A generated title card reading Qwen4-Max at 70% of List, with the subtitle A 30% discount that is exactly 30%, and three price cards reading Input $1.05 per 1M, Output $3.15 per 1M and Cache read $0.0525 per 1M.
Guides & Insights

Qwen4-Max Is 30% Off Alibaba's List Price: A Rate Card Before the Endpoint

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Qwen4-Max shipped on 16 November 2026. Three days later, on 19 November, a third-party inference platform put the model on its price list at $1.05 per million input tokens, $3.15 per million output and $0.0525 for a cache read. Divide those by A​libaba's own Singapore card — $1.50, $4.50 and $0.075 — and every one of them comes back at exactly 0.70. Not approximately 30% off. Exactly 30% off, on all three lines, to the last decimal place. That is not how an independent deployment prices itself; it is how a reseller prices itself. And the same page that carries the rate card also carries a "Launching soon" prompt and a sign-up box asking to be notified when Qwen4-Max is available, which makes this a price list published ahead of the thing it prices.

What the 30% actually is

Three numbers, one multiplier, and the multiplier is the whole story.

• Input — $1.05 per million tokens on the third-party card, against $1.50 on A​libaba's Singapore card. Ratio 0.700.

• Output — $3.15 against $4.50. Ratio 0.700.

• Cache read — $0.0525 against $0.075. Ratio 0.700.

A platform that had bought its own 2.4-trillion-parameter hardware, written its own sparse-attention kernels and negotiated bandwidth would land on numbers that cluster near A​libaba's, not on three figures that each divide cleanly by ten-sevenths. Independent pricing has independent noise in it: a rounding here, a margin there, a cache rate that reflects a different storage tier. Three-for-three exactness is the signature of a margin applied to a single upstream card, and the upstream card it was applied to is the Singapore one rather than the Global one — $1.413 and $4.24 would have produced $0.989 and $2.968, and the published figures are neither.

Which tells you the discount is promotional. Somebody is paying the 30% out of a launch budget, not out of a different cost structure, and the thing to ask about any promotional rate is how long it survives.

Three ways a platform gets 30% under list

There is no fourth explanation worth considering, and the three are worth being able to tell apart because they have different expiry dates.

The first is a launch discount funded by the vendor. A​libaba wants Qwen4-Max in front of developers who already have an API key somewhere, and the fastest way to get there is to subsidise a platform's first month. This is the most likely of the three and the one with the shortest clock, because it is the platform's balance sheet absorbing the difference for a defined window.

The second is genuine wholesale pricing — the platform buys at a negotiated rate below list and passes part of it through. This is the only version of the 30% that is stable, and it is also the version that would show up in a reseller agreement rather than in a marketing page. If the rate holds past January, this is what it was.

The third is a loss leader on a model that is expensive to serve and cheap to advertise. Qwen4-Max activates 71 billion parameters per token out of a 2.4-trillion-parameter backbone — a serving cost that does not shrink because the price did. A platform selling at 70% of list to win routing share is a plausible business, but it is a business with a floor beneath it.

All three produce the same price today, and none of them are visible from the rate card. That is the honest limit of what can be read off a page: the number is real and the arithmetic behind it is not published.

The limits the price does not carry

Here is the part that matters more than the discount, and it is the part a price-led comparison always drops. A​libaba's card for Qwen4-Max is not just money — it is a set of ceilings, and a reseller copies a price far more reliably than it copies a spec sheet.

• Context — A​libaba documents 1,000,000 tokens, with a 991,808-token input ceiling in non-thinking mode dropping to 983,616 in thinking mode. The third-party card advertises a 1M-token context without the input ceiling, which is the number that actually binds a long document.

• Output — A​libaba caps non-thinking output at 131,072 tokens and puts thinking plus answer inside a shared 262,144-token budget. A page that advertises "1M context" and stops there has told you nothing about how much answer you can get back.

• Modality — the vendor card is text, image and video in, text out, with native video understanding rather than an extractor bolted on. Third-party listings frequently ship the text and image paths first and the video path later, and the listing does not always say which one you are getting.

None of this is an accusation. It is the ordinary gap between a model and a deployment of a model, and on a four-day-old release there is exactly one source for the ceiling — A​libaba's own Model Studio page — because nobody has run the second deployment long enough to publish against it.

Practical consequence: if you are sizing a long-context job, take the context figure from the vendor and take the price from whoever is cheapest, but do not assume they describe the same product until you have sent one 400,000-token request through and watched where it truncates.

