Categories AI and Tools

Qwen4-Max Is Live: Alibaba’s Flagship Is 25% Cheaper

Alibaba’s Qwen4-Max is out of preview and on general sale as of November 16, 2026 — and the headline is the price. The new flagship costs $1.50 per million input tokens and $4.50 per million output tokens on QwenCloud, a straight 25% cut from the Qwen3.8-Max it replaces. But the launch comes with a twist every freelancer and small team should know before switching: Alibaba’s own benchmark chart puts Qwen4-Max level with Claude Opus 5.5, while the independent Artificial Analysis index scores it 43 — 16 points behind Opus 5.5. Both numbers are real. They just measure different things, and the gap between them is the whole story.

This is the freelancer’s decision piece on the launch: the exact prices, the worked monthly math against Opus 5.5, Kimi K3.1 Pro, Claude Fable 5.2 and GPT-6 Astra, the honest benchmark-disagreement box, and the “hosted only, no weights” warning that matters if you self-host.

What launched on November 16

Qwen4-Max is the first full model in Alibaba’s Qwen4 family and the successor to Qwen3.8-Max. It arrived five days after a Token Plan-only preview opened on November 11, and the production model is a different thing from that preview in one important way: it runs on ordinary pay-as-you-go Model Studio keys, with a one-million-token free quota valid for 90 days. The model ID is qwen4-max.

The specs, as Alibaba lists them:

  • 2.4 trillion parameters in the backbone, with 71 billion activated per token (a sparse mixture-of-experts design), plus 328 billion N-gram embedding parameters.
  • 1 million tokens of context, with a maximum input of 991,808 tokens without thinking and 983,616 with thinking switched on. Maximum output is 131,072 tokens without thinking, and a 262,144-token budget shared between reasoning and the final answer when thinking is on.
  • Native multimodality: text, images and video go in; text comes out.
  • Availability: QwenCloud Model Studio in six regions — Singapore, Beijing, Hong Kong, Frankfurt, US Virginia and Tokyo — with OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope APIs.

Alibaba positions it as a long-horizon agent model rather than a chat model. Most of the launch material is devoted to multi-hour autonomous runs in engineering, finance, law and drug discovery — the kind of work where a model chews through a huge context window and produces long outputs. That positioning matters for the price math below, because long agent runs are exactly where output tokens (the expensive side) pile up.

One date to put in your calendar now: the preview identifiers (qwen4-max-preview and qwen4.0-max-preview) stop working at 24:00 on November 25, 2026 (UTC+8). If you tested the preview last week, your code needs the production model ID before then. Alibaba’s own preview documentation page now recommends the production version in its place.

The price, in full

QwenCloud’s rate card (Singapore region) lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens. Cached input — tokens the provider has already processed — costs $0.075 per million, a 95% discount off the input rate. The other five regions list slightly lower prices: $1.413 input and $4.24 output. The same card applies whether thinking mode is on or off. The rates above come from Alibaba Cloud’s Model Studio pricing page — check it before you commit, since list prices can change.

Compare that with the model it replaces. Qwen3.8-Max listed at $2.00 input and $6.00 output in Singapore, so Qwen4-Max is exactly 25% cheaper on both sides of the meter. Alibaba also reports that on its own Work360 benchmark, the average cost per task fell from $4.43 to $1.67 — about 60% lower — and that the model uses 35% fewer tokens for the same work. Those are Alibaba’s own numbers, so treat them as the vendor’s claim, not an independent measurement. The independent cost-per-task number tells a more complicated story, and I’ll get to it below.

Two more pricing details worth knowing. First, the free quota: one million tokens on QwenCloud, available in the Singapore region for about 90 days, which is enough to run a real evaluation before you spend anything. Second, a rumor to file under “unconfirmed”: one source claims Qwen4-Max prices will double on December 1. I have not seen a second source for that, and Alibaba’s own pages list the $1.50/$4.50 card as the standing rate — so don’t plan around a doubling, but do check the rate card again before you commit a big workload to it in December.

