Qwen 4 Max vs Claude Opus 5.5
Qwen 4 Max is Alibaba’s new flagship model. It ran as a limited preview from 11 November 2026 and opened for pay-as-you-go API use on 16 November under the model ID qwen4-max. Claude Opus 5.5 is Anthropic’s current Opus, released on 22 September 2026. Alibaba’s own launch chart compares Qwen 4 Max with Opus 5.5 line by line, and “qwen 4 max vs opus 5.5” is already a search suggestion, so this is the comparison people want first.
This page puts four kinds of evidence side by side: independent benchmarks, the numbers Alibaba published, hands-on tests by other developers, and seven tasks we ran ourselves on both models with the same prompts.
We run apimodels.app, which sells both models through one API key, so we have no reason to favour either. The benchmark scores below are not ours: they come from Artificial Analysis, Vals AI, Alibaba and the reviewers named in each table, with links at the end. The seven same-prompt tests are ours; we ran them on 17 November 2026 through our own API with each model’s default thinking setting.
Qwen 4 Max
Ties on everyday work for about a third of the cost
- ·Artificial Analysis Intelligence Index 43 (Opus 5.5: 59)
- ·Leads on HLAB legal tasks 21.3 vs 9.2 and Agents’ Last Exam 46.7 vs 43.4 (Alibaba’s chart)
- ·$1.50 / $4.50 per 1M tokens, one price across the whole 1M window
- ·Text, image and video input; up to 131K output tokens
Claude Opus 5.5
Stronger on hard and multi-step work
- ·Artificial Analysis Intelligence Index 59 at max effort, 48 at medium
- ·Terminal-Bench 4.0 59.6 vs 48.5; ProgramBench 65.0 vs 31.5 (Alibaba’s chart)
- ·$2.40 / $12.00 per 1M tokens on apimodels.app (list $4 / $20)
- ·Text and image input; up to 128K output tokens
Pick Claude Opus 5.5 for multi-file and terminal coding, long agent runs and anything where a wrong answer is expensive. Pick Qwen 4 Max for well-scoped tasks at volume, legal and document agent work, video input, and when cost per task is what decides.
Specs and prices side by side
On paper the two are close: a one-million-token context window, a similar output cap, thinking switched on by default, tool calling and structured output. The differences that matter are price, video input and how each one controls thinking.
Qwen 4 Max charges the same rate across its whole window, and its thinking can be switched off per request. Opus 5.5 also keeps one rate across the window, but its thinking cannot be switched off; you lower the effort level instead. One detail to watch on Qwen: Alibaba’s international page lists a 131K output cap (262K counting the thinking), while its mainland page lists 393K, so check the page for the region you call.
| Qwen 4 Max | Claude Opus 5.5 | |
|---|---|---|
| Maker, release | Alibaba, 16 Nov 2026 (preview from 11 Nov) | Anthropic, 22 Sep 2026 |
| List price per 1M tokens (input / output) | $1.50 / $4.50 | $4 / $20 |
| Cached input, list | $0.075 | $0.20 |
| Price on apimodels.app | $1.50 / $4.50 | $2.40 / $12.00 |
| Context window | 1M | 1M |
| Max output | 131K (262K with thinking) | 128K |
| Input types | Text, image, video | Text, image |
| Thinking | On by default, can be switched off | Always on; effort low to max, default medium |
| Open weights | No | No |
Independent benchmarks
Two independent labs have both models on the same tests, and both put Opus 5.5 ahead. Artificial Analysis scores Opus 5.5 at 59 and Qwen 4 Max at 43 on its Intelligence Index, which places Qwen 38th of 81 models. Vals AI has them closer: 70.14% against 63.20% on the Vals Index, with Qwen 11th of 38.
The settings are not equal, and that changes how to read the gap. Artificial Analysis ran Opus 5.5 at max effort and has tested only one setting for Qwen 4 Max so far. At medium effort, which is what Opus 5.5 uses unless you ask for more, its score is 48, five points above Qwen rather than sixteen.
