Qwen4-Max costs $1.50 in and $4.50 out, and a task costs 67% less than on Qwen3.8-Max. Gemini 4 Hydrogen still finishes one for less.
Alibaba's new flagship is a real upgrade on its own family. Against everyone else, the cheap output price runs into a model that likes to talk.

Photo by Jeffrey Zhang on Unsplash
Qwen4-Max went on sale overnight, and OpenRouter listed it at 02:18 UTC on November 16. Alibaba's Model Studio price list puts it at $1.50 per million input tokens and $4.50 per million output on the international deployment. That's 25% under Qwen3.8-Max's $2 and $6, and cache hits fell further, from $0.25 to $0.075.
Artificial Analysis had it benchmarked by morning. It scores 43 on the Intelligence Index, nine points above Qwen3.8-Max, and a task costs $3.27 where 3.8 cost $10.02. If you run Qwen3.8-Max today, switch. That part is easy.
The harder question is whether it beats the other $1-to-$2 models, and on a per-task basis it mostly doesn't. Gemini 4 Hydrogen charges the same for input, 67% more for output, and still finishes AA's tasks for 11 cents less with an 11-point higher score.
What Alibaba charges, region by region
| Per 1M tokens | Singapore (International) | Global and Beijing |
|---|---|---|
| Input | $1.50 | $1.413 |
| Cache read (implicit or explicit) | $0.075 | $0.071 |
| Cache write, 5 minutes | $1.875 | $1.767 |
| Cache write, 1 hour | $3.00 | $2.827 |
| Output (thinking included) | $4.50 | $4.24 |
| Free quota | 1M tokens, 90 days | None |
Alibaba Cloud Model Studio pricing and QwenCloud, read November 16, 2026. "Global" covers Hong Kong, Frankfurt, US Virginia and Tokyo.
One price runs from the first token to the millionth, and thinking mode bills the same as non-thinking, so there's no long-context surcharge to plan around. The Global deployment is 5% to 6% cheaper on every line, so if your traffic doesn't need to land in Singapore, it's free money.
Cached input at $0.075 is 5% of the input price, the same rate and the same discount Gemini 4 Hydrogen charges. Implicit caching is on by default, so repeated prefixes get the discount without you writing any cache code. Explicit writes cost 1.25x input for five minutes and 2x for an hour, the same multipliers Anthropic uses.
No, it isn't $0.075 per million
A dev.to post calls $0.075 the input price and says output rates are "undetailed". Both are wrong. $0.075 is the cache-read rate. A prompt Qwen4-Max hasn't seen costs twenty times that, and output is on Alibaba's price page at $4.50.
If you budgeted off that headline, multiply your input line by 20 for any traffic that misses the cache, then add output.
Against Qwen3.8-Max: a clean win
The card is 25% cheaper, but the bigger saving is in tokens. Alibaba's launch post says Qwen4-Max uses about 35% fewer tokens on its Work360 agent test at high effort and about 49% fewer on Terminal-Bench 4.0. Combine that with the lower rates and Alibaba gets average cost per Work360 task down from $4.43 to $1.67, a 62% cut, while the score rises from 47.74 to 52.69.
We usually discount vendor cost claims. This one held up. On Artificial Analysis, Qwen3.8-Max (0902) costs $10.02 a task and Qwen4-Max costs $3.27, a 67% drop that's slightly bigger than Alibaba claimed, with the index score up from 34 to 43.
Qwen3.8-Max hasn't been deprecated and its price hasn't changed. There's no reason to stay on it unless you depend on its batch endpoint: OpenRouter sells 3.8 at $1.15/$3.45 in batch mode, and Qwen4-Max has no batch price yet.
Against everyone else: verbose for its price
| Model | AA effort | Index | $ per task | $ per point |
|---|---|---|---|---|
| Qwen4-Max | not stated | 43 | $3.27 | $0.076 |
| Qwen3.8-Max (0902) | default | 34 | $10.02 | $0.295 |
| Kimi K3.1 Pro | default | 46 | $3.68 | $0.080 |
| Gemini 4 Hydrogen | xhigh | 54 | $3.16 | $0.059 |
| GPT-6.1 Terra | max | 51 | $0.96 | $0.019 |
| GPT-6.1 Sol | max | 55 | $1.27 | $0.023 |
| GLM-5.3 | max | 32 | $5.01 | $0.157 |
| Mistral Large 4 | high | 26 | $2.06 | $0.079 |
Artificial Analysis Intelligence Index v5, read November 16, 2026. Dollars per point is our division.
AA labels Qwen4-Max "very verbose": it wrote 127 million output tokens across the index, against a median of 78.6 million. That's where the cheap $4.50 output price goes. It writes about 60% more than a typical model, so the per-token saving gets eaten on the way to an answer.
GPT-6.1 Terra is the uncomfortable comparison. It costs $0.96 a task for a 51, so each index point costs about a quarter of what Qwen4-Max charges. Gemini 4 Hydrogen at xhigh lands 11 points higher for slightly less money. Only Kimi K3.1 Pro, at $3.68 for a 46, sits close on price and score.
