Qwen 4 Max review
Model page: Qwen 4 Max API — pricing, specs and the playground
Qwen 4 Max is the first model of Alibaba’s Qwen 4 generation. It ran as a time-limited preview from 11 November 2026 and became generally available on 16 November under the model ID qwen4-max. Alibaba positions it for complex reasoning, coding and agent work, and published its launch numbers next to Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1.
This review gathers what is known a day after release: the specs and price, two independent benchmark suites, Alibaba’s own chart, what reviewers in English and Chinese found when they used it, and seven tasks we ran ourselves on Qwen 4 Max and four competitors with the same prompts.
We run apimodels.app, which sells Qwen 4 Max at Alibaba’s list price alongside the models it is compared with here. The benchmark scores are not ours; they come from the sources named in each table and linked at the end. The seven same-prompt tests are ours, run on 17 November 2026 through our own API with every model at its default thinking setting.
What Qwen 4 Max is: specs and price
The price is the first thing people ask about. Qwen 4 Max costs $1.50 per million input tokens and $4.50 per million output tokens on Alibaba’s international platform, with cached input at $0.075, and the rate does not rise for long prompts. That is a quarter below its predecessor Qwen3.8-Max ($2 / $6), under half of Claude Opus 5.5 ($4 / $20) and below GPT-6 Sol ($2 / $10).
Thinking is on by default and counts as output. You can switch it off per request, which matters for both speed and cost; more on that below. Alibaba’s two regional pages disagree on the output cap: the international page says 131K tokens (262K including thinking) and the mainland page says 393K. It is not open-weight; Alibaba has said smaller Qwen 4 models will be released openly later, without a date.
| Qwen 4 Max | |
|---|---|
| Maker | Alibaba (Qwen / Tongyi) |
| Released | Preview 11 Nov 2026; generally available 16 Nov 2026 |
| Model ID | qwen4-max |
| List price per 1M tokens | $1.50 input / $4.50 output; cached input $0.075 |
| Mainland China list price | ¥10 input / ¥30 output per 1M tokens |
| Context window | 1M tokens, one price across the whole window |
| Max output | 131K (262K with thinking) on the international page |
| Input types | Text, image, video |
| Thinking | On by default; can be switched off per request |
| Tools | Function calling, JSON Schema output, built-in web search |
| Open weights | No |
Independent benchmarks: mid-table overall, strong in finance and legal
Artificial Analysis scores Qwen 4 Max at 43 on its Intelligence Index, thirty-eighth of the 81 models it tracks and just above the median of 41. That is level with GPT-6 Sol (44) and well behind Claude Opus 5.5 (59) and Gemini 4 Carbon (60). The number drew a sharp reaction on Hacker News, where some readers expected more from a model of its size; others pointed out that Artificial Analysis has so far tested only one thinking setting.
Vals AI, which leans towards professional tasks, is kinder: 63.20% on the Vals Index, eleventh of 38, with its best placings in finance (eighth of 41 on both the Finance Index and Finance Agent) and on the legal agent benchmark HLAB (seventh of 39). Its Terminal-Bench 4.0 score, 48.48%, is fifteenth of 65.
One figure explains much of what follows: Qwen 4 Max produced 127 million output tokens to complete the Artificial Analysis index, where the median model produced 78.6 million. It thinks a lot.
| Measure | Run by | Qwen 4 Max | Rank |
|---|---|---|---|
| Intelligence Index v5.0.1 | Artificial Analysis | 43 | 38th of 81 |
| Cost per index task | Artificial Analysis | $3.27 | — |
| Output tokens to run the index | Artificial Analysis | 127M (median 78.6M) | 65th of 80 for brevity |
| Vals Index v2.2 | Vals AI | 63.20% | 11th of 38 |
| Finance Agent v2.1 | Vals AI | 66.68% | 8th of 41 |
| Legal Index | Vals AI | 70.71% | 17th of 39 |
| HLAB-Vals-1.1 (legal agent) | Vals AI | 21.25% | 7th of 39 |
| Vibe Code Bench v1.1 | Vals AI | 86.66% | 15th of 129 |
| Terminal-Bench 4.0 | Vals AI | 48.48% | 15th of 65 |
| Code Arena WebDev (Elo, preliminary) | Arena | 1750 | 6th |
Alibaba’s own chart against four rivals
Alibaba published sixteen benchmarks with four competitors alongside. These are vendor-reported numbers, but they are unusually candid: Qwen 4 Max is not the top score on most rows. Against Claude Opus 5.5 it leads on six of sixteen; against GPT-6.1 Sol on nine.