The price table: Qwen4-Max vs the flagships

Per-token prices only tell you the sticker. Here’s how Qwen4-Max lines up against the models a freelancer would actually compare it with, using published standard rates checked in November 2026:

Model Input / 1M tokens Output / 1M tokens Cached input / 1M
Qwen4-Max (Alibaba) $1.50 $4.50 $0.075
Kimi K3.1 Pro (Moonshot) $2.00 $10.00 $0.20
Claude Opus 5.5 (Anthropic) $4.00 $20.00 $0.20
Claude Fable 5.2 (Anthropic) $8.00 $40.00 —
GPT-6 Astra (OpenAI) $10.00 $50.00 $1.00

On the sticker, Qwen4-Max undercuts every Western flagship on both input and output. It is roughly a third of Opus 5.5’s price, a little over a tenth of GPT-6 Astra’s, and less than half of Kimi K3.1 Pro’s output price. If you only ever looked at the rate card, the decision would be easy. It isn’t, because the sticker is not the bill.

The monthly math: what a real workload costs

Let me work the numbers for a realistic freelancer workload — say, 10 million input tokens and 2 million output tokens a month, all uncached, at standard rates. That’s roughly a small agency running a coding agent and a document pipeline every working day:

Model Input cost Output cost Monthly total
Qwen4-Max 10 × $1.50 = $15 2 × $4.50 = $9 $24
Kimi K3.1 Pro 10 × $2 = $20 2 × $10 = $20 $40
Claude Opus 5.5 10 × $4 = $40 2 × $20 = $40 $80
Claude Fable 5.2 10 × $8 = $80 2 × $40 = $80 $160
GPT-6 Astra 10 × $10 = $100 2 × $50 = $100 $200

Qwen4-Max costs $24 a month for work that runs $80 on Opus 5.5 and $200 on GPT-6 Astra. That is a genuine, large saving — if the model does the work to the same standard. And that “if” is where the launch gets interesting, because the independent benchmarks say it doesn’t, at least not yet.

Two things soften the sticker gap in practice. First, caching: in real agent traffic, a large share of input tokens are repeats of system prompts and conversation history. OpenRouter’s live provider data shows Alibaba Cloud International’s effective blended input price at $0.278 per million, not $1.50, because 85.9% of input tokens are being served from cache. Qwen4-Max’s 95% cache discount is the deepest in the table, so heavy-caching workloads benefit more from Qwen than the sticker suggests. Second, the reverse: Qwen4-Max’s output price has no cache discount, and long agent runs generate a lot of output. If a model is verbose — and early testers describe Qwen4-Max as slow and verbose — the output side eats the input savings.

The benchmark fight: 75.7 vs 43

Here is the part of the launch you need to read carefully, because two honest measurements disagree sharply.

Alibaba’s own chart puts Qwen4-Max level with Claude Opus 5.5 on DeepSWE v1.1, a coding-agent benchmark: 75.7 to Opus 5.5’s 74.2. On that chart, Qwen4-Max beats Opus 5.5 and GPT-6 Astra on several tests, and Alibaba’s announcement says it scored near the top of coding benchmarks, ahead of Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro. Bloomberg reported that framing. But the comparison table Alibaba published lists other rivals, and the model trails on most agentic-coding rows in it — the “outscored Fable 5.1” claim is, at best, selective, and the benchmark, harness and effort setting behind it have not been published.

Artificial Analysis, the independent harness that ran the model itself within hours of launch, scores Qwen4-Max 43 on its Intelligence Index (v5.0.1), which combines nine tough evaluations. That puts it 12th overall: 3 points behind Kimi K3.1 Pro (46), 16 behind Opus 5.5 (59) and 22 behind the leader, Claude Fable 5.2 (65) — a model Alibaba left off its own chart. It is also 9 points up on Qwen3.8-Max (34), so the generational improvement is real; it’s just not a leap to the top tier. On Terminal-Bench 4.0 specifically, Qwen4-Max scores 48%, behind Kimi K3.1 Pro at 67% and ahead of Tencent’s Hy4 at 37%. Vals ranked it 11th.

