Qwen4-Max released: a quarter off Qwen3.8-Max’s price — and no League Score yet.
Qwen4-Max, the first model of Alibaba’s Qwen4 generation, became generally available today, November 16, 2026, at $1.50/$4.50 per million tokens — both headline rates exactly 25% below Qwen3.8-Max, with cached input cut far deeper. It enters our database unranked: no tracked League benchmark has published for it, and the headline 43 in the launch coverage is an index the League does not score. This post sets out the dated record, works the price maths in full, and names the bar it has to clear.
By AI Pro League · Published

Qwen4-Max is the newest entry in our model database — one of 64 models we track. Our November 16, 2026 changelog records the release from Alibaba’s official Model Studio pricing page (Alibaba Cloud Model Studio, read today): released Nov 16, 2026 at $1.50 input / $4.50 output per million tokens, cached input $0.075. It is the Qwen4 flagship — a 2.4-trillion-parameter mixture-of-experts model per Alibaba’s launch post — and it arrives early against its own preview: the Token Plan preview that opened on November 11, 2026 was due to run until its retirement on November 25, and the generally available model, qwen4-max on Model Studio’s pay-as-you-go API, has shipped nine days before that date. Preview-era guidance that the model has no per-token price is already stale. What follows is computed from the database at build time.
Qwen4-Max: the release record, dated and sourced
| Measure | Qwen4-Max |
|---|---|
| Release date | Nov 16, 2026 |
| Model ID / access | qwen4-max, Alibaba Cloud Model Studio pay-as-you-go |
| Input / output per 1M tokens | $1.50 / $4.50 |
| Cached input per 1M tokens | $0.075 |
| Blended rate (3:1 input:output) | $2.25 |
| Context window | 1,000,000 tokens |
| Max output | 131,072 tokens without thinking; 262,144 for thinking plus the answer |
| Inputs | Text, image, video; thinking on by default |
| Weights | Not open weight — proprietary; smaller Qwen4 models promised open weights, no date |
| League Score | Unranked — no tracked benchmark yet |
The record’s shape is a flagship’s: the same 1M-token window as its predecessor, text, image and video inputs, and thinking on by default — served through the cloud, not as downloadable weights. The full rate card sits in our AI model pricing index, and the launch closes the question our Qwen 4 release date analysis of November 8 left open: the series has a date now because it has a model.
The maths: 25% off the headline, 95% off the cache
| Model | Input / output per 1M | Cached per 1M | Blended | Workload cost | Cached-heavy cost | League Score |
|---|---|---|---|---|---|---|
| Qwen4-Max | $1.50 / $4.50 | $0.075 | $2.25 | $24.00 | $12.60 | Unranked |
| Qwen3.8-Max | $2.00 / $6.00 | $0.25 | $3.00 | $32.00 | $18.00 | 68.5 (65% coverage) |
Worked in full: the headline cut is ($2.00 − $1.50) ÷ $2.00 = 25% on input, and the same 25% on output. On the standard workload the new flagship costs $24.00 — 10 × $1.50 + 2 × $4.50 — against $32.00 for Qwen3.8-Max, a saving of $8.00 on identical tokens; the blended rates tell the same story, $2.25 against $3.00. The cached rate is cut deeper still: $0.075 against $0.25 is a 70% cut in the rate itself, and measured as a discount off each model’s own input rate it deepens from 87.5% to 95.0%. On the cached-heavy run agent loops actually execute, Qwen4-Max costs $12.60 — 8 × $0.075 + 2 × $1.50 + 2 × $4.50 — against $18.00, 30% less. The field-wide pattern behind that gap is priced in our prompt caching pricing analysis.
The bar: what Qwen3.8-Max holds in the League
A cheaper successor is only a better one when it matches the record it replaces, and that record is specific. Qwen3.8-Max, released Aug 3, 2026, carries a League Score of 68.5 — #28 overall — on 65% coverage: three of the five components the League scores, their weights summing to the coverage figure itself.
| Component | Raw | Normalised | League weight | Source |
|---|---|---|---|---|
| LiveBench overall | 60.9 | 54.3 | 40% | LiveBench, measured Nov 10, 2026 |
| GPQA Diamond | 92.6 | 85.2 | 15% | BenchLM GPQA leaderboard, measured Oct 14, 2026 |
| MMLU-Pro | 88.6 | 100.0 | 10% | MMLU-Pro, measured Oct 14, 2026 |
Each component has its own explainer — LiveBench, which carries 40% of the League Score, GPQA Diamond and MMLU-Pro, where Qwen3.8-Max’s 88.6 is the highest raw score in that component’s field. That is the bar, stated exactly: when a tracked leaderboard publishes on Qwen4-Max, its normalised components will be weighed against a 68.5 at 65% coverage — not against the launch headlines.
Why Qwen4-Max has no League Score yet
Our changelog records no result for Qwen4-Max on any of the five benchmarks the League scores — no LiveBench, SWE-bench Verified, GPQA Diamond, MMLU-Pro or LMArena figure from a tracked leaderboard source. The headline 43 in the launch coverage is the Artificial Analysis Intelligence Index, which the League does not score, so it cannot be converted into a League Score without inventing a mapping — and we do not invent mappings. The model therefore enters unranked at 0% coverage, exactly as our score coverage explainer describes for launch-day models. It is not alone: 13 of the 64 models we track currently carry no League Score, against 51 scored models in the AI model rankings. Muse Ember entered the same way last week, on the same headline index.
What the unranked launch does not yet prove
- Every quality claim so far is vendor-published or off-League. Alibaba’s launch figures — its internal professional-work benchmark, its token-efficiency gains — are the maker’s own measurements, and the independent numbers circulating today sit on boards the League does not track. Independent re-measurement on a tracked leaderboard is what turns launch claims into a record.
- Cheaper is a price finding, not a capability one. The 25% headline cut and the 95.0% cached discount are exact, and they are all that is exact today: the predecessor is a known quantity at 68.5 on 65% coverage, while Qwen4-Max’s counter-offer is its rate card and its context window — not a measured score.
- The output ceiling bundles thinking with the answer. The 262,144-token maximum is thinking plus the final answer combined; without thinking, the cap is 131,072 — the same figure our database records as Qwen3.8-Max’s maximum output. Long-answer budgets should be planned on the split, not the headline.
- No self-hosted Qwen4 follows from this launch. Qwen4-Max itself is not open weight; Alibaba says smaller Qwen4 models will get open weights, with no date. A plan that needs downloadable weights today still runs on the previous generation.
None of that dims the release itself: the Qwen4 generation has opened with its flagship at a quarter below its predecessor’s price and the deepest cached discount in the family’s record. The score to watch is the first one — whichever tracked leaderboard publishes on Qwen4-Max first will decide whether the price cut is also a capability step. The model’s full record lives on its Qwen4-Max page, the predecessor’s on the Qwen3.8-Max page, and the lab’s history on the Alibaba Qwen profile.