Alibaba's new flagship AI model is cheap to use but trails US leaders on an independent test
Alibaba unveiled Qwen4-Max on Sunday, Reuters reported. It is the company's new top-of-the-line AI model and the successor to Qwen3.8-Max from August. Alibaba pitches it as a cheaper, more efficient engine for professional and scientific work. Its list prices undercut those of the leading American models.
For buyers, the question is whether a low price buys enough capability. The first independent results suggest a real upgrade over Alibaba's last model, but not a match for the best American systems. And a low price per token does not always mean a cheap job.
What Alibaba is selling
Qwen4-Max is available only as a paid service, through Alibaba's QwenCloud and its own apps, including QwenWork, QwenStudio and Qoder. Unlike many Qwen models, its weights (the trained model files) are not available to download. Alibaba says smaller Qwen4 models will be released free later, without giving a date.
Alibaba lists the model at $1.50 per million input tokens and $4.50 per million output tokens. Tokens are the word fragments AI services bill by. Text the system has recently seen and stored, known as cached input, costs $0.075 per million tokens. Prices differ by region and by plan, so buyers should check the specific deployment they would use.
The company says the model has 2.4 trillion parameters, a rough measure of size, but uses only about 71 billion for any given step, a design meant to keep running costs down. Alibaba says it also reads up to about a million tokens at once, roughly several long books, and understands images and video.
Alibaba's headline efficiency claim comes from Work360, a test it designed itself. On that test, it says average cost per task fell 62%, from $4.43 to $1.67, compared with Qwen3.8-Max. It says the new model used 24% to 35% fewer tokens, depending on how hard it was set to think. None of those figures has been checked by an outside party.
Alibaba also stressed safety. It says it kept training data for defensive security work, such as finding and fixing software flaws, and removed data meant to improve hacking. On one secure-coding benchmark, Alibaba reports the pass rate rising from 22% to 31.8%. The company says internal testing found no capabilities posing a large-scale cyber risk.
What independent testing shows
Artificial Analysis, an independent firm that runs the same battery of tests on major models, scored Qwen4-Max at 43 on its Intelligence Index. That is nine points above the latest version of Qwen3.8-Max, a large gain for one generation. It places the model 38th of 81 ranked models, just above a median of 41.
The leaders sit well above that. On the same index, at their highest settings, Anthropic's Claude Fable 5.2 scores 65 and Claude Opus 5.5 scores 59. OpenAI's GPT-6 Astra scores 57 and GPT-6.1 Sol scores 55. Qwen4-Max is 16 points behind Opus 5.5.
Alibaba also trails a domestic rival. Moonshot's Kimi K3.1 Pro, a Chinese rival, scores 46 on the same index. Techzine, citing Bloomberg's report on Alibaba's announcement, said Qwen4-Max reportedly outperforms Kimi K3.1 Pro. The independent index does not bear that out.
Some online claims that Qwen4-Max beats Opus 5.5 do not survive scrutiny either. A review by OrcaRouter, an AI routing service, found the comparison works only against Opus at its lowest effort setting. OrcaRouter also notes that Alibaba's own launch post compared the model only with its predecessor.
Why a cheap price isn't always a cheap job
Price per token is only half of the bill. The other half is how many tokens a model burns to finish a task. Artificial Analysis found Qwen4-Max produced 127 million output tokens across its test suite, against a median of about 79 million. The firm labels the model very verbose.
That shows up in its cost estimates. Artificial Analysis puts Qwen4-Max at about $3.27 per test task. That is roughly a third of Qwen3.8-Max's $10.02, a genuine saving for existing Alibaba customers. But OpenAI's GPT-6.1 Sol, which scores 12 points higher, costs about $1.27 per task, despite list prices of $2 and $10 per million tokens.
Against the most capable systems, the math looks different. On Artificial Analysis's estimates, Opus 5.5 costs about $10 per task and Fable 5.2 about $26.50. For companies that need near-top performance at a fraction of those prices, Qwen4-Max is a credible option, but not the only one.
There are caveats on both sides. Artificial Analysis graded Qwen4-Max at a single setting for how hard it thinks, while Alibaba lets customers choose, so everyday results and costs may differ. Speed data was not yet available at launch. And one company's test suite is not your workload.
The strategic picture is clearer. Alibaba is keeping its best model behind a paywall while promising to give away smaller ones. It is competing on price with OpenAI's lower-cost models and with Chinese rivals at once. The first independent numbers show it closing the gap on its own past, not yet on the leaders.