Qwen 4 Max is also a talkative thinker. It produced 127 million output tokens across the Artificial Analysis index, against a median of 78.6 million for the models tested, so its lower price per token does not carry over one-to-one to the bill.
| Measure | Run by | Qwen 4 Max | Claude Opus 5.5 |
|---|---|---|---|
| Intelligence Index v5.0.1 | Artificial Analysis | 43 | 59 (max); 48 (medium) |
| Cost per index task, list prices | Artificial Analysis | $3.27 | $10.04 |
| Vals Index v2.2 | Vals AI | 63.20% (11th of 38) | 70.14% |
| Cost per Vals test | Vals AI | $9.33 | $24.77 |
| Time per Vals test | Vals AI | 64 min 39 s | 84 min 31 s |
| Terminal-Bench 4.0 | Vals AI | 48.48% | 65.15% |
| Code Arena WebDev (Elo, preliminary) | Arena | 1750 | 1816 |
What Alibaba’s own chart shows
Alibaba published sixteen benchmarks with Opus 5.5 in the next column, and it did not pick only the ones it wins: by its own numbers Qwen 4 Max leads on six and Opus 5.5 on ten. These are vendor-reported figures, so read them as Alibaba’s claim, not as an independent result.
The pattern is consistent. On a single well-defined software issue the two are level (DeepSWE 75.7 against 74.2). Once the job spans a whole repository or a terminal session, Opus pulls well ahead: ProgramBench 65.0 against 31.5, Terminal-Bench 4.0 59.6 against 48.5. Qwen’s clearest win is legal agent work, 21.3 against 9.2 on HLAB, and it edges ahead on Agents’ Last Exam and AutomationBench. Although Qwen accepts video and Opus does not, Opus scores higher on Alibaba’s chart and video understanding tests.
| Benchmark (reported by Alibaba) | Qwen 4 Max | Claude Opus 5.5 |
|---|---|---|
| DeepSWE v1.1 | 75.7 | 74.2 |
| Agents’ Last Exam v1.1 | 46.7 | 43.4 |
| AutomationBench-AA | 72.5 | 69.5 |
| HLAB-Vals-1.1 (legal) | 21.3 | 9.2 |
| BioMysteryBench-1.1 | 85.0 | 81.9 |
| HealthBench Professional | 60.0 | 59.9 |
| CodeArena (Elo) | 1750 | 1816 |
| Terminal-Bench 4.0 | 48.5 | 59.6 |
| ProgramBench | 31.5 | 65.0 |
| APEX-Agents v1.1 | 72.1 | 73.5 |
| AA-Briefcase v1.1 (Elo) | 1686 | 1801 |
| JobBench | 68.5 | 75.4 |
| Humanity’s Last Exam | 51.7 | 59.8 |
| Chartography | 54.2 | 66.3 |
| Video-Index | 68.1 | 74.8 |
Same jobs, run by other developers
The most careful head-to-head so far is Curtis Pyke’s at Kingy AI, who ran both models through the same coding tasks twice each. On eight real GitHub issues from SWE-bench Verified, Opus 5.5 resolved 11 of 16 runs and Qwen 4 Max 9 of 16, and on the hidden test cases the gap was wider, 37 of 46 against 24 of 46. Opus was also about five times quicker in active time, 86 seconds against 431 at the median. Qwen’s bill was 76% lower: $4.42 against $18.65.
On small self-contained repositories the gap nearly closed, 8 of 8 against 7 of 8, and Qwen’s one miss was a detail: it raised a KeyError where the documentation asked for a ValueError. On 18 simple tasks both models passed every check. That matches what we saw in our own tests below.
One finding from the same review is worth keeping in mind before you build a routing rule. Sending every task to Qwen first and escalating the failures to Opus cost 2.8 times as much as using Opus alone in his runs, because Qwen spent a long time on the tasks it then failed. A lower price per token did not mean a lower price per solved problem.
Chinese reviewers who tested the preview reported strong document work and slow long jobs. ZPedia’s twelve-task run had Qwen read a financial report in eight minutes with all six spot-checked figures correct and answer five needle-in-a-haystack questions over 1,600 support tickets without a miss; it also left one debt item out of a prospectus summary and one product line out of a slide deck. On Hacker News, Simon Willison’s SVG test only came out right at the highest thinking setting, which took 7 minutes 23 seconds.
| Test (who ran it) | Qwen 4 Max | Claude Opus 5.5 |
|---|---|---|
| 8 SWE-bench Verified issues × 2 runs (Kingy AI) | 9 of 16 resolved | 11 of 16 |
| Hidden test cases, same issues (Kingy AI) | 24 of 46 | 37 of 46 |
| Median active time per issue (Kingy AI) | 431 s | 86 s |
| Cost of the whole run (Kingy AI) | $4.42 | $18.65 |
| 4 small repositories × 2 runs (Kingy AI) | 7 of 8 | 8 of 8 |
| 18 simple tasks (Kingy AI) | 48 of 48 | 48 of 48 |
Our seven same-prompt tests: a tie on answers, not on time or cost
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 has to call tools and 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 task has an automatic check, 93 checks in total, and we verified the checks against reference answers before spending anything.
