Gemini 4 Carbon gets its first tracked score: LiveBench 83.5, League #3.
Gemini 4 Carbon went generally available on November 10, 2026, and today our database records its first result on a tracked leaderboard — 83.5 on LiveBench 2.0, #4 on that board — which normalises to a League Score of 97.3 and overall #3. This post works the normalisation maths in full, states the 40% coverage the score rests on, and prices what the December 1 rate step does to a standard workload.
By AI Pro League · Published

Gemini 4 Carbon has a League Score. Our November 12, 2026 changelog records its first result on a tracked leaderboard: 83.5 on LiveBench 2.0 (results dated 2026-11-06; read today), ranking #4 on LiveBench itself. The November 11, 2026 changelog records the step before it: Carbon went generally available on November 10, 2026 per Google’s official announcement (blog.google, read November 11), ending the Fairwind-only preview that began Oct 28, 2026, with API access via Google AI Studio, the Gemini API and Vertex AI. What follows is computed from the database at build time.
Gemini 4 Carbon: the record, dated and priced
| Measure | Gemini 4 Carbon |
|---|---|
| First release (Fairwind preview) | Oct 28, 2026 |
| Generally available | Nov 10, 2026 |
| LiveBench raw score (LiveBench 2.0) | 83.5 (#4 of 34) |
| League Score / overall rank | 97.3 / #3 of 51 scored models |
| Score coverage | 40% — one component (LiveBench) |
| Input / output per 1M tokens (introductory) | $2.00 / $10.00 |
| Cached input per 1M tokens | $0.10 |
| From Dec 1, 2026 (per Google’s pricing page) | $4.00 / $20.00, cached $0.20 |
| Context window | 1,048,576 tokens |
| Inputs | Text, image, audio, video, PDF |
Two details in that record were corrected at GA, per the same November 11 changelog entry: the context window is 1,048,576 tokens, and PDF joins the recorded inputs. Carbon entered this database unranked — as Muse Ember and Inkling 2 Small did — because no tracked leaderboard had published for it. That wait is now over, on exactly one leaderboard.
How 83.5 becomes a League Score of 97.3
Carbon’s League Score rests on this single component, so it is checkable by hand. Normalisation stretches the recorded LiveBench field — 34 models, raw scores from 32.4 to 84.9, median 68.9 — across a 0–100 scale: (83.5 − 32.4) ÷ 52.5 × 100 = 97.3. With one recorded component at 40% weight, the League Score is that normalised value: 97.3. LiveBench carries 40% of the League Score — the joint-largest weight of the five components — for the reasons set out in our LiveBench explainer; the full field sits on the LiveBench leaderboard.
The top of the League now — and why #4 on LiveBench is #3 here
| Rank | Model | League Score | Coverage |
|---|---|---|---|
| 1 | Claude Fable 5.2 | 100.0 | 40% |
| 2 | GPT-6 Astra | 99.4 | 55% |
| 3 | Gemini 4 Carbon (new) | 97.3 | 40% |
| 4 | Claude Opus 5.5 | 97.1 | 40% |
| 5 | Claude Fable 5.1 | 96.0 | 55% |
| 6 | Claude Opus 5 | 95.6 | 90% |
Carbon is #4 on LiveBench raw but #3 in the League, and the gap between those two ranks is the composite working as designed. Claude Fable 5.1 posts the higher raw LiveBench score — 84.0 against Carbon’s 83.5 — but its second component, a GPQA Diamond record, normalises below its LiveBench component and pulls its composite down to 96.0 at 55% coverage. Carbon, with nothing to pull it down yet, sits above it. Above Carbon, GPT-6 Astra holds #2 at 99.4 on a deeper 55% record, and Claude Fable 5.2 leads at 100.0 — the release priced in our Claude Fable 5.2 analysis. The standings themselves live in the AI model rankings.
What the December 1 price step does to the workload maths
| Model | League Score | Workload cost |
|---|---|---|
| Gemini 4 Carbon (to Nov 30) | 97.3 | $40.00 |
| Gemini 4 Carbon (from Dec 1) | 97.3 | $80.00 |
| Claude Opus 5.5 | 97.1 | $80.00 |
| Claude Fable 5.2 | 100.0 | $160.00 |
| GPT-6 Astra | 99.4 | $200.00 |
The step doubles every rate — input, output and cached — so the standard workload doubles with it: $40.00 becomes $80.00 on December 1, 2026. Where that lands matters more than the doubling itself: from December, Carbon’s workload costs exactly what Claude Opus 5.5 costs — $80.00 — and the two models sit 0.2 League points apart, while the two models above Carbon cost $160.00 and $200.00 on the same tokens. Cached input softens both versions of the bill: at $0.10 per million, cached tokens cost 95% less than fresh input, so a cached-heavy run of 8M cached + 2M fresh input + 2M output costs $24.80 at the introductory rates and $49.60 from December — the pattern priced across the field in our prompt caching pricing analysis. The full rate card sits in our AI model pricing index.
What the #3 ranking does not yet prove
- The record is one benchmark deep. Coverage is 40% — LiveBench only. The BenchLM GPQA leaderboard (updated 2026-11-11) still carries no Carbon row, and no SWE-bench Verified, MMLU-Pro or LMArena result is recorded either, so LiveBench is Carbon’s only tracked score — the standing caveat in our score coverage explainer. A promising rumour vs a finding: 97.3 is a finding about one benchmark, not yet about the other four.
- The margin below is 0.2 points. Claude Opus 5.5 sits at 97.1 on the same single-component, 40% shape of record. One LiveBench re-measurement — this database revised more than twenty LiveBench rows under LiveBench 2.0 on November 10 — can swap #3 and #4 without either model changing.
- The introductory price is a 20-day window. Every cost above halves or doubles on December 1. A budget set on November’s $40.00 workload is a $80.00 budget three weeks later, at unchanged token counts.
None of that dims the score itself: on the freshest general benchmark we track, a model that was unranked yesterday enters at #3 — and until December it does so at half the workload cost of the model 0.2 points behind it. The score to watch is the second one: whichever tracked leaderboard reports on Carbon next will test whether 97.3 is a ceiling or a starting point. The model’s full record lives on its Gemini 4 Carbon page.