Qwen 4 Max is much cheaper than Claude Opus 5.5 and much slower. Per million tokens it costs $1.50 for input and $4.50 for output, against $4 and $20 for Opus 5.5. On our coding task both passed all 59 hidden tests. Qwen 4 Max at high cost a quarter to a third less than Opus 5.5 at medium and took about six times as long.
Alibaba released Qwen 4 Max on November 16, 2026. Opus 5.5 is the model Anthropic tells most users to start with. We ran both on the same task in the same agent, three runs per setting, and gathered the independent scores in one place.
Which should you use?
Pick Opus 5.5 if you are watching the agent work. In our test it finished in 4.2 minutes at medium and 9.6 at high, and Qwen 4 Max needed 25 minutes at high. Opus 5.5 is also the safer choice for hard tasks, given the scores below.
Pick Qwen 4 Max for long jobs you start and come back to, such as overnight runs or batches of routine fixes. At high it cost a quarter to a third less than Opus 5.5 at medium, and it finished every run. Leave it at high. On our task xhigh doubled the time and added nothing.
Here is what we measured, averaged over three runs of each setting:
| Model and effort | Hidden tests | Time | Steps | Cost per task |
|---|---|---|---|---|
| Qwen 4 Max, high (Alibaba’s documented default; we set it explicitly) | 59/59 in all 3 runs | 25.0 min | 76 | about $0.65 (mean of 2 fully recorded runs) |
| Qwen 4 Max, xhigh | 59/59 in all 3 runs | about 47 min, one run stopped at 60 | 75 | $0.84 to $1.28 recorded, real cost higher |
| Opus 5.5, medium (API default) | 59/59 in all 3 runs | 4.2 min | 10 | $0.97 |
| Opus 5.5, high | 59/59 in all 3 runs | 9.6 min | 17 | $1.70 |
Our runs in Atomic Agent. Opus 5.5 on November 9, Qwen 4 Max on November 16, same task and same agent build. The Qwen high cost is the mean of the two fully recorded runs, $0.62 and $0.68. The third and longest run, 92 steps and 27.2 minutes, was not fully recorded and likely cost more.
What the numbers show:
- Qwen 4 Max at high cost a quarter to a third less than Opus 5.5 at medium and took six times as long. Those are the documented defaults of both models.
- Qwen 4 Max works in many short steps. It took 66 to 92 steps at high, while Opus 5.5 took about 10 at medium and 17 at high. Each Qwen step averaged about 20 seconds, against about 25 for Opus 5.5 at medium. There were just far more of them.
- The price gap per task is much smaller than per token. Qwen 4 Max output is 4.4 times cheaper per token, but it spent more tokens and steps to reach the same result.
- Opus 5.5 at high took 9.6 minutes for $1.70. Qwen 4 Max at xhigh took about 47 minutes, and its recorded cost of $0.84 to $1.28 is a lower bound.
What we measured
We ran all four settings through Atomic Agent, our open-source agent, on a build included in the 0.7.0 release, in cloud mode through OpenRouter. Opus 5.5 was pinned to Anthropic. For qwen/qwen4-max, Alibaba is the only provider on OpenRouter. Each run got a fresh, isolated copy of the same Python project and the same one-line prompt. The time limit was 60 minutes (20 for the Opus 5.5 medium runs, which finished in about 4).
The task is a warehouse stock library of about 1,400 lines with 13 planted bugs and one missing feature, stock holds. The agent has to read a written spec, fix the bugs, build the feature, and get the test suite passing. We then score the result against 59 hidden tests the agent never sees. The untouched code passes 20 of them.
Cost is the total OpenRouter reported for each logged request. Here is how the runs varied:
- Qwen 4 Max at high took 24.1, 27.2, and 23.7 minutes, with 66, 92, and 69 steps. The two fully recorded runs cost $0.62 and $0.68, a mean of about $0.65. The third and longest run was not fully recorded, so we left its cost out, and it likely cost more.
- Qwen 4 Max at xhigh took 47.3 and 47.4 minutes in two runs. The 60-minute limit stopped the third, and its code already passed all 59 tests. Every xhigh run is missing part of its requests in the log, so the recorded $0.84, $0.92, and $1.28 are too low.
- Opus 5.5 runs were all fully recorded. At medium the cost ranged from $0.96 to $1.00. At high it ranged from $1.54 to $1.98, with 13 to 25 steps.
Cache reads showed up on the Qwen 4 Max bill. About 2.7 million input tokens were read from cache in each fully recorded high run. At the full $1.50 per million, that alone would cost over $4, and the runs cost $0.62 to $0.68. Atomic Agent adds cache markers only for Anthropic models, so this caching came from Alibaba’s side. Alibaba’s docs list implicit caching for the model, while OpenRouter’s implicit caching flag for it says false.