Both models got all 93 checks right. The function passed all 50 hidden cases on both; both found all five bugs planted in the cache class; both refunded the defective keyboard for exactly $129, refused the change-of-mind mouse that was 46 days old, and left the shipping fee alone; both counted the eight payment timeouts in the log and found the one request ID that crossed three services.
The differences were in how they got there. Opus 5.5 started answering sooner on most tasks, usually within 8 to 26 seconds, and finished the set in 341 seconds against 415 for Qwen 4 Max. Qwen thinks first and then writes, so its first visible word arrived after 15 to 62 seconds. On the three puzzles both thought for about three minutes, and Opus used the entire 20,000-token allowance we set, which made that one task cost 24 cents against 5 cents on Qwen.
The bill for all seven tasks, at apimodels.app prices, was $0.64 on Opus 5.5 and $0.21 on Qwen 4 Max, about a third. Both models pad their answers with thinking tokens: Opus wrote 37,700 output tokens in total and Qwen 22,100.
One thing the checks did not score: in the product copy, both invented specifications we never gave them. Qwen added a 250-gram weight, bone-conduction pickup and “ten minutes of charging for three hours of playback”; Opus added five noise-cancelling levels and auto-pause. If you use either for product pages, give it the real spec sheet and tell it to use nothing else.
| Task (checks) | Qwen 4 Max | Claude Opus 5.5 |
|---|---|---|
| Write a timeline-merge function (50 hidden cases) | 50/50 · 51 s · $0.012 | 50/50 · 32 s · $0.044 |
| Fix a TTL cache class with 5 planted bugs (10) | 10/10 · 27 s · $0.009 | 10/10 · 27 s · $0.040 |
| Refund agent with tools and a policy (7) | 7/7 · 16 s · $0.006 | 7/7 · 12 s · $0.022 |
| Five questions over a 1,500-line log (5) | 5/5 · 63 s · $0.120 | 5/5 · 52 s · $0.233 |
| Chinese group chat to JSON (13) | 13/13 · 19 s · $0.004 | 13/13 · 10 s · $0.017 |
| Three counting puzzles (3) | 3/3 · 196 s · $0.049 | 3/3 · 179 s · $0.242 |
| Chinese product copy, five hard constraints (5) | 5/5 · 43 s · $0.010 | 5/5 · 29 s · $0.039 |
| All seven | 93/93 · 415 s · $0.21 | 93/93 · 341 s · $0.64 |
What a task costs
Per token, Qwen 4 Max is priced at well under half of Opus 5.5: $1.50 against $4 for input and $4.50 against $20 for output at list. Per task the gap narrows, because Qwen writes more thinking tokens, but it stays large. Artificial Analysis measured $3.27 per index task on Qwen and $10.04 on Opus at list prices; Vals measured $9.33 against $24.77.
On apimodels.app Opus 5.5 is $2.40 / $12.00, which brings an Artificial Analysis task to about $6.02, and Qwen 4 Max is at Alibaba’s list price, about $3.27. So on our prices Qwen costs a little over half as much per hard task, and about a third on the lighter tasks in our own set.
Two things move these numbers in practice. Prompt caching matters more on Opus, where a cached token is billed at a twentieth of the input price; long agent sessions that resend the same context get most of their input at that rate. And Kingy AI’s routing result above is the warning for the other direction: on hard tasks Qwen can burn time and tokens and still fail, so measure cost per finished task on your own work, not cost per token.
Speed
Opus 5.5 is the faster of the two once it starts writing. Artificial Analysis measured 116 output tokens per second for Opus 5.5 and has not yet published a speed for Qwen 4 Max; OpenRouter’s live figures show about 30 tokens per second for Qwen, and reviewers of the preview measured between 33 and 53.
On our own platform over the last 30 days, Opus 5.5 answered with a first token after about 5 seconds at the median and wrote at roughly 170 tokens per second (production calls, sample size not disclosed). Qwen 4 Max went on sale here on 16 November, so we have no comparable production figure yet; in our tests, leaving out the puzzles, its first visible word took 15 to 62 seconds because it finishes thinking before it writes.
If a person is waiting for the reply, lower the thinking on either model. Qwen 4 Max lets you switch thinking off for a request; Opus 5.5 lets you drop effort to low.
Which to use
Multi-file refactors, terminal agents, long coding sessions: Claude Opus 5.5. It leads by a wide margin on ProgramBench and Terminal-Bench 4.0, and in Kingy AI’s runs it solved more issues in a fifth of the time.
