Qwen4-Max Is Here: Alibaba’s New Flagship AI Model Costs a Fraction of GPT and Claude

A glowing neural network orb with flowing data streams hovering above a laptop on a desk

Alibaba has a new flagship AI model, and it is priced to make the rest of the industry uncomfortable. Qwen4-Max went generally available on Monday, November 16, 2026, five days after a preview opened on Alibaba’s Token Plan subscription. It is the first model built on the new Qwen4 architecture, the successor to Qwen3.8-Max, and it arrives with a 1-million-token context window, native vision, and an API price of $1.50 per million input tokens and $4.50 per million output tokens — less than a quarter of what Anthropic charges for output tokens on its top model.

The headline is not just the sticker price. Alibaba says Qwen4-Max does the same work with up to 50% fewer tokens than its predecessor, and that the average cost of a task on its in-house Work360 benchmark fell from $4.43 to $1.67 — a 62% cut. Independent benchmarks tell a more measured story: the model scores 43 on the Artificial Analysis Intelligence Index, nine points above Qwen3.8-Max but still 16 points behind Claude Opus 5.5. Here is what Qwen4-Max actually is, what it costs, how it compares, and who should try it.

What Qwen4-Max Actually Is

Qwen4-Max is a mixture-of-experts model with 2.4 trillion parameters in total, of which 71 billion are active for any given token — the sparse design that lets a very large model answer at a manageable cost. It is multimodal from the ground up, accepting text, image and video input, and reasoning is switched on by default. The context window is 1 million tokens, with up to 393,216 tokens of output, which puts it in the same league as the longest-context frontier models for digesting whole codebases, long documents and extended agent sessions.

Alibaba positions it as a workhorse for coding, office work, front-end design and complex agent tasks rather than a chatbot showpiece. It is available through the QwenCloud API, and inside Alibaba’s own apps — Qwen Studio, QwenWork and Qoder. The model ID for the API is qwen4-max, and the API follows the OpenAI-compatible format, so existing tooling can point at it with minimal changes. That matters because the model is aimed squarely at developers and teams already running coding agents: it slots into the same workflows as tools like Cowork, which let Claude Code, Codex and Gemini CLI share one task queue, and it competes directly with the cheap, fast tier that Claude Haiku 5.5 now occupies for everyday apps.

How Much Qwen4-Max Costs — and What the 62% Really Means

Glowing price comparison bars in teal and orange on a dark navy background with floating coins, illustrating lower AI model pricing

On QwenCloud, Qwen4-Max costs $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million implicitly cached input tokens. Against its predecessor, Qwen3.8-Max, that is 25% cheaper on input and output and 70% cheaper on cached input ($2, $6 and $0.25 respectively). The bigger saving comes from token efficiency: because the model needs fewer tokens to finish the same job, Alibaba reports the average Work360 task cost dropping from $4.43 to $1.67.

Two caveats belong next to that 62% figure. First, it compares Qwen4-Max with Alibaba’s own older model, not with its rivals — it is a measure of Alibaba’s progress, not a head-to-head win. Second, independent cost-per-task measurements tell a subtler story: Artificial Analysis puts Qwen4-Max at $3.27 per task on its index, against $3.16 for Google’s Gemini 4 Hydrogen, so on that yardstick it is competitive rather than dominant. The honest summary is that Qwen4-Max is dramatically cheaper than its predecessor and roughly on par with the cheapest frontier options — a strong position for a flagship, just not the rout the headline number suggests.

Against the top tiers, the price gap is still striking. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, so Qwen4-Max’s output price is less than a quarter of Anthropic’s. Google’s Gemini 4 Argon and Carbon cost $2 and $10 respectively — and those prices are set to double on December 1. Even OpenAI’s budget-friendly GPT-6.1 Terra charges $5 per million output tokens, more than Qwen4-Max. For teams running agents all day, where output tokens are the expensive half of every bill, that gap adds up fast.

How Qwen4-Max Benchmarks Against GPT, Gemini and Claude

On the Artificial Analysis Intelligence Index, Qwen4-Max scores 43. That is nine points above Qwen3.8-Max at 34, three points behind Kimi K3.1 Pro, and 16 points behind Claude Opus 5.5 at 59. In other words, Alibaba’s new flagship is clearly the best model China has fielded on this index, and clearly not yet at the level of the best Western frontier models on raw capability.

Alibaba’s own benchmark sheet is more flattering, as vendor sheets usually are. The company published around sixteen evaluations in which Qwen4-Max wins most of them against its predecessor, with the biggest gains on professional and agentic work. The pattern is consistent with the pricing story: this is a model optimised for getting real work done cheaply, not for topping a leaderboard. It is the same trade-off shaping the whole market right now — OpenAI’s GPT-5.2 is pitched at professional work, DeepSeek is opening its app tools to developers at very low prices, and every vendor is fighting over cost per task rather than cost per token alone.

