Alibaba released Qwen4-Max to Token Plan subscribers on November 11 in preview mode, followed by general availability on November 16. The model has earned attention for its verbose output: when testing firm Artificial Analysis ran its benchmark suite, Qwen4-Max produced 127 million output tokens—roughly double the median for competing models. Claude Opus 5.5 generated 218 million output tokens under maximum effort with default fallback, but that configuration differs from its standard medium-effort setting.
Pricing favors Qwen4-Max significantly. Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens, compared to Anthropic's $4 and $20 rates respectively. Cached input pricing shows similar gaps: Qwen4-Max costs $0.075 per million cached tokens versus Opus 5.5's $0.20. The volume of tokens Qwen4-Max generates does not translate to higher bills because of these low rates.
Cheaper only wins if you don't give up accuracy to get it.
Benchmarks published by Alibaba show mixed results against Opus. On agent tasks—where models work through problems step-by-step using tools—Qwen4-Max shows wins and losses. Coding performance is similarly uneven: Qwen4-Max trails on Terminal-Bench 4.0 with 48.5% versus Opus's 59.6%, but leads on DeepSWE v1.1. These published results use different evaluation setups and are not direct controlled comparisons.
To assess whether Qwen4-Max serves as 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 scoring script graded responses against prepared answer keys, and GPT-5 served as a baseline to ensure tests were sufficiently difficult.
Test scenarios
- Multi-app workflow (76 points): Models acted as agents across a fake company's ticket tracker, code repository, chat, 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 overlapping outages occurring simultaneously. The dataset included misleading deploys, two servers with incorrect time settings, and a deceptive alert.
- Agent bug fix (47 points): Models debugged an accounting library containing 11 intentional bugs. Six bugs produced no visible test failures, requiring comparison with documentation. Three of the library's own tests were incorrect, and models should report them without modification.
Multi-app workflow results
Opus 5.5 achieved a perfect score of 76 points. Qwen4-Max scored 71, with all five missed points stemming from a single error. When three questions asked about code ownership, Qwen4-Max selected the most specific matching rule rather than following the repository's documented rule that the last matching entry takes precedence. All other work, including timezone-aware deadline calculations and both required refusals, was correct.
…Qwen4-Max can handle a long, messy office job, but one misread rule followed it through the whole task.
In practical terms, this mistake would route tickets to incorrect teams in a real organization. Qwen4-Max generated nearly twice the output tokens as Opus but cost less than one-third, though it required more than three times longer to complete with roughly half as many conversational turns.
- Opus 5.5: 44 turns, 2,451,342 input tokens, 26,617 output tokens, $2.13, 248 seconds
- Qwen4-Max: 25 turns, 1,746,170 input tokens, 51,264 output tokens (41,644 reasoning), $0.62, 791 seconds
Incident investigation results
Qwen4-Max narrowly outperformed on this task, scoring 59 of 60 points versus Opus 5.5's 58. Both models successfully disentangled 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, misattributing a customer-facing error code to the wrong service—Opus made this mistake twice, Qwen4-Max once.
This represented Qwen4-Max's strongest performance. It achieved one additional point over Opus while costing approximately 31% of the price. However, Opus 5.5 demonstrated superior speed, completing the investigation in under 4 minutes compared to Qwen4-Max's 9.5 minutes. A single test run cannot establish broader advantages in agent-based investigation work.
- Opus 5.5: 22 turns, 1,216,769 input tokens, 23,642 output tokens, $1.41, 221 seconds
- Qwen4-Max: 18 turns, 1,104,031 input tokens, 33,036 output tokens (23,153 reasoning), $0.44, 570 seconds
Agent bug fix results
Neither model achieved perfect accuracy on this coding task. Opus 5.5 scored 45 of 47 points while Qwen4-Max scored 40. This result aligns with Opus's published advantage on Terminal-Bench 4.0, though Alibaba's own 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 the same test and lost five additional points. It fixed 10 of 11 bugs, missing one involving payment splitting across invoices that loses a cent. It also left a scratch file in the test folder, which violated the rules.
Tool logs revealed different working patterns. Opus made its first fix on turn 4 and completed the task in 12 turns. Qwen4-Max spent 15 turns reading files before making its first edit and required 30 total turns. The additional reading time did not improve the score. The cost gap narrowed on this test: Qwen4-Max consumed more than three times the input tokens, resulting in a bill approximately two-thirds of Opus's cost.
- Opus 5.5: 12 turns, 228,824 input tokens, 6,433 output tokens, $0.38, 64 seconds
- Qwen4-Max: 30 turns, 827,545 input tokens, 15,813 output tokens (6,701 reasoning), $0.25, 310 seconds
Overall findings
Qwen4-Max's reputation for verbose output held true across testing. The model generated 77% more output tokens than Opus across all three tasks. These totals include reasoning tokens and reflect different tokenization methods between models, so they do not measure only visible answer length.
What Qwen4-Max really costs is time and accuracy.
Despite generating substantially more tokens, Qwen4-Max's low pricing meant the total cost remained approximately one-third of Opus's. The actual trade-off involves time and accuracy. Qwen4-Max required 3.14 times longer overall and scored 170 of 183 points compared to Opus's 179—a nine-point deficit.
Recommendations
Since each model ran only once, small score differences could shift in subsequent runs. Before deploying Qwen4-Max to long-running agent jobs, testing on internal workloads is advisable. Answers should be verified against source material, and code must undergo testing and human review; agreement from a second model does not establish correctness.
Opus 5.5 remains the better choice for coding tasks and any work where incorrect answers create downstream problems. It demonstrated higher accuracy on two of three tests and completed all tasks in roughly one-third of Qwen4-Max's total time.
Qwen4-Max has earned consideration for investigation work involving large data volumes where additional processing time is acceptable. Its one-point advantage on the investigation task is promising, but single-run testing cannot validate Alibaba's broader agent capabilities or establish definitive coding performance rankings.