Gemini 4 Neon LiveBench score: 87.4 normalises to a League 100.0.
Two days after Gemini 4 Neon entered our database unranked, its first tracked score has landed: 87.4 on LiveBench 2.0, the new maximum of the tracked field, which normalises to a League Score of 100.0 and overall #1 — on a single component, 40% coverage, past a #2 whose score rests on two. This post works the normalisation in full, records the six scores the same re-read revised down, and prices the new #1 on the standard workload.
By AI Pro League · Published

The Gemini 4 Neon LiveBench score is 87.4. Our November 20, 2026 changelog records it as the model’s first result on a tracked leaderboard: 87.4 per LiveBench 2.0 (livebench.ai via BenchLeader, results dated 2026-11-06; read today), ranking #1 on LiveBench itself, 2.5 points ahead of the 84.9 tie between GPT-6.2 Astra and Claude Fable 5.2, as the BenchLeader table grows from 37 to 38 configurations. Wednesday’s Gemini 4 Neon release analysis recorded the model entering our database of 68 models unranked; today it joins the 56 scored models at the top of the AI model rankings. The BenchLM GPQA leaderboard (updated 2026-11-19) still carries no Neon row, so LiveBench is its only tracked score. What follows is computed from the database at build time.
The Gemini 4 Neon LiveBench score, worked in full
LiveBench carries 40% of the League Score on this site — more than any other single component — and a raw LiveBench score becomes a component by normalisation against the tracked field: 38 models in our database carry a LiveBench component, running from 32.4 (Gemma 4 31B) to 87.4 (Gemini 4 Neon), a span of 55.0 points. Worked step by step for Neon: (87.4 − 32.4) ÷ 55.0 × 100 = 100.0. Because LiveBench is the only component Neon has, that normalised figure is also its League Score: 100.0, resting on 40% coverage — one component of the five the League scores. It enters at #1 overall and #27 on value, at a blended $6.00 per million tokens. How the board itself works is set out in our LiveBench explainer, and the standings themselves live on the current LiveBench leaderboard.
One component at #1, two components at #2
| Model | LiveBench raw | Normalised | Normalised before today | League Score | Overall rank | Coverage | Components |
|---|---|---|---|---|---|---|---|
| Gemini 4 Neon (new) | 87.4 | 100.0 | 104.8 | 100.0 | #1 | 40% | 1 |
| GPT-6.2 Astra | 84.9 | 95.5 | 100.0 | 96.7 | #2 | 55% | 2 |
| Claude Fable 5.2 | 84.9 | 95.5 | 100.0 | 95.5 | #3 | 40% | 1 |
Read the top of the table carefully. Neon’s 100.0 rests on 1 component at 40% coverage. The model it passes, GPT-6.2 Astra, scores 96.7 at 55% coverage on 2 components — its GPQA Diamond component still normalises to 100.0, which is what cushions its fall. The new #1 therefore carries less evidence than the #2 behind it: a promising rumour, however large, next to a finding. Claude Fable 5.2, the other 100.0 of Nov 18, falls to 95.5 and #3 with no new measurement of its own — its raw 84.9 is unchanged, and only the field around it moved. Our score coverage explainer sets out why we print the coverage beside every score on this site.
A new ceiling re-scales the field — and six scores were revised
Normalisation is relative, so a new maximum is never only the new model’s news. Before today the field ran to 84.9, a span of 52.5; Neon stretches the span to 55.0, and every other LiveBench component is re-scaled against it — Claude Fable 5.2’s component falls from 100.0 to 95.5 on an unchanged raw, and the same compression runs down the whole table above. Separately, the BenchLeader re-read that added Neon also revised six tracked scores down by 0.2 each; our changelog records every one, and the League consequences are larger than 0.2 suggests, because each revised raw is then normalised against the wider span:
| Model | LiveBench raw before | LiveBench raw now | Normalised before | Normalised now | League Score now | Overall rank now |
|---|---|---|---|---|---|---|
| GPT-6 Astra | 84.5 | 84.3 | 99.2 | 94.4 | 94.8 | #4 |
| Gemini 4 Carbon | 83.5 | 83.3 | 97.3 | 92.5 | 92.5 | #7 |
| GPT-6.1 Sol | 81.4 | 81.2 | 93.3 | 88.7 | 88.7 | #10 |
| GPT-6.1 Terra | 78.8 | 78.6 | 88.4 | 84.0 | 84.0 | #12 |
| Claude Sonnet 5.5 | 77.0 | 76.8 | 85.0 | 80.7 | 80.7 | #16 |
| Qwen4-Max | 73.3 | 73.1 | 77.9 | 74.0 | 74.0 | #19 |
Gemini 4 Carbon takes the largest League fall of the six: its raw step of 0.2 becomes a League move from 97.3 to 92.5 once the wider span is applied. Every other tracked LiveBench score in the re-read matches the database exactly — the changelog says so, and the six above are the whole of the movement.