Security Guardrails: Trained to Defend, Not to Attack

One of the more unusual parts of the launch is how Alibaba talks about safety. The company says it filtered Qwen4-Max’s training data by purpose: material useful for defensive work, such as finding and fixing vulnerabilities, was kept, while data intended to improve attack capabilities was removed. Internal experts then reviewed long-running task sessions, including the model’s full chains of thought, and Alibaba concludes that “no capabilities posing large-scale cybersecurity risks have been observed.”

There is also a concrete number. On the Last Secure Code Bench, which tests whether generated code both works and is secure, Qwen4-Max’s pass rate rose from 22% to 31.8% over its predecessor. That is still a long way from safe-by-default, but it is a meaningful jump, and it fits a launch that leans heavily on coding and agent use. Two honest limits: these are Alibaba’s own assessments, and no external red team has tested the model yet — so treat the safety claims as a statement of intent backed by internal testing, not as an independent verdict.

Where to Try Qwen4-Max — and What Is Still Closed

A developer working at a desk with two monitors showing code and an AI assistant chat interface in a warm home office at dusk

The full model is available now through the QwenCloud API on a pay-as-you-go basis, and inside Qwen Studio, QwenWork and Qoder. The November 11 preview ran on the Token Plan subscription only and is scheduled to retire when the preview window ends on November 25, so the public API is now the way in. Because the API is OpenAI-compatible, developers can try it in Claude Code, Codex, OpenCode and similar agent tools by pointing them at the QwenCloud endpoint.

What you cannot do yet is download it. Qwen4-Max itself is proprietary and stays closed — Alibaba has not announced weights for the Max tier, and reports indicate it will remain hosted-only. Smaller Qwen4 models are expected to be open-sourced later, following the pattern of earlier Qwen releases, but there is no date for that. If your workflow depends on self-hosting or fine-tuning weights, Qwen4-Max is not your model; if you want a cheap hosted flagship for coding agents and long-context work, it is now one of the cheapest options with a 1M-token window.

Should You Switch to Qwen4-Max?

Switch or trial it if you run coding agents or long-context workloads at volume and your bill is dominated by output tokens — that is where the price advantage is largest, and the 1M-token context removes a real constraint for codebase-scale tasks. It is also worth a look if you already use Alibaba’s cloud or its Qwen apps, where the integration is native.

Hold off if you need the absolute best reasoning or agentic performance, where Opus 5.5 still leads by a wide margin on independent benchmarks, or if you need open weights. And remember the context: this launch is part of a pricing war, not a capability leap. The model to watch is not just Qwen4-Max but the response it forces — rivals have already been cutting prices and doubling context windows, and a flagship that is both capable and this cheap raises the floor for everyone.

Qwen4-Max does not dethrone the frontier leaders. It does something almost as disruptive: it makes flagship-class AI cheap enough that cost stops being the reason not to use it.

Frequently Asked Questions

How much does Qwen4-Max cost?

$1.50 per million input tokens and $4.50 per million output tokens through the QwenCloud API, with implicitly cached input at $0.075 per million tokens. That is 25% cheaper than Qwen3.8-Max on input and output, and 70% cheaper on cached input.

What is the Qwen4-Max context window?

1 million tokens, with up to 393,216 tokens of output. Reasoning is enabled by default.

When was Qwen4-Max released?

A preview opened on November 11, 2026 through Alibaba’s Token Plan subscription. General availability with a public pay-as-you-go API followed on November 16, 2026.

How does Qwen4-Max compare with Claude Opus 5.5?

On the Artificial Analysis Intelligence Index, Qwen4-Max scores 43 against Opus 5.5’s 59 — 16 points behind on capability. On price it is far cheaper: $1.50/$4.50 per million tokens against Opus 5.5’s $4/$20.

Is Qwen4-Max open source?

No. Qwen4-Max is proprietary and available only through Alibaba’s hosted services. Smaller Qwen4 models are expected to be open-sourced later, but Alibaba has not given a date.

Where can I use Qwen4-Max?

Through the QwenCloud API (model ID qwen4-max), and inside Qwen Studio, QwenWork and Qoder. The API is OpenAI-compatible, so it also works with agent tools such as Claude Code, Codex and OpenCode.

Sources: Qwen — Meet Qwen4-Max (official launch post, November 16, 2026); Digital Trends — Qwen4-Max is China’s newest flagship AI model (November 16, 2026); Artificial Analysis — Qwen4-Max model page (Intelligence Index 43); SiliconNoon — Qwen4-Max vs Opus 5.5 deep dive (November 16, 2026).

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top