GPT-6.2 Astra released: League 100.0 at #1 — on two components, not one.
OpenAI released GPT-6.2 Astra on November 17, 2026 at $8/$40 per million tokens — 20% below GPT-6 Astra and exactly Claude Fable 5.2’s headline rates. It enters the League at 100.0, tied with Fable 5.2 at #1 — but its 100.0 rests on two tracked components at 55% coverage, where Fable’s rests on one at 40%. This post works both normalisations in full, prices the tie, and records the Sol paradox alongside it: more evidence, a slightly lower score.
By AI Pro League · Published

GPT-6.2 Astra is OpenAI’s new frontier model, released November 17, 2026. Our November 18, 2026 changelog records the release from OpenAI’s official announcement and API pricing and model pages (developers.openai.com, read today): the model is available in ChatGPT Work and Codex and as gpt-6.2-astra in the API, with a 1,050,000-token context window (922,000 max input), 128,000 max output, a May 18, 2026 knowledge cutoff and text+image inputs. It enters our database of 67 models — 54 of them scored — at the top of the AI model rankings, and unlike most launch-day entries it does not enter unranked: two tracked leaderboards already carry its results. What follows is computed from the database at build time.
The two normalisations behind the GPT-6.2 Astra score, worked in full
LiveBench carries 40% of the League Score on this site — more than any other single component. Astra records 84.9 on LiveBench 2.0 (livebench.ai via BenchLeader, results dated 2026-11-06; read today) — the maximum of our tracked field of 37 models, tied with Claude Fable 5.2. Normalised against that field, which runs from 32.4 to 84.9: (84.9 − 32.4) ÷ 52.5 × 100 = 100.0. GPQA Diamond carries 15%. Astra records 97.0 per the BenchLM GPQA leaderboard (benchlm.ai, updated 2026-11-18) — the new maximum of our 27-model GPQA field, which runs from 73.0 to 97.0: (97.0 − 73.0) ÷ 24.0 × 100 = 100.0. The League Score is the weighted mean of the components present, re-based on the coverage they carry: (100.0 × 0.40 + 100.0 × 0.15) ÷ 0.55 = 100.0, at 55% coverage — two components of the five the League scores. The boards themselves are set out in our LiveBench explainer and GPQA Diamond explainer, and the standings live on the LiveBench leaderboard and the GPQA Diamond leaderboard.
Two models at 100.0: the coverage difference
| Model | League Score | Overall rank | Coverage | Components (normalised) | Workload cost |
|---|---|---|---|---|---|
| GPT-6.2 Astra | 100.0 | #1 | 55% | LiveBench 100.0, GPQA Diamond 100.0 | $160.00 |
| Claude Fable 5.2 | 100.0 | #2 | 40% | LiveBench 100.0 | $160.00 |
| GPT-6 Astra | 98.3 | #3 | 55% | LiveBench 99.2, GPQA Diamond 95.8 | $200.00 |
Read the tie honestly. Both 100.0s are real, and they are not the same claim. Fable 5.2’s 100.0 — recorded in our Claude Fable 5.2 release analysis — rests on the single LiveBench component at 40% coverage: one component, measured once, at the field maximum. Astra matches that raw 84.9 and adds a second field maximum on an independent board, so its 100.0 is corroborated at 55% coverage. Neither is a complete record — our score coverage explainer sets out why we print the coverage beside every score — but of the two, Astra’s is the finding and Fable’s remains the promising rumour. The predecessor is not displaced from sale: GPT-6 Astra stays current at 98.3 and #3, at unchanged prices.
What GPT-6.2 Astra costs on a workload
The price record, checked Nov 18, 2026: $8.00 / $40.00 per million tokens, cached input $0.40 — 20% below GPT-6 Astra on input and output, and 60% below it on cached input. On the standard workload Astra costs $160.00 — 10 × $8.00 + 2 × $40.00 — against $200.00 for its predecessor and $160.00 for Fable 5.2, whose headline rates are identical to Astra’s. The one rate that differs is cached input: $0.40 against Fable’s $0.20 — a 95.0% discount against 97.5%. On the cached-heavy run (8M cached + 2M fresh input + 2M output) that gaps shows up in full: Astra costs $99.20 — 8 × $0.40 + 2 × $8.00 + 2 × $40.00 — against $97.60 for Fable 5.2 and $128.00 for GPT-6 Astra. At the flagship tier the headline price is now a tie; the cached rate is where the two 100.0s part. Full rate cards sit in our AI model pricing index.
GPT-6.2 Sol: more evidence, a slightly lower score
| Model | LiveBench raw | Normalised | GPQA raw | Normalised | League Score | Coverage | Overall rank | Workload cost |
|---|---|---|---|---|---|---|---|---|
| GPT-6.2 Sol | 80.3 | 91.2 | 96.0 | 95.8 | 92.5 | 55% | #9 | $40.00 |
| GPT-6.1 Sol | 81.4 | 93.3 | — | — | 93.3 | 40% | #8 | $40.00 |
The everyday tier shipped the same day at $2.00 / $10.00 — unchanged from GPT-6.1 Sol on every rate — and it carries the release’s quiet paradox. GPT-6.2 Sol has the deeper record: a GPQA Diamond 96.0 (normalised 95.8, tied with GPT-6 Astra on the raw board) takes it to 55% coverage, where its predecessor has only LiveBench at 40%. Yet its League Score is lower — 92.5 at #9 against 93.3 at #8 — because the component both models share moved the other way: LiveBench 80.3 (normalised 91.2) against 81.4 (normalised 93.3). More evidence, slightly less of it at the top: that is what a second component does to a score, and it is the honest shape of a maturing record.
What the #1 ranking does not yet prove
- Three of five components are unmeasured. Astra has no tracked SWE-bench Verified, LMArena or MMLU-Pro result, so 45% of the League Score’s weight — coding, crowd preference and knowledge — has not been tested for it at all.
- The #1 is shared, and shared narrowly. Astra and Fable 5.2 sit at the same 100.0; the ordering between them rests on coverage, not on a score gap, and a single new component for either model re-orders the top.
- Normalisation moves with the field. Both 100.0s are computed against today’s fields (LiveBench 32.4–84.9, GPQA 73.0–97.0); a re-measurement or a new field maximum re-scales every component, including these.
- The launch benchmarks are not the tracked ones. OpenAI’s headline claims for this release rest on its own evaluations; the League scores only the two tracked boards above, and treats the rest as unmeasured until a tracked source publishes.
The score to watch next is Astra’s third component: whichever of SWE-bench Verified, LMArena or MMLU-Pro publishes first will either corroborate the 100.0 or pull it back toward the field. The model’s full record lives on its GPT-6.2 Astra page, its sibling’s on the GPT-6.2 Sol page, and the model it ties with on the Claude Fable 5.2 page.