Read across the rows and a profile appears. On a single well-defined software issue it is level with everyone (DeepSWE 75.7). On legal agent work it is far ahead of all four (HLAB 21.3, the next best is 11.7). On agent benchmarks built around business tasks it is at or near the top. Then come the weak rows: ProgramBench, which asks for whole programs, at 31.5 against 65.0 for Opus 5.5; Terminal-Bench 4.0 at 48.5, the lowest of the five; and chart and video understanding, where it trails both OpenAI models and Opus.
Alibaba also reports that on its internal Work360 suite the cost per task fell from $4.43 on Qwen3.8-Max to $1.67, with a higher score (52.69 against 47.74). We have no way to check that figure.
| Benchmark (reported by Alibaba) | Qwen 4 Max | Claude Opus 5.5 | GPT-6.1 Sol | GPT-6 Astra | Kimi K3.1 |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 75.7 | 74.2 | 75.2 | 74.1 | 75.5 |
| HLAB-Vals-1.1 (legal) | 21.3 | 9.2 | 6.3 | 7.1 | 11.7 |
| Agents’ Last Exam v1.1 | 46.7 | 43.4 | 42.6 | 43.4 | — |
| BioMysteryBench-1.1 | 85.0 | 81.9 | 84.4 | 78.9 | 80.6 |
| AutomationBench-AA | 72.5 | 69.5 | 64.9 | 68.5 | 73.3 |
| APEX-Agents v1.1 | 72.1 | 73.5 | 60.0 | 64.7 | 65.5 |
| CodeArena (Elo) | 1750 | 1816 | 1742 | 1763 | 1686 |
| Humanity’s Last Exam | 51.7 | 59.8 | 54.0 | 53.6 | 51.1 |
| Terminal-Bench 4.0 | 48.5 | 59.6 | 56.1 | 59.1 | 67.2 |
| ProgramBench | 31.5 | 65.0 | 52.5 | 55.0 | 25.0 |
| Chartography | 54.2 | 66.3 | 67.8 | 71.1 | 50.0 |
| Video-Index | 68.1 | 74.8 | 73.8 | 76.7 | 63.2 |
What reviewers found when they used it
The reports agree on more than they disagree. Qwen 4 Max is careful and accurate on documents and well-scoped code, and it is slow. Most Chinese-language reviews tested the preview, so treat their timings as early.
On documents: ZPedia had it read a financial report in eight minutes with all six spot-checked figures correct, answer five needle-in-a-haystack questions over 1,600 support tickets without a miss, and clean a 44-row spreadsheet into 35 rows with real formulas in the summary cells. In the same run it left a debt item due within twelve months out of a prospectus summary and a product line out of a slide deck, and delivered a web page when asked for a plain SVG.
On code: Kingy AI’s head-to-head with Claude Opus 5.5 is the most rigorous so far. Qwen resolved 9 of 16 runs on real GitHub issues against 11 for Opus, passed every check on 18 simple tasks, and missed one small-repository task by raising the wrong exception type. It took a median 431 seconds of active time per issue against 86 for Opus, and cost $4.42 against $18.65. A Chinese reviewer who gave the same build task to seven coding tools found the preview the slowest at 75 minutes (the quickest took 8) while using the fewest tokens of the seven.
On visual taste: this was the most consistent complaint in China. Reviewers at Guixingren and Zhidongxi found its generated interfaces and 3D scenes functional but rough, with harsh lighting and coarse geometry, behind Kimi K3.1 Pro and well behind Claude Fable 5.2. On Hacker News, Simon Willison’s standing SVG test came out right only at the highest thinking setting, after 7 minutes 23 seconds and 26,295 output tokens.
