.1%, below Carbon's 65.2 and Kimi K3.1 Pro's 67.2 but above Sonnet 5.5's 63.6. Pricing is $8/$40 with cached input at $0.4 (95% discount), a fifth below the previous generation's $10/$50/$1.00 and 60% lower on cache; the vendor claims +5.2 points over GPT-6 Astra on Agents' Last Exam at under a third of the cost. 1M context, with the output ceiling unpublished. GPQA, Vals and LiveBench all read null in AA's record (not run), so they stay blank rather than guessed.
At the same price
The four models closest on blended price, side by side.
| Model | v5 | Blended | tok/s | Score gap |
|---|---|---|---|---|
| Claude Fable 5.2Anthropic | 65.0 | $16.00 | 95 | +3.0 |
| Claude Opus 5.5Anthropic | 59.1 | $8.00 | 116 | -2.9 |
| Muse WonderMeta | 44.0 | $8.00 | 113 | -18.0 |
| Claude Sonnet 5.5Anthropic | 54.5 | $4.00 | 146 | -7.5 |
Compare side by side
Line it up against other models on price, speed and every benchmark.
The comparison tool lines up to 4 models side by side with switchable normalisation — built for purchasing decisions.
Open the comparison tool → Back to leaderboardRelated benchmarks
Data sources & methodology
Full methodology & sources →Every number on this page comes from published evaluations. We run no tests of our own and accept no vendor-submitted scores. Main sources:
Evaluation data in the wider industry also comes from public datasets, leaderboard platforms and their evaluation harnesses, private in-house sets, production logs and user feedback, and synthetic data. This site's current rankings use only the parts that are public, independently checkable and clearly defined.