The reasoning token counts suggest a higher setting reached Alibaba. The xhigh runs used 78K to 133K reasoning tokens each, against 34K to 41K in the fully recorded high runs.
This is one task with three runs per setting. Every finished run scored 59 out of 59, so this test says nothing about which model is smarter. It shows what each setting spent in money, time, and steps to reach the same result.
Timing is rough. The Opus 5.5 medium runs went one after another, while runs of five other models went at the same time. The Opus 5.5 high runs went fully in parallel, alongside three Fable 5.2 runs. All six Qwen 4 Max runs shared one machine and one API key at the same time, a week later.
We did not test Opus 5.5 at low, xhigh, or max, or Qwen 4 Max at low. On a task hard enough that one model fails, the result could look very different.
Pricing
| Per 1M tokens | Qwen 4 Max | Opus 5.5 |
|---|---|---|
| Input | $1.50 | $4 |
| Output, including thinking | $4.50 | $20 |
| Cache read | $0.075 | $0.20 |
| 5-minute cache write | $1.875 | $5 |
| 1-hour cache write | $3 | $8 |
| Batch input / output | not offered | $2 / $10 |
| Fast mode input / output | not offered | $8 / $40 |
| Cost per task, our test, high / medium | about $0.65 (high) | $0.97 (medium) |
| Cost per task, our test, next level up | $0.84 to $1.28 recorded (xhigh) | $1.70 (high) |
Qwen 4 Max prices from Alibaba Cloud Model Studio (International) and OpenRouter on November 16. Opus 5.5 prices from Anthropic’s pricing page on November 9.
Per token, Qwen 4 Max input is about 2.7 times cheaper and output about 4.4 times cheaper. Model Studio gives 1 million free tokens in its International region, valid for 90 days.
Opus 5.5 has a Batch API at half price, and Alibaba’s docs list batch inference for Qwen 4 Max as not supported. Batch suits one-off bulk requests that can wait for results. It does not fit an agent loop, where each step waits for the reply to the previous one, so the batch row does not change the agent comparison.
Independent results
Every outside evaluator we checked puts Opus 5.5 ahead. These are their figures as listed on November 16:
| Board | Qwen 4 Max | Opus 5.5 |
|---|---|---|
| Artificial Analysis Intelligence Index | 43 (level not stated) | 59 at max, 52 at high, 48 at medium |
| Artificial Analysis cost per task | $3.27 | $10.04 at max |
| Vals Index | 63.20% (xhigh) | 70.14% |
| Vals Terminal-Bench 4.0 | 48.48% | 65.2% |
| Vals cost per test | $9.33 | $24.77 |
| Arena Code WebDev | 1750 (xhigh, preliminary) | 1816 (max) |
| Arena Agent, net improvement | 7.46% (preliminary) | 18.19% (max) |
On Artificial Analysis, Qwen 4 Max scores 16 points below Opus 5.5 at max for about a third of the cost per task. It also scores 5 points below Opus 5.5 at medium, the setting our Opus runs used. Artificial Analysis does not say which effort level it ran Qwen 4 Max at, and it lists no time per task for it. For Opus 5.5 at max it lists 24.9 minutes.
On Vals, the gap is 7 points on the overall index and about 17 on Terminal-Bench 4.0. Vals ran Qwen 4 Max at xhigh, and its models do not all run at the same effort, so read the cost row as what each entry cost in that run.
Arena Agent lists a median of $1.47 and 15.7 minutes per task for Qwen 4 Max, ranked 14th. Opus 5.5 at max has more than twice its net improvement. Arena’s Qwen ratings are preliminary and move as votes come in.
Benchmarks from Alibaba
Anthropic has not published scores for Qwen 4 Max. Alibaba’s launch table includes Opus 5.5, and these are its two columns:
| Benchmark | Qwen 4 Max | Opus 5.5 |
|---|---|---|
| DeepSWE v1.1 | 75.7 | 74.2 |
| Terminal-Bench 4.0 | 48.5 | 59.6 |
| ProgramBench | 31.5 | 65.0 |
| CodeArena WebDev (Elo) | 1750 | 1816 |
| Agents’ Last Exam | 46.7 | 43.4 |
| APEX-Agents | 72.1 | 73.5 |
| JobBench | 68.5 | 75.4 |
| AutomationBench (AA) | 72.5 | 69.5 |
| AA-Briefcase (Elo) | 1686 | 1801 |
| HLAB-Vals | 21.3 | 9.2 |
| AgentArena (net) | 7.5 | 18.2 |
| HLE | 51.7 | 59.8 |
| BioMysteryBench | 85.0 | 81.9 |
| HealthBench Professional | 60.0 | 59.9 |
| Chartography | 54.2 | 66.3 |
| Video-Index | 68.1 | 74.8 |
Compiled by Alibaba in the Qwen 4 Max blog post, November 16, 2026. Alibaba takes some rows from Arena, Artificial Analysis, Vals, Mercor, and the Agents’ Last Exam leaderboard, and ran the rest itself. Its footnotes do not state effort levels. These are the vendor’s numbers.