| Who tested | What they ran | Result |
|---|---|---|
| Kingy AI | 8 SWE-bench Verified issues × 2, vs Opus 5.5 | 9 of 16 resolved (Opus 11); 431 s vs 86 s; $4.42 vs $18.65 |
| Kingy AI | 18 simple tasks | 48 of 48 checks, same as Opus 5.5 |
| ZPedia (preview) | 12 office and coding tasks | Report figures 6 of 6; ticket search 5 of 5; missed one debt item and one product line |
| MindStudio | Full-stack bug hunt and builds | Found the off-by-one bug; sketch-to-simulation in about 15 minutes |
| Ai Xuexi de Laozhang (preview) | One build task across 7 coding tools | Slowest at 75 minutes; fewest tokens used |
| Guixingren (preview) | Web and 3D generation | Code improved; lighting and geometry rough |
| Simon Willison, Hacker News | SVG drawing at four thinking levels | Correct only at the highest level, 7 min 23 s |
Our seven same-prompt tests against four rivals
We wrote seven tasks modelled on what developers most often send through our API: writing a function, fixing a buggy class, a support agent that calls tools and must follow a refund policy, questions over a 1,500-line production log, turning a Chinese group chat into JSON, three counting puzzles, and Chinese product copy with hard constraints. Each has an automatic check, 93 checks in total, verified against reference answers before we ran anything. We ran them on Qwen 4 Max, Claude Opus 5.5, Gemini 4 Carbon, GPT-6 Sol and GPT-6.1 Sol.
Qwen 4 Max got all 93 checks right. So did Opus 5.5 and both Sol models; Gemini 4 Carbon got 89. Qwen’s function passed all 50 hidden cases, it found all five bugs we planted in a cache class, it refunded only the defective keyboard and turned down a return that was 16 days past the policy window, it counted eight payment timeouts in the log and found the one request that crossed three services, and it worked out that “before Friday” meant 20 November and “next Wednesday” the 25th.
So on correctness it sits with the best of the group. On time it does not. The set took Qwen 415 seconds; GPT-6 Sol did it in 189 and Opus 5.5 in 341. It also waits longer than the others before showing anything, 15 to 62 seconds outside the puzzles, because it finishes thinking first.
On cost it lands in the middle: $0.21 for the seven tasks, about a third of Opus 5.5’s $0.64 and a little over twice the $0.09 of the Sol models, which use far fewer thinking tokens. Qwen wrote 22,100 output tokens across the set; GPT-6 Sol wrote 4,350.
One habit to know about: in the product copy it invented specifications we never supplied, a 250-gram weight, bone-conduction pickup and a ten-minute fast charge. Opus 5.5 invented some too. The copy read well and met every constraint, but for real product pages you need to give it the spec sheet and tell it to stay inside it.
| Model | Checks passed | Total time | Output tokens | Cost at apimodels.app prices |
|---|---|---|---|---|
| Qwen 4 Max | 93 / 93 | 415 s | 22,100 | $0.21 |
| Claude Opus 5.5 | 93 / 93 | 341 s | 37,700 | $0.64 |
| GPT-6 Sol | 93 / 93 | 189 s | 4,350 | $0.09 |
| GPT-6.1 Sol | 93 / 93 | 197 s | 3,680 | $0.09 |
| Gemini 4 Carbon | 89 / 93 | 1,501 s | 15,300 | $0.17 |
What it costs in practice
Three numbers are worth separating. The price per token is low: $1.50 / $4.50. The cost per task is moderate, because thinking is billed as output and Qwen 4 Max thinks a lot: Artificial Analysis measured $3.27 per index task, above GPT-6 Sol ($2.47) and GPT-6.1 Sol ($1.27) and well below Opus 5.5 ($10.04) and Gemini 4 Carbon ($5.48). And the cost per finished task depends on how hard your work is, since on jobs it fails it can spend a long time first.
Some arithmetic at list price. A request with 10,000 input tokens and 2,000 output tokens, thinking included, costs 10,000 × $1.50 / 1M + 2,000 × $4.50 / 1M = $0.024. A long-document request with 200,000 input tokens and 3,000 output tokens costs $0.3135, and the same request again with the document cached costs about $0.03, because cached input is billed at $0.075 per million.
The price was checked on 17 November 2026. On apimodels.app Qwen 4 Max is sold at this same list price, and built-in web search through the Responses endpoint is billed at $0.012 per search on top of the tokens.
Speed, and how to make it faster
Slowness is the most common complaint, and the public numbers bear it out. OpenRouter’s live median is about 30 output tokens per second; reviewers of the preview measured 33 to 53. Claude Opus 5.5 and GPT-6 Sol are measured at 116 and 125 by Artificial Analysis, which has not published a figure for Qwen 4 Max yet.