The cost-per-task numbers make the trade-off concrete. Artificial Analysis measures Qwen4-Max at $3.27 per Intelligence Index task, against $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. That is a real improvement — a third of its predecessor’s cost per completed task. But GPT-6.1 Sol at medium reasoning scores 46 on the same index at just $0.36 per task. In other words, a model that scores three points higher costs roughly a ninth as much per completed unit of work. Cheap tokens are not the same as cheap work, and Qwen4-Max’s verbosity is the reason: the evaluation generated 127 million output tokens, and at $4.50 per million, output spend dominates the bill.

The community reaction was quick and skeptical. The Hacker News thread on the launch drew 848 points and 447 comments within a day, and the top-voted observations cut both ways: one commenter noted the model “scores a 43 on Artificial Analysis with an average cost of $3.27” against GPT-6.1 Sol’s 46 at $0.36, while another pointed out that the Artificial Analysis charts show maximum effort by default, “which isn’t how most people use them,” and that on the cost-per-task pareto chart “Qwen4-Max is not in a good spot.” A third tester reported that in a short-horizon data-analyst exam, Qwen4-Max “scored a few points worse than Qwen 3.8 27B from August (which is open weights) and cost a bit more.”

My read: the truth sits between the two charts. Alibaba’s chart is vendor-compiled and selective; the independent index is real but measures maximum-effort settings that don’t match how most people run a model day to day. What both agree on is that Qwen4-Max is a big step up from Qwen3.8-Max and a long way from the frontier leaders on hard reasoning and agentic coding.

The “no open weights” warning

This is the change that matters most to a specific group of readers, and it’s the one most launch coverage buries. Qwen4-Max’s weights are not open. It is served through QwenCloud’s API (plus QwenWork, Qwen Studio, Qoder, OpenRouter and the Vercel AI Gateway) only. Fine-tuning and model deployment are both marked unsupported in Alibaba’s documentation. Smaller Qwen4 variants are promised as open source later, with no date.

That is a break from Qwen’s tradition. The Qwen family built its reputation on open weights — developers could download the model, run it on their own hardware, fine-tune it and inspect it. Qwen4-Max keeps the name and drops the openness. If you self-host, if you fine-tune, or if you need to inspect a model’s behavior for compliance reasons, Qwen4-Max is not your model, whatever the price says. Kimi K3.1 Pro, by contrast, published its weights — and its open-weight status is one reason its $2/$10 rate card punches above its weight in the comparison above.

For API-only users the closed weights matter less day to day, but they matter for trust: when a vendor’s own benchmark chart is the only evidence for a claim, and outsiders can’t download the model to check it, the vendor’s numbers carry less weight. That is exactly the dynamic playing out in the benchmark fight above.

Who should switch — and who shouldn’t

Switch or try it now if: you run high-volume, input-heavy work — document summarization, long-context retrieval, bulk classification — where Qwen4-Max’s $1.50 input and $0.075 cached input dominate the bill. The one-million-token free quota makes a real evaluation free, and the 1M-token context window is genuinely useful for whole-codebase or whole-contract work. If your current bill is mostly Opus 5.5 input tokens, the $24-vs-$80 monthly math above is worth testing with your own prompts.

Try it with a measured test if: you run coding agents. Qwen4-Max’s DeepSWE number says it can hang with Opus 5.5; the independent Terminal-Bench number says it can’t. The only way to resolve that for your stack is to run your own tasks through both and compare completion rates, not benchmark scores. Budget for verbosity — measure tokens per completed task, not tokens per dollar.

Don’t switch if: you need the strongest reasoning or agentic coding available (Fable 5.2 and Opus 5.5 still lead the independent index by 16–22 points), you need open weights or fine-tuning (Qwen4-Max offers neither), or your workload is output-heavy with strict latency needs — early testers call the model slow, and at 53 tokens per second on the Artificial Analysis run, it is not a speed pick.