Even in Alibaba’s own table, Opus 5.5 leads on 10 of 16 rows. Qwen 4 Max leads on DeepSWE by 1.5 points, on Agents’ Last Exam, AutomationBench, HLAB-Vals, and BioMysteryBench, and on HealthBench Professional by 0.1. The widest gap is ProgramBench, where Opus 5.5 scores more than twice as high.
Alibaba’s Opus 5.5 figure for Terminal-Bench 4.0, 59.6, does not say where it comes from. Vals has Opus 5.5 at 65.2% on the same benchmark, while its Qwen 4 Max score of 48.48% matches Alibaba’s 48.5. By the Vals numbers the gap is about 17 points, not 11.
Effort levels are not the same scale
| Qwen 4 Max | Opus 5.5 | |
|---|---|---|
| API default | high | medium |
| Levels | low, high, xhigh | low, medium, high, xhigh, max |
If you send medium | you get high | you get medium |
If you send max | you get xhigh | you get max |
| Thinking | on by default | always on |
From the Qwen 4 Max documentation and Anthropic’s models overview, November 16.
The word high means different things on the two models. Each lab sets its own scale, and Qwen 4 Max has three levels where Opus 5.5 has five. So we compared the settings by what they cost and how long they took, not by name. Qwen 4 Max at high cost less than Opus 5.5 at medium and took more than twice as long as Opus 5.5 at high. Qwen 4 Max at xhigh, its top level, is the closest counterpart to Opus 5.5 at max by name only.
One shared setting does not line the two models up. With medium in your config, Opus 5.5 runs at medium and Qwen 4 Max at high, because Alibaba maps it up. With max, Opus 5.5 runs at max and Qwen 4 Max at xhigh.
On Qwen 4 Max, most published scores are at xhigh. Vals and Arena list xhigh, and Alibaba’s Terminal-Bench figure appears to be at xhigh too. Our test ran both levels, and high was enough. The Qwen 4 Max review has more on the effort mapping.
Limits and other differences
| Qwen 4 Max | Opus 5.5 | |
|---|---|---|
| Context window | 1M tokens, input up to 983,616 with thinking | 1M tokens |
| Output per request | 131,072 on OpenRouter | 128,000, or 300,000 on Message Batches with a beta header |
| Knowledge cutoff | not published | June 2026 |
| Weights | closed | closed |
| Input types | text, image, video | text, image |
From Alibaba’s model documentation, OpenRouter, and Anthropic’s models overview, November 16.
Alibaba’s docs allow 262,144 tokens for thinking and answer together, and the default thinking budget is 131,072 tokens. On OpenRouter the completion cap is 131,072, so a long xhigh request could hit it. None of our runs did.
Qwen 4 Max accepts the Anthropic Messages format as well as OpenAI Chat Completions, OpenAI Responses, and Alibaba’s own DashScope protocol. Neither model can run on your own machine. Qwen 3.8 Max had open base weights, and Qwen 4 Max does not.
How to try both
To repeat our comparison on your own tasks, download Atomic Agent for Mac, Windows or Linux, or install the terminal version. Our test ran on a build included in the 0.7.0 release.
curl -fsSL https://atomicagent.io/install | shatomic-agentOn Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell. The source is on GitHub.
Connect OpenRouter with your own key and choose qwen/qwen4-max or anthropic/claude-opus-5.5 from the live model list; see our OpenRouter guide. Atomic Agent sends no effort value by default, so Opus 5.5 runs at medium, as in our test. For Qwen 4 Max we set high explicitly. To set the level, add one line to your openrouter entry under llm.providers in ~/.atomic-agent/config.json:
"extraBody": { "reasoning": { "effort": "high" } }That line applies to every model on the entry, and as the section above shows, the same word lands on different levels on each model. Change it when you switch. More in our models guide.
For the full Qwen 4 Max review, see Qwen 4 Max: price, benchmarks, and the effort setting. Opus 5.5 against Anthropic’s top model is in Fable 5.2 vs Opus 5.5, and another Chinese model on the same task is in our Kimi K3.1 Pro review.
The takeaway
On our task, Qwen 4 Max at high did the same job as Opus 5.5 at medium for about $0.65 instead of $0.97, and took 25 minutes instead of 4. Its token price is 4.4 times lower on output, but the many extra steps ate most of that gap.
Keep Opus 5.5 for interactive work and hard problems. Try Qwen 4 Max at high for background jobs where time is cheap. Do not pay for xhigh until your own tasks show a gain, because on ours it only doubled the time.


