A generated title card reading Qwen4-Max - the partner card, read line by line, with three figures set as a rate strip: $1.413 input, $4.24 output and $0.071 cache read per million tokens, labelled Global-scope rates, alongside a chip reading 262,144 shared budget.
Guides & Insights

Qwen4-Max's New Partner Card Advertises Four Capabilities. Three of Them Cost You Something Later.

Author

jinhao.song

Date Published

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

A second platform is now selling Qwen4-Max, and its announcement reads like a product launch: a 1M-token context, text, image and video input, tool calling, structured output, reasoning, and $1.413 per million input tokens, $4.24 per million output, $0.071 per million cached. The model went generally available on A​libaba's own API on 16 November 2026, so a second seller two days later is fast but not surprising. What is worth reading closely is the feature list, because three of those four capabilities behave differently on a partner platform than they do on A​libaba's own, and the price is the clue that tells you why.

Start with the price, because everything else follows from it. $1.413 and $4.24 are not new numbers. They are A​libaba's Global-scope rates — Beijing, Hong Kong, Frankfurt, US Virginia and Tokyo — to three decimal places, sitting 5.8% below the Singapore card of $1.50 and $4.50. The cached rate is the tell: $0.071 is the Global figure, and Singapore's is $0.075. Nobody arrives at $0.071 independently. A partner card quoting the vendor's regional rate to the thousandth of a dollar is quoting a pass-through, and a pass-through means one deployment of the model with a second billing layer on top, not a second training run or a second set of kernels.

That distinction is not pedantry. It sets the expectation for every other row on the card: prices survive a pass-through, and behaviour does not, unless the partner has done the engineering to make it.

A generated two-column comparison card titled Qwen4-Max - the advertised card, checked, with the partner-listing column reading 1M tokens context, output ceiling not stated, $1.413 input, $4.24 output, video input yes and reasoning dial not stated, against the vendor column reading 1M with 984k usable, 262,144 shared, $1.413 Global, $4.24 Global, video yes, and low, high and xhigh on the dial.

Claim one: 1M context. True, and still the least useful number on the page

Every Qwen4-Max listing advertises a million tokens, and every one of them is technically right. A​libaba's own documentation gives 1,000,000 tokens of context with a 991,808-token input ceiling when thinking is off and 983,616 when it is on, and Artificial Analysis independently records a 984k-token context window. Those three numbers are the same 1M with different amounts of the truth exposed.

The row that matters is not the input ceiling but the output budget, which no partner card in this category advertises. Qwen4-Max has a 262,144-token budget shared between the thinking and the final answer, of which 131,072 is available when thinking is off. On a 900,000-token input, that leaves you at most 91,808 tokens of response in thinking mode before the window closes — and reasoning tokens count against it. A team sizing a long-context job against "1M in, some out" is sizing against a number the vendor never published.

This is the first place a partner listing can quietly cost you money. A card that says 1M says nothing about whether the platform enforces the shared budget, whether it truncates reasoning rather than the answer, or whether it silently stops at 128K of output. Those are the failure modes that show up as a truncated JSON object at 2am, not as an error code.

Claim two: video input. This is the row that decides whether the listing is a real port

Text, image and video input is the capability in this announcement that is hardest to fake. Serving video to Qwen4-Max is not a matter of accepting a URL and passing it along: frames have to be sampled, encoded into the model's visual token stream at a resolution and frame rate the model was trained on, and billed against a context that is already the scarce resource. A platform that has only wired the text path will either reject video outright or accept it and quietly degrade it.

Which is why the video row is the one to test first if you are evaluating a partner. A one-minute 30fps clip at a conservative sampling rate is thousands of visual tokens; drop it into a 700,000-token document prompt and you are at the input ceiling with a shorter answer than you planned for. Qwen4-Max's video support is real on A​libaba's own API — the model is documented for it, and Artificial Analysis records text, image and video as its modalities — but a partner card is an assertion, not a test result.

The image path has a milder version of the same problem. Resolution handling, tiling and the token cost of a high-DPI screenshot are all platform decisions, and none of them appear on a rate card that prices input per million tokens.

Claim three: tool calling and structured output. The one place a partner can be better than the vendor

Structured output and tool calling are where a pass-through platform can add real value, and it is worth saying so plainly because the rest of this piece is cautious. A​libaba documents structured output, parallel tool calling, prefix completion, web search and both explicit and implicit caching across four protocol surfaces — O​penAI Chat Completions, O​penAI Responses, A​nthropic Messages and DashScope. A partner that exposes only the O​penAI shape is offering less surface, not a worse model.

But a partner that has built a schema-constrained decoder on top can deliver something the vendor's own endpoint does not guarantee: JSON that parses on the first attempt, every time, without a retry loop. For a pipeline that extracts fields from invoices at volume, one avoided retry is 2× the input tokens saved, and that dwarfs the 5.8% regional difference between the Singapore and Global price cards.

