From Preview to Pay-Per-Token: Qwen4-Max Goes Live
Five days ago, Alibaba showed the world its most powerful AI model yet, then handed it a 14-day expiry date. Qwen4-Max arrived on November 11 as a preview you could only reach through a subscription plan.
It came with no price card, no published benchmarks and no promise it would still be there after November 25. On Monday, the countdown ended early. The model is now generally available on Alibaba Cloud, with a public API and a published price.
Alibaba says the numbers put it within reach of the best models in the world. The switch matters because the preview left the biggest questions unanswered. Developers could try the model inside Alibaba’s own apps, but they could not buy it by the token.
They could not see what it cost to run in production, and they could not compare it against anything else. The general release answers all three, and the answers are cheaper than the model it replaces.
What Qwen4-Max Actually Is
Qwen4-Max is the flagship of Alibaba’s new Qwen4 family, and it is built on a scale that puts it among the largest models ever offered through a public API. It is a sparse mixture-of-experts model with 2.4 trillion parameters in total, of which 71 billion activate for any given token.
A further 328 billion parameters sit in N-gram embedding layers, where they are looked up rather than multiplied. Three Gated DeltaNet layers compress the model’s history into a fixed state for each layer of sparse attention.
What that engineering buys is a context window of one million tokens. The model takes text, images and video as input and returns text, with reasoning, tool calling and structured outputs built in.
It is aimed squarely at long-running, tool-heavy work: codebase analysis, document extraction, interface generation and research agents that keep state across many steps.
There is one change from the preview that developers will want to note. The shared budget for thinking and the final answer is now 262,144 tokens, down from 393,216 on the preview card.
Non-thinking output caps at 131,072 tokens, with an input ceiling of 991,808 tokens. In thinking mode the input ceiling drops slightly, because the reasoning and the answer draw from the same bucket.
The Price Card: 25 Percent Cheaper Than Qwen3.8-Max
The headline number is $1.50 per million input tokens and $4.50 per million output tokens in Singapore, flat across the full one-million-token window. Cached input reads cost $0.075 per million, and the other five regions list $1.413 and $4.24.
Against Qwen3.8-Max, the model Qwen4-Max replaces, that is a straight 25 percent cut on both sides of the meter in Singapore. The model is live in six regions: Beijing, Singapore, Frankfurt, US Virginia, Tokyo and Hong Kong.
It speaks four protocol surfaces — OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope — so existing code written for those APIs can point at it with minimal changes. Batch inference is available, but only in Beijing, a limitation worth knowing before planning a large job around it.
Alibaba’s own arithmetic says the savings run deeper than the price cut. On Work360, its internal benchmark of industry tasks across 26 domains, Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max.
It does so while using roughly 24 percent fewer tokens at the highest effort setting and 35 percent fewer at high effort. Put the two together and the average cost per task falls from $4.43 to $1.67, a 62 percent reduction. Those are Alibaba’s numbers, and they should be read that way, but the direction is clear enough.
What the Benchmarks Say — and What They Don’t
According to The Business Times, Qwen4-Max scored near the top of coding benchmarks, outperforming Anthropic’s Claude Fable 5.1 and homegrown rival Moonshot’s Kimi K3.1 Pro. Independent leaderboards published their first results on launch day, and they tell a more measured story.
On Artificial Analysis, Qwen4-Max lands at 43 on the Intelligence Index at its highest reasoning effort. That is nine points ahead of Qwen3.8-Max but three behind Kimi K3.1 Pro at 46, at about $3.27 per index task.
The gap widens on Terminal-Bench 4.0, where Artificial Analysis puts Qwen4-Max at 48 percent against 67 percent for Kimi K3.1 Pro. On the Code Arena WebDev leaderboard it sits at number six with a preliminary 1,750 points, 74 above Qwen3.8-Max.
Leaderboard scores this early can move as more votes and runs come in, and none of this is a final verdict. What it suggests so far is a model that is strong where execution, iteration and tool use matter, and less dominant as a general planner than the very top frontier systems.
That is a useful distinction for anyone deciding where to spend their tokens. A model that is cheaper per task and good at long, tool-heavy work does not need to win every benchmark to be the right choice for agents that run for hours.
Alibaba appears to be positioning Qwen4-Max exactly there, as a workhorse rather than a showpiece.
Where You Can Use Qwen4-Max Now
The model identifier is now simply qwen4-max. The preview strings, qwen4-max-preview and qwen4.0-max-preview, are documented for retirement at midnight on November 25, Beijing time, so any configuration still pointing at them is on a nine-day clock.
The Token Plan that gated the preview is gone as the only route in; per-token metering is the standing model. Access runs through QwenCloud, and the model also appears in Alibaba’s own apps, including QwenWork, Qoder and Qwen Chat. OpenRouter lists it as well.
For developers, the practical test is straightforward: point an existing agent at the new model ID, watch the token counts on a real task and compare the bill against whatever it was running before.
The Catch: No Open Weights, and a Smaller Output Budget
Two things the launch does not do are worth stating plainly. Qwen4-Max itself is not open weight, so nobody can download it and run it on their own hardware. Alibaba says smaller versions of the Qwen4 series will follow, and those are the ones expected to carry open weights, but no date has been given.
The Qwen family remains the most downloaded open-weight model family in the world, which makes the flagship staying closed a deliberate commercial choice rather than an accident.
The second catch is the shrunken output budget. A shared thinking-plus-answer limit of 262,144 tokens is still enormous, but it is a third smaller than the preview offered, and long reasoning chains will hit it sooner. Teams building agents that think for a long time before answering should test where that ceiling actually bites.
Why This Launch Matters
Qwen4-Max did not arrive alone. On the same day, Tencent launched Hy4 flash, a 224-billion-parameter model with 12 billion active per token, priced at $0.070 per million input tokens and $0.281 per million output.
That is a 21-fold gap on input between two models that both claim native vision and both accept a million tokens of context. The Chinese model race is now being fought on price and efficiency as much as on capability, and the gap between a flagship and a value model has never been wider.
For Alibaba, the stakes are concrete. The company has spent years building the Qwen family into the default open-weight choice for developers worldwide, and Qwen4-Max is its bid to keep those developers when the work moves from downloading weights to buying API calls.
A flagship that costs 25 percent less than its predecessor, uses fewer tokens per task and scores near the top of coding benchmarks is a credible pitch. Whether the independent leaderboards keep it there is the question the next few weeks will answer, one token at a time.
Source: The Business Times, November 16, 2026.