Single functions, bug fixes, structured extraction, tool-calling agents with clear rules: either. Both were perfect on our seven tasks and on Kingy AI’s 18 simple ones; Qwen 4 Max does it for about a third of the cost.
Legal and contract agent work: try Qwen 4 Max first. Its lead on HLAB, 21.3 against 9.2, is the largest margin in its favour, and Vals independently measured 21.25%.
Video input: Qwen 4 Max. Opus 5.5 takes text and images only.
Hard research and knowledge questions, charts: Claude Opus 5.5, which leads on Humanity’s Last Exam (59.8 against 51.7) and Chartography (66.3 against 54.2).
Chinese-language work: we found no head-to-head data from anyone, and in our two Chinese tasks both models scored full marks. Test with your own material.
On apimodels.app switching between them is a change of the model string on the same key, so the practical approach is to run a sample of your own tasks on both and compare the cost per finished task.
Caveats
Qwen 4 Max has been generally available for one day at the time of writing. Several of the hands-on reports above tested the preview, Artificial Analysis has not yet measured its speed and has run only one thinking setting, and some leaderboards (SWE-bench’s own, Aider, LiveBench) do not list it yet. Expect these numbers to move.
Our seven tasks are everyday-difficulty work, run once per model at default settings. They show that both models handle that level reliably; they do not measure the hard end, which is where the benchmarks above find the gap. Sources also disagree on one Opus 5.5 figure: Alibaba’s chart gives it 59.6 on Terminal-Bench 4.0 and Vals gives 65.15%.
Sources
Specs and list prices: Alibaba’s QwenCloud model page and Anthropic’s pricing page, read on 17 November 2026. Independent scores, task costs and speeds: Artificial Analysis and Vals AI, read the same day. Alibaba’s sixteen-benchmark chart as reproduced by DataCamp. Hands-on reports: Kingy AI (opened and checked), ZPedia on Tencent News, and the Hacker News discussion. The seven same-prompt tests are our own, run on 17 November 2026.
Inicio rápido
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": "Find the bugs in this class: ..."}]
}'
# same key, same endpoint: change "model" to "claude-opus-5-5" to run Opus 5.5Preguntas frecuentes
Is Qwen 4 Max better than Claude Opus 5.5?
Not overall. Claude Opus 5.5 scores 59 on the Artificial Analysis Intelligence Index against 43 for Qwen 4 Max, and leads on terminal and repository-scale coding, hard knowledge questions and chart reading. Qwen 4 Max leads on legal agent tasks and a few automation tests and matches Opus on single-issue fixes and everyday tasks: in our seven same-prompt tests both scored 93 out of 93.
How much does Qwen 4 Max cost compared with Opus 5.5?
Qwen 4 Max lists at $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075. Claude Opus 5.5 lists at $4 and $20, with cached input at $0.20; on apimodels.app it is $2.40 and $12.00. Per task the gap is smaller than per token because Qwen writes more thinking tokens: Artificial Analysis measured $3.27 per task on Qwen and $10.04 on Opus at list prices.
Is Qwen 4 Max good for coding?
It is strong on well-scoped coding and weaker on large, multi-step jobs. On single-issue fixes it matches Opus 5.5 (DeepSWE 75.7 against 74.2), and in our tests it passed all 50 hidden cases on a function and found all five planted bugs in a class. On repository-scale and terminal work it falls behind: 31.5 against 65.0 on ProgramBench and 48.5 against 59.6 on Terminal-Bench 4.0, by Alibaba’s own chart.
Which is faster, Qwen 4 Max or Opus 5.5?
Claude Opus 5.5 is faster. Artificial Analysis measured 116 output tokens per second for Opus 5.5, while public figures for Qwen 4 Max range from about 30 to 53. In our seven tests Opus finished in 341 seconds and Qwen in 415, and Opus began writing sooner on most tasks because Qwen completes its thinking before it outputs anything.
What is the context window of Qwen 4 Max?
Qwen 4 Max has a one-million-token context window with one price across the whole window, the same size as Claude Opus 5.5. Alibaba’s international page lists up to 131K output tokens, or 262K counting thinking, and text, image and video input.
Can I use Qwen 4 Max and Claude Opus 5.5 with one API key?
Yes. Both run on one apimodels.app key. Call POST https://api.apimodels.app/v1/chat/completions with model qwen4-max or claude-opus-5-5; Opus 5.5 also works on the Anthropic-format /v1/messages endpoint.