Most of the wait is thinking. In our tests the thinking tokens outnumbered the answer tokens on every task that reports the split; on the three puzzles it thought through 10,777 tokens to deliver a 24-token answer. If your task does not need that, switch thinking off for the request: answers start sooner and the output bill drops with it. Keep it on for multi-step reasoning, where the accuracy is the point.
When not to use it
Repository-scale coding and terminal agents. ProgramBench at 31.5 against 65.0 for Opus 5.5 and Terminal-Bench 4.0 at 48.5 are its weakest published results, and Kingy AI found it slower and less reliable than Opus on real issues.
Anything where a person is waiting, unless you switch thinking off. With thinking on, expect tens of seconds before the first word.
Design-led front-end and 3D work. Several reviewers found the visual results rough next to Kimi K3.1 Pro and Claude Fable 5.2.
Charts and screenshots. It trails both OpenAI models and Opus 5.5 on Chartography, and it cannot take audio or PDF files as input.
How it compares, one line each
Against Claude Opus 5.5: Opus is stronger on hard and multi-step work, Qwen ties on everyday tasks at about a third of the cost. The full comparison is on the Qwen 4 Max vs Claude Opus 5.5 page linked below.
Against GPT-6 Sol: about level on intelligence; Sol is roughly twice as fast and uses far fewer tokens, Qwen is stronger on agent and legal tests and keeps one price past 272K tokens.
Against Gemini 4 Carbon: Carbon scores much higher on benchmarks and takes audio and PDF; Qwen followed a written policy correctly in our tests where Carbon did not, and its price does not double on 1 December.
To call it on apimodels.app, send the model name qwen4-max to the OpenAI-compatible chat completions endpoint with the same key you use for the other models; the language model docs linked below cover thinking controls and web search.
Sources
Specs and prices: Alibaba’s QwenCloud international and mainland model pages and the preview documentation, read on 17 November 2026. Independent scores: Artificial Analysis and Vals AI, read the same day; live speed from OpenRouter. Alibaba’s chart as reproduced by DataCamp. Hands-on reports: Kingy AI, MindStudio, ZPedia on Tencent News, Ai Xuexi de Laozhang and Guixingren on NetEase, and the Hacker News discussion; the Chinese-language reports tested the preview. The seven same-prompt tests are our own, run on 17 November 2026.
Quick start
cURL
curl https://api.apimodels.app/v1/chat/completions \
-H "Authorization: Bearer $APIMODELS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen4-max",
"messages": [{"role": "user", "content": "Review this contract clause for conflicts with the policy below: ..."}]
}'FAQ
Is Qwen 4 Max good?
For everyday work, yes. In our seven same-prompt tests Qwen 4 Max got all 93 checks right, the same as Claude Opus 5.5 and GPT-6 Sol, and reviewers found it accurate on documents and well-scoped code. It is weaker on repository-scale coding, charts and visual design, and it is slow: Artificial Analysis scores it 43, thirty-eighth of 81 models.
How much does Qwen 4 Max cost?
Qwen 4 Max costs $1.50 per million input tokens and $4.50 per million output tokens on Alibaba’s international platform, with cached input at $0.075, and the price does not rise for long prompts. In mainland China it is ¥10 and ¥30. Thinking tokens are billed as output. apimodels.app sells it at the same $1.50 / $4.50.
Is Qwen 4 Max open source?
No. Qwen 4 Max is available only through an API. Alibaba has said smaller models in the Qwen 4 family will be released with open weights later, without giving a date.
What is the context window of Qwen 4 Max?
Qwen 4 Max has a one-million-token context window. Alibaba’s international page lists a maximum output of 131K tokens, or 262K including thinking; its mainland page lists 393K, so check the page for the region you use.
Why is Qwen 4 Max slow, and can I speed it up?
Most of the wait is thinking: Qwen 4 Max thinks by default and finishes thinking before it writes, so the first word can take tens of seconds. You can switch thinking off per request, which makes answers start sooner and lowers the output bill. Its writing speed, about 30 to 53 tokens per second in public measurements, is also below Claude Opus 5.5 and GPT-6 Sol.
How can I use Qwen 4 Max outside China?
Qwen 4 Max is served from Alibaba Cloud’s international regions, and it is also available through apimodels.app with one API key: call POST https://api.apimodels.app/v1/chat/completions with model qwen4-max. No Alibaba Cloud account is needed for that route.
