Qwen4-Max: 7 Essential Facts and 1 Big Catch

Alibaba's Qwen4-Max launched on 16 November at $1.50 input and $4.50 output per million tokens. Seven facts on its architecture, benchmarks, independent rankings and the cost-per-task catch.

This guide is based on Alibaba’s launch materials, QwenCloud pricing and independent evaluations from Artificial Analysis and Vals as reported in November 2026. I have not run Qwen4-Max on my own tasks, so every benchmark is attributed and vendor numbers are labelled.

Qwen4-Max is the cheapest serious flagship model you can call today. Alibaba released it on 16 November 2026 at $1.50 per million input tokens and $4.50 per million output, roughly a tenth of GPT-6 Astra’s output price and well under half of Kimi K3.1 Pro’s.

The catch is the one that always comes with cheap tokens: Qwen4-Max uses a lot of them. Independent testers call it slow and verbose, and on Vals’s index, OpenAI’s GPT-6.1 Sol was both cheaper per test and slightly more accurate. So Qwen4-Max is a genuine bargain for some workloads and a false economy for others. Here are the seven facts that decide which side you are on.

Table of Contents

Qwen4-Max cost per task chart
Price per token is the headline. Cost per task is the bill, and Qwen4-Max uses far more tokens than its rivals.

Fact 1: It Is Closed, Not Open

Qwen has a reputation for open weights, so this surprises people. Alibaba’s new flagship is a closed model. Alibaba has not announced open weights for Max, though DataCamp reports the company says smaller Qwen4 models will be open-sourced later.

The launch came in two stages. A preview, qwen4-max-preview, opened on 11 November through Alibaba’s subscription Token Plan only, with no per-token price. The full model, qwen4-max, followed on 16 November with pay-as-you-go pricing, and the preview identifier retires on 25 November.

If you were hoping to download and self-host Alibaba’s best model, that is not on offer. For downloadable options this month, see our roundup of the best local LLMs, which covers what actually fits on your hardware.

Fact 2: The Price Undercuts Every US Flagship

Here is QwenCloud’s international list price against the models it competes with, as compiled by DataCamp.

Model Input per 1M tokens Output per 1M tokens
Qwen4-Max $1.50 $4.50
GPT-6 Astra $10 $50
Claude Fable 5.2 $8 $40
Claude Opus 5.5 $4 $20
GPT-6.1 Sol $2 $10
Kimi K3.1 Pro $2 $10
Gemini 4 Carbon $2 (rising to $4 on 1 Dec) $10 (rising to $20)
Grok 4.7 $2 $6

Caching makes it cheaper still. Cached input and explicit cache reads cost $0.075 per million tokens, a 70% drop from the previous generation’s $0.25, according to XenoSpectrum. Writing to the cache costs $1.875 per million for a five-minute window or $3.00 for an hour. The rate limit is five million tokens per minute, and new accounts get one million free tokens in the Singapore region for about 90 days.

For subscription users, third-party reports list Token Plan tiers from $6 to $43 a month, with promotional prices of $3 and $5 on the two cheapest tiers through 30 November. The personal plan is Singapore-only.

Fact 3: A 2.4T Model With 71B Active

Qwen4-Max is a sparse mixture-of-experts model with a 2.4-trillion-parameter backbone and about 71 billion parameters active per token. It also carries an extra 328 billion N-gram embedding parameters, which is an unusual design choice.

Alibaba describes three new components: a hybrid attention scheme it calls GDN plus QSA, combining Gated DeltaNet layers with sparse attention; a Gated Residual design with parallel paths per layer; and the N-gram Embedding. The point of the hybrid attention is to keep a one-million-token window affordable, by compressing older context into fixed-size memory while sparse layers pull out only the relevant parts.

The company says the 2.4T training run had no loss spikes, and that the base model reaches 5 to 13% lower bits-per-byte than Qwen3.8-Max-Base while activating 25% fewer parameters. Those are in-house numbers, and DataCamp notes the comparison is not strictly like for like. Qwen4-Max takes text, image and video in, and produces text.

Fact 4: One Million Tokens, With Conflicting Output Limits

The context window is one million tokens. Maximum input is 991,000 tokens, or 983,000 with thinking enabled. Thinking is on by default, and the model can spend up to 262,000 reasoning tokens drawn from its output budget.

The output limit is where the documentation disagrees with itself. The model page lists 393,000 tokens in both modes, while the API docs list 131,072 without thinking and 262,144 with it. If your workload depends on very long outputs, test the real ceiling before you build around either number.

That default thinking matters for cost. Reasoning tokens bill as output, and Qwen4-Max thinks a lot. Turning thinking off for simple tasks is the single easiest saving available.