The family ordering, on a tracked board
| Model | LiveBench raw | Normalised | League Score | Overall rank | Workload cost |
|---|---|---|---|---|---|
| Gemini 4 Neon (new #1) | 87.4 | 100.0 | 100.0 | #1 | $60.00 |
| Gemini 4 Carbon | 83.3 | 92.5 | 92.5 | #7 | $40.00 |
| Gemini 4 Argon | 77.7 | 82.4 | 82.4 | #14 | $40.00 |
| Gemini 4 Hydrogen | 77.0 | 81.1 | 81.1 | #15 | $30.00 |
Google’s own launch coverage compares Neon with rival flagships, not with its siblings — the vendor publishes no Neon-versus-Carbon score. On this tracked board the family ordering is now measured: the adapted mid-tier, Neon at 87.4, sits above the flagship it was adapted from, Gemini 4 Carbon at 83.3, with Gemini 4 Argon and Gemini 4 Hydrogen below. One board is not the family’s whole record — Carbon’s case rests on boards this site does not track — but on the one independent board all four have faced, the middle tier leads it.
What the new #1 costs on a workload
The price record from the release post is unchanged — $3.00 / $15.00 per million tokens, cached input $0.15, checked Nov 19, 2026 — so the new #1 is also the cheaper #1. On the standard workload Neon costs $60.00 — 10 × $3.00 + 2 × $15.00 — against $160.00 for GPT-6.2 Astra and $160.00 for Claude Fable 5.2. On the cached-heavy run it costs $37.20 — 8 × $0.15 + 2 × $3.00 + 2 × $15.00 — against $99.20 and $97.60. Within its own family the ordering inverts: Gemini 4 Carbon runs the same standard workload for $40.00 at its introductory rate, so Neon’s #1 costs half again as much per workload as the sibling it outscores — until Dec 1, when Carbon’s standard rate doubles its workload cost and Neon becomes the cheaper of the two. The full rate cards sit in our AI model pricing index.
What a second tracked score would change
Four of the five components the League scores still have no Neon result: GPQA Diamond, MMLU-Pro, SWE-bench Verified and LMArena. The first of them to publish does two things at once — it raises coverage above 40%, and it re-weights the League Score away from a single normalisation, in whichever direction the new component points. GPT-6.2 Astra shows the shape of that record at #2: LiveBench’s 40% plus GPQA Diamond’s 15% is the deeper 55% record Neon now leads on less. Until a second board publishes, the honest summary of Neon is the narrow one: the field’s best first component, an unfinished record.
What the #1 ranking does not yet prove
- It is one component. LiveBench carries 40% of the League Score; the other 60% — reasoning, knowledge, coding and crowd-preference components — is unmeasured for Neon, and a model can be strong on LiveBench and ordinary elsewhere.
- The #2 carries more evidence. GPT-6.2 Astra’s 96.7 rests on two components at 55% coverage; Neon’s 100.0 rests on one at 40%. The ranking orders the scores, not the depth of the records behind them.
- Normalisation moves with the field. The 100.0 is computed against today’s field (32.4–87.4); the same re-read that produced it re-scaled every other component and revised six raws — a future re-measurement can move this score without Neon being re-measured at all, exactly as Fable 5.2’s fell today.
- The GPQA gap is a gap, not a zero. The BenchLM GPQA leaderboard carrying no Neon row (updated 2026-11-19) means no tracked result exists — it says nothing about how the model would score there.
The score to watch next is the second one: whichever tracked leaderboard publishes on Neon next will decide whether 100.0 at 40% coverage is the start of a record or a single component’s high-water mark. The model’s full record lives on its Gemini 4 Neon page, the release-day record in Wednesday’s Gemini 4 Neon release analysis, and the family table on the Gemini 4 Carbon page.