One caveat in Qwen's favour: AA doesn't say which effort setting it ran. The API defaults to high and accepts low and xhigh. Alibaba's own chart shows low effort using roughly a quarter fewer tokens than xhigh on Work360, so a low-effort run might cost less and score less. We'll update this when AA publishes per-effort results.
Same tokens, seven rate cards
Strip out verbosity and compare rate cards on equal token counts: 50K of context, 45K of it a cache hit, and 2K tokens back.
| Model | Per turn | 100K turns |
|---|---|---|
| Qwen4-Max (Singapore) | $0.0199 | $1,988 |
| Qwen4-Max (Global, Beijing) | $0.0187 | $1,874 |
| GPT-6.1 Terra | $0.0173 | $1,725 |
| Gemini 4 Hydrogen | $0.0259 | $2,588 |
| Qwen3.8-Max | $0.0333 | $3,325 |
| Claude Sonnet 5.5 | $0.0345 | $3,450 |
| Kimi K3.1 Pro | $0.0390 | $3,900 |
Our arithmetic from each provider's list price. Same tokens on every row, so this prices the cards, not the models.
On identical tokens, Qwen4-Max is 40% cheaper than Qwen3.8-Max, 23% cheaper than Hydrogen, and 42% and 49% cheaper than Sonnet 5.5 and Kimi K3.1 Pro. Only Terra undercuts it, by 13%, because Terra's input and cache rates are lower. Qwen4-Max wins on paper. It loses ground once it starts thinking at length.
Where Alibaba's own chart has it winning
The launch chart compares Qwen4-Max with Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1. No Gemini model appears. The results split cleanly by task type.
| Benchmark | Qwen4-Max | Opus 5.5 | GPT-6.1 Sol | Kimi K3.1 |
|---|---|---|---|---|
| DeepSWE v1.1 | 75.7 | 74.2 | 75.2 | 75.5 |
| Automationbench-AA | 72.5 | 69.5 | 64.9 | 73.3 |
| BioMysteryBench | 85.0 | 81.9 | 84.4 | 80.6 |
| Terminal-Bench 4.0 | 48.5 | 59.6 | 56.1 | 67.2 |
| ProgramBench | 31.5 | 65.0 | 52.5 | 25.0 |
| Humanity's Last Exam | 51.7 | 59.8 | 54.0 | 51.1 |
Vendor-reported, from the Qwen4-Max launch post. Six of the 16 rows shown.
On repository bug-fixing and workflow automation, Qwen4-Max is level with models that charge up to four times as much for output. On DeepSWE, the five current models in Alibaba's chart sit within 1.6 points, which says more about the benchmark than about any of them.
Terminal work and building programs from scratch are different. Qwen4-Max trails Opus 5.5 by 11 points on Terminal-Bench 4.0 and scores less than half of Opus's 65.0 on ProgramBench. If your agent lives in a shell, the cheaper tokens won't make up for that.
Five things the price page doesn't tell you
- Max output is unclear. The model page says 131,072 tokens without thinking and 262,144 for thinking plus answer. QwenCloud and OpenRouter say 393,216. Test before you rely on the bigger number.
- Thinking is on by default at high effort. Set
enable_thinking: falseorreasoning_effort: "low"on short calls if you want to cut down the verbosity. - The preview model was only ever on Token Plan, not pay-as-you-go, and ends at midnight UTC+8 on November 25. We'd move anything on
qwen4-max-previewtoqwen4-maxbefore then. - The weights are closed. Alibaba says smaller models in the family will be open-sourced, with no date.
- Rate limit is 5M tokens per minute on QwenCloud, with no published request limit. OpenRouter has one provider, Alibaba, at list price.
Where it makes sense
Anyone on Qwen3.8-Max, today. It's cheaper, better and the same API.
Teams that send long, mostly cached contexts and get short answers back should also look. That's the workload where per-token price dominates and verbosity matters least, and Qwen4-Max's flat 1M card is among the cheapest at that size. Retrieval-heavy pipelines with a strict output format fit well.
For open-ended agent work we'd test GPT-6.1 Terra first, and Gemini 4 Hydrogen if you need the extra points. Hydrogen's price is fixed, as far as Google has said, but Gemini 4 Argon and Carbon double on December 1, so if Qwen4-Max is your fallback for those, it just got more attractive.
Compare Qwen4-Max with every other model on the pricing table, see its model page, or plug your own token counts into the calculator. Our Gemini 4 Hydrogen post and GPT-6.1 Terra post cover the two models that beat it per task, and the November price changes roundup has the December 1 increases.
Sources
- Alibaba Cloud Model Studio pricing - per-region rates, cache write and read prices, free quota
- Qwen4-Max model page - context, max input and output, thinking defaults, regions
- Qwen4-Max preview page - Token Plan only, November 25 end date
- QwenCloud model page - Singapore rates, rate limit
- Qwen4-Max launch post - parameters, benchmark chart, Work360 cost per task, token reductions
- Artificial Analysis - index scores, cost per task and output tokens for every model in the comparison
- OpenRouter - listing, providers, max completion, Qwen3.8-Max batch price