Fact 5: It Wins Only Some of Alibaba’s Own Benchmarks

Even in Alibaba’s own launch chart, Qwen4-Max took first place in only four of 16 benchmarks, and Claude Opus 5.5 beat it on many of the coding tests, according to XenoSpectrum. Here are the most revealing rows, all vendor-reported.

Benchmark Qwen4-Max Best rival listed
Agents’ Last Exam v1.1 46.7 Opus 5.5 and GPT-6 Astra, 43.4
BioMysteryBench 1.1 85.0 (top) Below 85.0
DeepSWE v1.1 75.7 Five models within 1.6 points
Terminal-Bench 4.0 48.5 Kimi K3.1, 67.2
ProgramBench 31.5 Opus 5.5, 65.0
Humanity’s Last Exam 51.7 Opus 5.5, 59.8
Chartography 54.2 GPT-6 Astra, 71.1

The pattern is clear. The model is strong on broad agent and knowledge-work tasks and weak on terminal coding, program synthesis and visual chart reading. That shapes where it belongs in a stack far more than its headline price does.

Alibaba also reports big efficiency gains over Qwen3.8-Max, including around 48 to 49% fewer tokens on Terminal-Bench and an average Work360 cost per task falling from $4.43 to $1.67. DataCamp says it could not fully reconcile that last figure, so treat it as a claim.

Fact 6: Independent Tests Put It Ninth

Outside Alibaba’s chart, the picture is more modest. Artificial Analysis gives Qwen4-Max an Intelligence Index score of 43, ninth overall, at an average of $3.27 per task. That compares with $10.04 for Opus 5.5 and $1.27 for GPT-6.1 Sol. It also recorded 127 million output tokens across the index, against a median of 78.6 million, which is the verbosity problem in a single number.

Vals ranks it 11th of 38 on its index at 63.20%, the highest score of any Chinese model it has tested. Its Terminal-Bench run of 48.48% matches Alibaba’s figure, which is a good sign for the honesty of at least that number. The full Vals test cost $9.33, against $24.77 for Opus 5.5 and $3.61 for Sol, which scored 63.94%. Each test took about an hour and four minutes.

On specialist indices, Vals places it eighth of 41 for finance and 17th of 39 for legal. Testers describe it as slow and prone to overthinking simple tasks, though one reported it finishing a task faster than MiniMax M3.1 Flash.

Fact 7: Access Is Easy if You Already Use OpenAI or Anthropic APIs

You can call qwen4-max on QwenCloud through the standard API or Token Plan, using OpenAI-compatible and Anthropic-compatible endpoints. That means most existing code needs only a new base URL, key and model name. It is also available in QwenStudio (Qwen Chat), QwenWork and Alibaba’s Qoder coding product.

Alibaba released a small open-source companion, the Qwen Live Delivery plugin, under Apache 2.0. At version 0.1.0 it supports only Codex CLI, on Linux and macOS with Node.js 22 or later, so it is an early experiment rather than a reason to switch.

Before routing production traffic, check data residency. The free quota and personal Token Plan are tied to the Singapore region, and enterprise users should confirm where their requests are processed and what retention terms apply.

Where Qwen4-Max Sits in the Qwen Family

Alibaba first sketched a four-tier Qwen 4 line-up at its Apsara conference in September: Max, Plus, Flash and a dedicated 27B variant, according to analysis published by Yotta Labs. Max is the first to ship, and it is the only one so far with a public price.

That ordering is a change of habit. Earlier Qwen generations often arrived as open-weight models first, with the biggest hosted version following. This time the company led with a closed flagship and promised the smaller, downloadable models later. Read that as a sign of where the money is: the hosted top tier is what Alibaba sells, and the open models are what keep developers loyal.

For existing Qwen users, the practical upgrade is from Qwen3.8-Max. Alibaba claims its successor beats it while using up to around half the tokens on some tasks, and that the average cost per task on its Work360 benchmark fell by more than 60%. One independent tester rated the new model below Qwen3.8-Max on their own tasks, so run your own comparison before switching everything over.

If you rely on Qwen for local deployment, nothing changes yet. The strongest downloadable Qwen weights are still in the Qwen3.8 line, and the open Qwen4 models are a promise, not a release. Plan around what you can download today.

What the Launch Means for the Price War

The bigger story is pressure on everyone else. Flagship output prices now range from $50 per million tokens for GPT-6 Astra down to $4.50 for Alibaba’s new model. That is an elevenfold spread for models that sit within a few places of each other on independent leaderboards.