A generated single-column price card for Qwen4-Max with six rows: input $1.05 per 1M against list $1.50, output $3.15 against list $4.50, cache read $0.0525 against list $0.075, a discount of exactly 30% on all three lines, a 1M context with a 991,808-token input ceiling, and a status line reading price published, endpoint not yet live.

What a 30% discount is worth on an agent run

Token prices are the wrong unit for a model sold on long-running agentic work, so here is the arithmetic on a shape Qwen4-Max is explicitly positioned for — a coding agent holding a 200,000-token repository context across 40 turns, emitting 30,000 tokens of reasoning and patch per turn. List prices, no caching, Singapore rates on both sides.

• At A​libaba's card — 200,000 input tokens per turn at $1.50 per million is $0.30, and 30,000 output tokens at $4.50 per million is $0.135. That is $0.435 a turn, or $17.40 across 40 turns.

• At the third-party card — the same turn is $0.21 in and $0.0945 out, $0.3045 a turn, or $12.18 across 40 turns.

• The saving — $5.22 per 40-turn run, which is $0.13 per turn. On a workload of a thousand runs a month that is $5,220; on a workload of ten it is lunch.

Now add caching, because that is where the discount compounds and where Qwen4-Max's own economics are most interesting. A​libaba documents cache reads at $0.075 per million in Singapore and cache writes at $1.875 for a five-minute window. At 200,000 cached tokens per turn the read costs $0.015, so the input side of a warm turn collapses from $0.30 to $0.015 — a twentyfold cut before any discount at all. The 30% is real money, but it is 30% of a number that competent caching has already made small. On a run that re-reads a stable prefix, the caching structure is worth roughly twenty times what the platform discount is worth, and no rate card advertises that.

The place it stops working is a prefix that changes every turn. A tool-calling agent whose context is mostly fresh tool output gets no cache hits, pays the write premium on nothing, and falls back to the full input rate — at which point the 30% is the entire saving and it is the only lever left.

A screenshot of the Alibaba Cloud Model Studio documentation for qwen4-max, captured in an English interface, showing the qwen4-max latest-model heading, the Model List and Model Billing sidebar, and the Text, Vision, Speech and Embedding model categories.

The endpoint question

All of the above assumes you can call the model at that price, and as of 19 November that is the one thing the page does not settle.

The same model card carries three claims that do not sit comfortably together. It lists Qwen4-Max as serverless and available, it publishes a full metered rate card, and it shows a "Launching soon" state with a sign-up prompt offering to notify you when the model is available. Any two of those are consistent. All three at once is a page in transition, and the correct reading is that the pricing is published and the capacity is not yet — which is a normal state for a model four days old and an unusual thing for a price list to say out loud.

That matters for a decision in a specific way. A promotional rate that starts before an endpoint exists has a clock that starts later than the announcement, or it does not start at all. If you are building a cost model around $1.05, model it as a discount you have not yet been able to test, and keep A​libaba's $1.50 as the number your budget survives on.

OrcaRouter does not serve Qwen4-Max — the model is reachable through A​libaba's own Model Studio API and several third-party platforms, and the discount above belongs to one of those. The model we do route from the same family is Qwen3.8-Max, at its published list price with 0% markup passed through, which is the same pass-through structure that makes a vendor price change land here the same day it lands upstream. One API for 200-plus models, one key, and a routing DSL that can send a request to Qwen3.8-Max or to a model from another lab without a second contract — the useful shape when the model you actually want is on a price list nobody has finished standing up.

A screenshot of the OrcaRouter model page for qwen/qwen3.8-max, captured in an English interface, showing the model name, the qwen/qwen3.8-max identifier, text, image and video input with text output, and a $2.00 price figure in the page header.

What to watch, and by when

• Whether the 30% survives its first month. A promotional rate and a wholesale rate look identical in week one and nothing alike in February. The date to check is the end of December 2026.

• Whether the "Launching soon" state clears, and what the limits say when it does. A published 1M context with no stated input ceiling is not the same product as A​libaba's 991,808-token non-thinking ceiling, and the difference only shows up on the request that truncates.

• Whether a second platform publishes a card that is not 0.70 times A​libaba's. Two platforms landing on the same multiplier would mean a wholesale rate; two landing on different ones would mean two launch budgets, and the cheapest of them is the one with the earliest expiry.

• Whether A​libaba responds. A vendor that watches 30% come off its flagship within three days either pulls the wholesale rate or moves its own list price, and the second of those is the one a buyer wants.

The headline is that A​libaba's flagship is available at 70% of list, four days after it went generally available. The more useful sentence is that the 30% is the smallest number in the story — smaller than the caching structure it sits on top of, smaller than the context ceiling it does not mention, and smaller than the gap between a price list and an endpoint you can call. Price the discount. Budget the list.