So the question to ask about this capability is not "is structured output supported" — every listing says yes — but "is it constrained decoding or a prompt instruction". The first is a guarantee; the second is a strong suggestion that fails one time in fifty, usually on the longest and most expensive request in the batch.

Claim four: reasoning. The capability that ports worst, and the one nobody prices

Qwen4-Max ships with reasoning on and effort defaulting to high. The dial accepts low, high and xhigh, the thinking budget defaults to 131,072 tokens, and max_completion_tokens defaults to 262,144 in thinking mode. A request that does not set any of those three gets the expensive configuration by default, and that default is where most of the bill comes from: Artificial Analysis measured 104,940 output tokens per index task, of which 68,914 were reasoning, against a field median of 65.1 million output tokens for the whole nine-evaluation suite. Qwen4-Max generated 127 million.

On a partner platform, three things can happen to that dial and only one of them is visible. The reasoning effort parameter can be passed through unchanged, in which case everything above holds. It can be accepted and ignored, in which case your low effort request bills like a high one. Or it can be rejected, in which case you find out with a 400 and adjust. The first is fine, the third is annoying, and the second is the one that quietly doubles a monthly invoice — and it is the most common, because passing an unknown field through to a vendor API is often the default behaviour of an O​penAI-compatible proxy.

This is also where the price claim gets tested hardest. At $1.413 / $4.24 a partner is billing the vendor's Global rate, so the token price is honest. Whether the token count is honest depends on a parameter that does not appear on the card.

What a "collaboration" announcement tells you, and what it does not

The framing of this launch — a partnership with A​libaba's Qwen team rather than a resale — is worth taking at face value for what it covers and not a word more. It plausibly covers early access to weights, capacity allocation and a supported identifier. It does not cover the things that decide your bill: the thinking budget, the input ceiling in each mode, the video sampling policy, the constrained-decoding guarantee, or which regions the deployment sits in.

There is a testable version of this. Ask the platform for its input ceiling in thinking mode. A​libaba publishes 983,616. If the answer is "1M", the platform has not read the vendor's documentation, and the rest of its spec sheet deserves the same scepticism. That single question separates a partner that has ported the model from one that has ported the price.

The independent data is now better than it was at launch, which helps. Artificial Analysis scores Qwen4-Max at 43 on the Intelligence Index v5.0.1 — the median for comparable models — at $3.27 per completed index task, 32.1 minutes per task, 54.5 output tokens per second and 155.4 seconds to the first answer token. That is a real independent evaluation of a two-day-old model, and it is the number to quote against any vendor chart. What it does not tell you is whether a partner platform reproduces it: the tracker measures the model on one endpoint, and a second deployment with different batching, different kernels or a different reasoning default can land several points and a large multiple of cost away.

A screenshot of the Artificial Analysis page for Qwen4 Max captured 18 November 2026 with an English interface, showing the Alibaba provider tag and a November 2026 release, an Intelligence Index of 43, output speed of 55 tokens per second, 185.4 seconds to first token, a 984k-token context window, and a summary line describing the model as below average in intelligence, well priced, notably slow per task and very verbose.

How to buy this without guessing

OrcaRouter does not host Qwen4-Max. The model is reachable through A​libaba's own Model Studio API and several third-party platforms, and that is the honest description of the situation — no vendor-hosted route on this platform, so no claim to one. What OrcaRouter does provide is the comparison point: 200-plus model identifiers behind one API at 0% markup, provider list price passed through, so the $1.50 Singapore card and the $1.413 Global card are the same numbers here as on A​libaba's own pricing page, and a vendor price change is live the same day rather than at the next billing cycle.

A screenshot of OrcaRouter's models catalogue captured 18 November 2026 with an English interface, showing the header '226 models - 17 providers - one API key, one bill', the filter rail for input modalities, context length, input price, status, access tier, series, supported parameters, privacy and inference region, and the first card visible being kimi/kimi-k3.1-pro next to a two-step 'how to call any model' panel.

For a team deciding whether a partner listing is worth adopting, the useful setup is a side-by-side rather than a review. Send the same 50 requests — one long document with a video, one structured-extraction batch, one agentic tool-calling chain, one low-effort classification job — down the vendor's own endpoint and down the partner's, and compare tokens billed, not tokens returned. The two numbers diverge exactly when a parameter is being dropped, and the routing DSL is what lets you keep the partner for the workloads where it is genuinely better and fall back automatically for the ones where it is not.

The announcement is accurate. The price on it is A​libaba's own Global rate, the modality list matches the vendor's documentation, and a partner that quotes a cache read to three decimal places has clearly done the arithmetic. What it leaves out is the part you cannot verify from a card: a shared 262,144-token budget that shrinks the usable window, a reasoning dial that decides most of the invoice, a video path that is either real engineering or an accepted field, and a structured-output promise that is either constrained decoding or a wish. Those four things are worth more of your evaluation time than the 5.8% the partner price saves.