Qwen4-Max undercuts Claude Opus 5.5 on price, but pays for it in speed and accuracy
Alibaba's new Qwen4-Max model costs roughly a third as much as Anthropic's Claude Opus 5.5, but independent testing reveals the trade-offs: slower responses and lower accuracy on coding tasks.

Alibaba released Qwen4-Max to Token Plan subscribers on November 11 as a preview, followed by general availability on November 16. The model has earned a reputation for verbosity: when testing firm Artificial Analysis ran its standard evaluation, Qwen4-Max produced 127 million output tokens, roughly double the median for comparable models. Claude Opus 5.5 generated 218 million output tokens under maximum effort with default fallback enabled, though that represents Opus at its highest setting; the model's default medium effort uses fewer tokens.
Pricing differences amplify the token volume gap. Qwen4-Max charges $1.50 per million input tokens and $4.50 per million output tokens, while Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Cached input pricing also diverges: Qwen4-Max offers $0.075 per million cached input tokens against Opus 5.5's $0.20 per million. Cache writes carry separate rates on both platforms.
Cheaper only wins if you don't give up accuracy to get it.
Alibaba positions Qwen4-Max as an everyday workhorse, marketing it as "your new teammate." Published benchmarks show mixed results: the model wins and loses against Opus on agent tasks that require step-by-step tool use. Coding performance is similarly uneven—Qwen4-Max trails on Terminal-Bench 4.0 (48.5% versus Opus's 59.6%) but leads on DeepSWE v1.1. These published results employ different benchmarks and evaluation methodologies, making direct comparison difficult.
To evaluate whether Qwen4-Max offers a practical alternative for everyday work, independent testing was conducted across three agent-based tasks. Opus 5.5 ran through Anthropic's API while Qwen4-Max ran through OpenRouter on Alibaba's servers. Each model executed once per test at default settings with no turn limit. A custom grading script evaluated answers against a prepared answer key. All three tests were validated as sufficiently difficult by running them first through GPT-5 as a baseline.
The tests
- Multi-app workflow (76 points) – Models worked as agents across a fake company's ticket tracker, code repository, chat system, and calendar. Tasks included triaging a customer bug, locating the code change responsible, assigning ownership, calculating deadlines, counting tickets, and refusing two policy-violating requests.
- Incident investigation (60 points) – Models analyzed logs, metrics, and change history to explain three simultaneous outages. The dataset included misleading deploys, two servers with incorrect clocks, and a deceptive alert.
- Agent bug fix (47 points) – Models repaired an accounting library containing 11 planted bugs. Six bugs produced no visible test failures, requiring comparison with documentation. Three of the library's own tests were intentionally incorrect; models should report them without modification.
Multi-app workflow
Opus 5.5 achieved all 76 points. Qwen4-Max scored 71, with all five missed points stemming from a single error. When three questions asked which team owns specific code, Qwen4-Max selected the most specific rule from the ownership file rather than the last matching rule, as the file specified. All other work proved correct, including timezone-aware deadline calculations and both required refusals.
…Qwen4-Max can handle a long, messy office job, but one misread rule followed it through the whole task.
The performance metrics revealed stark contrasts. Qwen4-Max generated nearly twice the output tokens as Opus while costing less than one-third. However, it required more than three times longer to complete the task in roughly half as many conversational turns.
- Opus 5.5: 44 turns, 2,451,342 input, 26,617 output, $2.13, 248 seconds
- Qwen4-Max: 25 turns, 1,746,170 input, 51,264 output (41,644 reasoning), $0.62, 791 seconds
Incident investigation
Qwen4-Max narrowly outperformed on this task, scoring 59 of 60 points compared to Opus 5.5's 58. Both models successfully untangled the three outages, identified the two-step root cause of the primary incident, eliminated all decoys, and detected both clock synchronization problems. Both made identical error types: they attributed a customer-facing error code to the wrong service, with Opus making this mistake twice and Qwen4-Max once.
This represented Qwen4-Max's strongest performance across all tests. It achieved one additional point versus Opus while consuming approximately 31% of the reported cost. A single test run cannot establish broader advantages in agent work. Opus 5.5 demonstrated superior speed, completing the investigation in under four minutes compared to Qwen4-Max's 9.5 minutes.
- Opus 5.5: 22 turns, 1,216,769 input, 23,642 output, $1.41, 221 seconds
- Qwen4-Max: 18 turns, 1,104,031 input, 33,036 output (23,153 reasoning), $0.44, 570 seconds
Agent bug fix
Neither model achieved perfect accuracy on this coding task. Opus 5.5 scored 45 of 47 points while Qwen4-Max scored 40. This result aligned with Opus's advantage on Terminal-Bench 4.0, though Alibaba's published coding benchmarks showed mixed outcomes.
Opus successfully fixed all 11 bugs but failed to report the third incorrect test, costing two points. Qwen4-Max missed that same test and lost five additional points. It fixed 10 of 11 bugs, failing to catch one where splitting a payment across invoices loses a cent. It also left a scratch file in the tests folder, which the evaluation rules counted as modifying the test suite.
Tool logs revealed different working patterns. Opus made its first fix on turn 4 and completed the task in 12 turns total. Qwen4-Max spent 15 turns reading files before making its first edit and required 30 turns overall. The additional reading did not translate to better accuracy.
The cost differential narrowed on this test. Qwen4-Max consumed more than three times the input tokens, resulting in a bill approximately two-thirds of Opus's.
- Opus 5.5: 12 turns, 228,824 input, 6,433 output, $0.38, 64 seconds
- Qwen4-Max: 30 turns, 827,545 input, 15,813 output (6,701 reasoning), $0.25, 310 seconds
Results
Qwen4-Max's reputation for verbosity held up across testing. The model generated 77% more output tokens than Opus across all three tasks. These totals include reasoning tokens and do not measure only visible answer length; the models also employ different tokenizers.
What Qwen4-Max really costs is time and accuracy.
Despite high token output, Qwen4-Max's low pricing meant the total bill reached approximately one-third of Opus's cost. The real expense manifested in execution time and accuracy. Qwen4-Max required 3.14 times longer overall and scored 170 of 183 points, nine fewer than Opus's 179.
What do I think?
Each model ran only once, meaning point differences of one or two could shift in subsequent runs. Prospective users should test Qwen4-Max on their own long-running agent tasks before relying on it for code work. Answers should be checked against source material, and code should be verified through tests and human review; agreement from a second model alone does not confirm accuracy.
Opus 5.5 remains preferable for coding tasks and any work where errors carry downstream consequences. It demonstrated superior accuracy on two of three tests and completed all three tasks in roughly one-third of the total time.
Qwen4-Max has earned consideration for specific use cases. On investigation work involving extensive data analysis where additional processing time is acceptable, it matched Opus's performance at a fraction of the cost. Qwen4-Max's one-point advantage on the investigation task shows promise, but these single-run tests cannot validate Alibaba's broader agent claims or establish a definitive coding performance ranking.