Home  /  Blog  /  Best AI model — November 2026
League analysis

Best AI model — November 2026: Gemini 4 Neon leads at 100.0, on one component.

The best AI model in November 2026, by the League Score, is Gemini 4 Neon at 100.0 — ahead of GPT-6.2 Astra at 96.7 and Claude Fable 5.2 at 95.5. The ranking is real, and it is thin: Neon’s 100.0 rests on a single LiveBench component at 40% coverage, while the model in fifth place, Claude Opus 5, carries the joint-deepest record in the League at 90%. This post works the maths in full, prices the top five on one standard workload, and says what the #1 ranking does not yet prove.

By AI Pro League · Published

Best AI model November 2026 — Gemini 4 Neon leads the AI Pro League at a League Score of 100.0, ahead of GPT-6.2 Astra at 96.7 and Claude Fable 5.2 at 95.5

The question “which is the best AI model” has three different answers in November 2026, and the AI model rankings give all three. The highest score belongs to Gemini 4 Neon: a League Score of 100.0, overall #1, from a single LiveBench result of 87.4 measured Nov 20, 2026 on the LiveBench 2.0 board — the highest raw score in the 38-configuration tracked field. The best-tested model at the top is Claude Opus 5, fifth at 93.1, whose score rests on four components at 90% coverage. And the cheapest model at the top is Gemini 4 Carbon, seventh at 92.5, at $40.00 on the standard workload against Neon’s $60.00. What follows is computed from our database of 68 models, 56 of them scored, at build time.

The best AI model in November 2026 — the top seven, with coverage beside every score

The top of the League, from the AI Pro League database — standard workload is 10M input + 2M output tokens
RankModelLeague ScoreCoverageComponentsStandard workload
1Gemini 4 Neon100.040%1$60.00
2GPT-6.2 Astra96.755%2$160.00
3Claude Fable 5.295.540%1$160.00
4GPT-6 Astra94.855%2$200.00
5Claude Opus 593.190%4$100.00
6Claude Opus 5.592.740%1$80.00
7Gemini 4 Carbon92.540%1$40.00

Four of the top seven sit at 40% coverage — one component each, all of it LiveBench, which carries 40% of the League Score. Only two models in the top seven carry a second component, GPT-6.2 Astra and GPT-6 Astra, and only one, Claude Opus 5, carries more than two. That is the shape of the top of the table this month: a single board deciding nearly everything, and a long tail of models whose records are deeper but whose LiveBench raw sits lower. Our score coverage explainer sets out why we print the coverage beside every score on this site, and our LiveBench explainer sets out how the board itself is scored.

The maths worked in full

Neon’s 100.0 is a field maximum, not a perfect paper. LiveBench carries 40% of the League Score, and Neon’s raw 87.4 is the highest in the 38-configuration tracked field, which runs from 32.4 to 87.4 — worked step by step: (87.4 − 32.4) ÷ 55.0 × 100 = 100.0. The League Score is the weighted sum over the coverage the model actually has: 100.0 × 40% = 40.0, divided by the 40% coverage gives 100.0. A model that tops the field on its only component scores 100.0 by construction — the number says Neon leads the board, and says nothing about the four components it has not yet been measured on.

GPT-6.2 Astra shows what a second component does. Its LiveBench 84.9 normalises against the same field: (84.9 − 32.4) ÷ 55.0 × 100 = 95.5. Its GPQA Diamond 97.0 is the maximum of the 28-model tracked field, which runs from 73.0 to 97.0: (97.0 − 73.0) ÷ 24.0 × 100 = 100.0. The weighted sum is 95.5 × 40% + 100.0 × 15% = 53.20, divided by the 55% coverage gives 96.7 — 3.3 points behind Neon on a deeper record. Claude Fable 5.2 is the mirror image: the same LiveBench 84.9 as Astra, the same normalised 95.5, but no second component, so its 95.5 is the LiveBench number alone. How each board works is set out in our GPQA Diamond explainer and our LiveBench explainer.

Claude Opus 5 is the counterweight. Its four components normalise to LiveBench 90.4, SWE-bench Verified 100.0, GPQA Diamond 84.2 and LMArena 100.0; the weighted sum is 90.4 × 40% + 100.0 × 20% + 84.2 × 15% + 100.0 × 15% = 36.16 + 20.00 + 12.63 + 15.00 = 83.79, divided by the 90% coverage gives 93.1. No other model in the top seven has been measured on more than two boards, and only Gemini 3.1 Pro Preview shares Opus 5’s 90% coverage anywhere in the field. Opus 5 is also the coding leader at #1 — the only top-five model whose coding rank comes from a measured SWE-bench component rather than from the LiveBench proxy.

What the #1 ranking does not yet prove

Neon’s #1 ranking does not prove Neon is the best AI model. It proves Neon is the best model on LiveBench, on the November 20, 2026 reading, by 2.5 raw points over the next two models — a promising finding, not a verdict. A second component could move it either way: a GPQA Diamond score at the field median (normalised 77.5) would put the League Score at 93.9, while a second field maximum would hold it at 100.0 on a deeper record. The same caution applies in the other direction: Claude Opus 5.5 sits sixth at 92.7 on one component, and Gemini 4 Carbon seventh at 92.5 on one component — both could overtake Neon on a strong second score, and both could fall. Until the second components land, the honest statement is narrower than the headline: Neon leads the only board that has measured it, and leads it by the widest margin in the field. Read the gap honestly: the ranking is a snapshot of one board, not a judgement of the models.

The price of the top five

The top five priced on the standard workload — 10M input + 2M output tokens — and on the cached-heavy variant, where all 10M input tokens are cache reads
Model$/1M in / outStandard workloadCached-heavyLeague ScoreCoverage
Gemini 4 Neon$3.00 / $15.00$60.00$31.50100.040%
GPT-6.2 Astra$8.00 / $40.00$160.00$84.0096.755%
Claude Fable 5.2$8.00 / $40.00$160.00$82.0095.540%
GPT-6 Astra$10.00 / $50.00$200.00$110.0094.855%
Claude Opus 5$5.00 / $25.00$100.00$55.0093.190%

The cheapest model at the top is also the cheapest by a wide margin: Gemini 4 Carbon costs $40.00 on the standard workload, a quarter of GPT-6.2 Astra’s and Claude Fable 5.2’s $160.00 and a fifth of GPT-6 Astra’s $200.00, for a League Score 4.2 points below Neon. Neon itself is the value story of the top three: $60.00 against $160.00 for the two models immediately below it, and its cached input rate of $0.15 makes the cached-heavy workload $31.50 — cheaper than Claude Opus 5’s $55.00 at a higher League Score. The full value ranking, with the whole field priced, lives in our best value AI models analysis, and every model’s rates are on the pricing index.

The score to watch next is Neon’s second component. A GPQA Diamond or SWE-bench Verified result is the test that turns a 100.0 at 40% coverage into a finding with depth — or revises it down. Until then, the best AI model in November 2026 is Gemini 4 Neon on the board that has measured it, Claude Opus 5 on the evidence that has been gathered, and Gemini 4 Carbon on the bill. The models’ full records live on their pages: Gemini 4 Neon, GPT-6.2 Astra and Claude Opus 5.