Qwen4-Max LiveBench score: 73.3 normalises to a League 77.9.
One day after Qwen4-Max entered our database unranked, its first tracked score has landed: 73.3 on LiveBench 2.0, which normalises to a League Score of 77.9 and overall #16 — on a single component, 40% coverage. This post works the normalisation maths in full, states the coverage caveat against Qwen3.8-Max’s 68.5 at 65% coverage, and prices the now-scored model on the standard and cached-heavy workloads.
By AI Pro League · Published

The Qwen4-Max LiveBench score is 73.3. Our November 17, 2026 changelog records it as the model’s first result on a tracked leaderboard: 73.3 per LiveBench 2.0 (livebench.ai via BenchLeader, results dated 2026-11-06; read today), ranking #11 on LiveBench itself. Yesterday’s Qwen4-Max release analysis recorded the model entering our database of 64 models unranked; today it joins the 52 scored models in the AI model rankings. The BenchLM GPQA leaderboard (updated 2026-11-16) still carries no Qwen4-Max row, so LiveBench is its only tracked score. What follows is computed from the database at build time.
The Qwen4-Max 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: 35 models in our database carry a LiveBench component, running from 32.4 (Gemma 4 31B) to 84.9 (Claude Fable 5.2), a span of 52.5 points. Worked step by step for Qwen4-Max: (73.3 − 32.4) ÷ 52.5 × 100 = 77.9. Because LiveBench is the only component Qwen4-Max has, that normalised figure is also its League Score: 77.9, resting on 40% coverage — one component of the five the League scores. It enters at #16 overall and #12 on value, at a blended $2.25 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.
Where 73.3 sits in the LiveBench field
| Model | LiveBench raw | Normalised | League Score | Coverage |
|---|---|---|---|---|
| GPT-5.6 Sol | 76.3 | 83.6 | 88.8 | 70% |
| Qwen4-Max (new) | 73.3 | 77.9 | 77.9 | 40% |
| Claude Sonnet 5 | 73.0 | 77.3 | 78.9 | 75% |
In the raw ordering of our tracked field, 73.3 sits just above Claude Sonnet 5 at 73.0 and below GPT-5.6 Sol at 76.3 — credible frontier company for a first measurement, and a long way above the predecessor’s raw mark, which is where the comparison has to be read carefully.
Against Qwen3.8-Max: one component beats 68.5 — on less evidence
| Model | LiveBench raw | Normalised | League Score | Coverage | Overall rank | Workload cost | Cached-heavy cost |
|---|---|---|---|---|---|---|---|
| Qwen4-Max | 73.3 | 77.9 | 77.9 | 40% | #16 | $24.00 | $12.60 |
| Qwen3.8-Max | 60.9 | 54.3 | 68.5 | 65% | #29 | $32.00 | $18.00 |
Read the gap honestly. On LiveBench alone the generational step is large: a raw 60.9 (normalised 54.3) becomes 73.3 (normalised 77.9), and Qwen4-Max’s single component, 77.9, already sits above Qwen3.8-Max’s overall League Score of 68.5. The caveat is coverage, and it is explicit: the predecessor’s 68.5 rests on 65% coverage — LiveBench plus GPQA Diamond and MMLU-Pro components — while Qwen4-Max’s 77.9 rests on 40%. A score at 40% coverage is a promising rumour next to a finding at 65%: it is one component, measured once, and the components not yet measured are exactly the ones that could move it. Our score coverage explainer sets out why we print the coverage beside every score on this site.
What the scored model costs on a workload
The price record from yesterday’s release post is unchanged — $1.50 / $4.50 per million tokens, cached input $0.075, checked Nov 16, 2026 — so the scored model is also the cheaper one. On the standard workload Qwen4-Max costs $24.00 — 10 × $1.50 + 2 × $4.50 — against $32.00 for its predecessor. On the cached-heavy run it costs $12.60 — 8 × $0.075 + 2 × $1.50 + 2 × $4.50 — against $18.00. A 77.9 at 40% coverage for $24.00 a workload, against a 68.5 at 65% coverage for $32.00: that is the whole comparison, and the coverage figures are part of it, not footnotes to it. 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 Qwen4-Max 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. Qwen3.8-Max shows the shape of that record: its 65% coverage is LiveBench’s 40% plus GPQA Diamond’s 15% and MMLU-Pro’s 10%. Until a second board publishes, the honest summary of Qwen4-Max is the narrow one: an excellent first component, an unfinished record.
What the first score 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 Qwen4-Max, and a model can be strong on LiveBench and ordinary elsewhere.
- The predecessor comparison cuts both ways. Qwen4-Max’s single normalised component (77.9) beats Qwen3.8-Max’s overall League Score (68.5), but the 68.5 is corroborated across three components at 65% coverage and the 77.9 is not corroborated at all yet.
- Normalisation moves with the field. The 77.9 is computed against today’s field (32.4–84.9); a re-measurement of the field, or a new floor or ceiling, re-scales every LiveBench component, including this one.
- The GPQA gap is a gap, not a zero. The BenchLM GPQA leaderboard carrying no Qwen4-Max row (updated 2026-11-16) 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 Qwen4-Max next will decide whether 77.9 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 Qwen4-Max page, the predecessor’s on the Qwen3.8-Max page, and the release-day record in yesterday’s Qwen4-Max release analysis.