Western labs have already started moving. Anthropic cut Fable’s price by 20% with version 5.2, and Google’s Gemini 4 Carbon launched with introductory pricing, though that rate doubles on 1 December. Chinese labs have pushed in the other direction for most of the year, competing hard on price while closing the capability gap.

For buyers, this is good news with a footnote. The footnote is cost per task. As this launch shows, a model can be cheap per token and still expensive per job if it thinks at length. The labs know this, which is why every launch now leads with token efficiency as well as price.

My expectation is that by early 2027, list prices will matter less than they do today and published cost-per-task numbers will matter more. Independent benchmarks that report cost, such as Artificial Analysis and Vals, are already where the most useful comparisons live. Bookmark them, and check them before any price table.

The Big Catch: Cost per Task

Low token prices do not guarantee low bills. This model is cheap per token and expensive in tokens, and those two forces partly cancel out.

The Vals numbers show it plainly. Qwen4-Max cost $9.33 to run the full test and scored 63.20%. GPT-6.1 Sol cost $3.61 and scored 63.94%. Sol charges more per token and still came out cheaper and slightly better, because it used far fewer of them. Artificial Analysis tells the same story: $3.27 per task for Qwen4-Max against $1.27 for Sol.

Where it does save money is against the expensive end of the market. At $3.27 per index task versus $10.04 for Opus 5.5, it is about a third of the cost for a lower but respectable score. If your current default is a premium model and your tasks are broad knowledge work rather than terminal coding, that is a real saving.

The fix for verbosity is the same as for any thinking model: turn reasoning down or off where it is not needed, cap output length and cache long shared prompts. Our guide to LLM API cost optimisation covers those settings in detail.

Who Should Use Qwen4-Max

If you are Start with Why Link
Running broad agent or office tasks on a premium model Qwen4-Max with thinking tuned down About a third of Opus 5.5’s cost per task on Artificial Analysis Alibaba announcement
Doing terminal-heavy coding Kimi K3.1 Pro or Opus 5.5 Qwen4-Max scores 48.5 on Terminal-Bench 4.0 against 67.2 and 59.6 AI coding assistants
Optimising pure cost per task GPT-6.1 Sol Cheaper per task and slightly higher on Vals Vals index
Working in biology or research Q&A Qwen4-Max, after your own test Top score on Alibaba’s BioMysteryBench DataCamp breakdown
Needing downloadable weights Wait for smaller Qwen4 models Qwen4-Max itself is closed Best local LLMs

Related on this site: for more context, see MiniMax M3 and Gemma 4 Seven Months On.

FAQ About Qwen4-Max

When was Qwen4-Max released?
The full model launched on 16 November 2026, after a Token Plan preview that opened on 11 November. The preview model ID retires on 25 November.

How much does Qwen4-Max cost?
$1.50 per million input tokens and $4.50 per million output tokens on QwenCloud’s international list, with cached input at $0.075. Token Plan subscriptions are also available.

Is Qwen4-Max open source?
No. It is a closed model. Alibaba says smaller Qwen4 models will be open-sourced later, but has not announced open weights for Max.

Is Qwen4-Max better than Claude Opus 5.5?
On most independent measures, no. Artificial Analysis ranks it ninth with a score of 43, and Opus 5.5 beats it on many coding benchmarks. It is much cheaper per task, though.

Why is Qwen4-Max slow?
Thinking is on by default and the model is verbose. Artificial Analysis recorded 127 million output tokens across its index, well above the 78.6 million median. Reducing or disabling thinking on simple tasks helps.

Can I use Qwen4-Max with my existing OpenAI or Anthropic code?
Yes. QwenCloud offers OpenAI-compatible and Anthropic-compatible endpoints, so most integrations need only a new base URL, API key and model name.

Verdict

Try Qwen4-Max if you are paying premium prices for broad agent and knowledge-work tasks. It costs about a third as much per task as Opus 5.5 on Artificial Analysis, it plugs into existing OpenAI or Anthropic code with a base URL change, and Vals ranks it the strongest Chinese model it has tested.

The caveat is that it is not the cost leader it looks like on a price table. It is verbose, slow and weak at terminal coding, and GPT-6.1 Sol was cheaper per task and slightly more accurate on Vals. Cheap tokens only help if you do not need many of them.

Your next step: claim the free million tokens, run twenty of your own tasks with thinking on and off, and compare cost per correct answer against your current model. If Qwen4-Max wins with thinking off, you have found a real saving.

Sources and Further Reading

For more model coverage here, read our breakdowns of GPT-6 Astra, Claude Fable 5.1 and GLM-5.3-Flash, or compare the field in AI models compared.