Watch, don’t commit, if: you’re deciding between Qwen4-Max and Tencent’s Hy4 flash, which launched the same day at $0.07 input and $0.281 output — 21 times cheaper on input — but with no model card, no architecture paper and no independent benchmark at all. Two cheap Chinese flagships landing on the same day is a story in itself; neither has been independently verified enough to bet a production pipeline on yet.

What to do this week

  1. Grab the free quota. One million tokens on QwenCloud, 90 days, Singapore region. Run your three most common tasks through qwen4-max and your current model side by side, and record tokens per completed task — not just the per-token price.
  2. Migrate off the preview IDs before November 25. If any code still calls qwen4-max-preview, switch it to qwen4-max now; the preview retires at 24:00 UTC+8 on November 25, 2026.
  3. Check the rate card again in early December. The unconfirmed report of a December 1 price doubling is worth one minute of your time to rule out before you scale up.
  4. Don’t re-architect around it yet. Qwen4-Max is a strong value play for input-heavy work and a genuine improvement on its predecessor. It is not, on the independent evidence available today, a frontier replacement — and with no open weights, it doesn’t inherit the Qwen family’s old self-hosting advantage.

If you want the deeper launch detail, our preview piece from November 11 covers what the preview was and how the Token Plan worked — the production launch supersedes it, but the context is there: Qwen4-Max preview launch, November 11, 2026. And if you’re weighing the Claude side of the same decision, our breakdown of Claude Fable 5.2’s pricing is here: Claude Fable 5.2 price cut, 20 percent, for freelancers.

Frequently Asked Questions

How much does Qwen4-Max cost?

Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens in the Singapore region on QwenCloud, with cached input at $0.075 per million. Other regions (Beijing, Hong Kong, Frankfurt, Virginia, Tokyo) list $1.413 input and $4.24 output. That is 25% below Qwen3.8-Max, which listed at $2.00/$6.00.

When did Qwen4-Max launch?

Alibaba launched Qwen4-Max on November 16, 2026, five days after a Token Plan-only preview opened on November 11. The production model runs on ordinary pay-as-you-go Model Studio keys, and the preview identifiers retire at 24:00 on November 25, 2026 (UTC+8).

Is Qwen4-Max open source?

No. Unlike earlier Qwen models, Qwen4-Max’s weights are not public — it is available only through Alibaba’s hosted API and partner gateways. Fine-tuning and self-deployment are not supported. Alibaba says smaller Qwen4 variants will be open-sourced later, with no date given.

How does Qwen4-Max compare to Claude Opus 5.5?

On price, Qwen4-Max is much cheaper: $1.50/$4.50 per million tokens against Opus 5.5’s $4/$20. On benchmarks, the two disagree: Alibaba’s own chart has Qwen4-Max level with Opus 5.5 on DeepSWE v1.1 (75.7 vs 74.2), but the independent Artificial Analysis Intelligence Index scores Qwen4-Max 43 against Opus 5.5’s 59. For input-heavy work the savings are real; for hard reasoning and agentic coding, Opus 5.5 still leads on independent evidence.

How many tokens of context does Qwen4-Max support?

One million tokens. Maximum input is 991,808 tokens without thinking and 983,616 with thinking on; maximum output is 131,072 tokens without thinking, or a 262,144-token budget shared between reasoning and the answer when thinking is on. It accepts text, images and video as input.

Should freelancers switch to Qwen4-Max?

For input-heavy, long-context work — summarization, retrieval, bulk document processing — yes, test it: the free one-million-token quota makes a real evaluation cost nothing, and the monthly math can cut an Opus 5.5 bill by two-thirds. For frontier reasoning, hard agentic coding or anything needing open weights or fine-tuning, no — stick with the leaders and re-check Qwen4-Max after independent benchmarks mature.

Leave a Reply

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