Qwen3.8-Max: a million tokens of context at $2 per input

On 3 August Alibaba opened access to Qwen3.8-Max. The 2.4-trillion-parameter model holds up to 1 million tokens at once and understands text, images and video. Pricing is $2 per million input tokens and $6 per million output tokens. Open weights are promised for the following week.

The number worth reading on for is not the parameter count. It is the million tokens of context.

Why do long documents no longer need to be cut into pieces

A four-hundred-page contract used to be too large to fit into a model whole. It was sliced up, each piece was processed separately, and the answers were then glued back together. Cross-references were lost in the gluing: clause 4.2 points to an annex, and the annex ended up in a different request.

A million tokens is roughly thousands of pages at once. Instead of slicing — a single request. More on how the context window is counted and why a token is not the same as a word.

The second difference is the promised open weights. They are due to be published in the week after release, so somewhere around 10 August. What open weights are and why they matter for a business: the model can be downloaded and run on a company's own hardware, without calling anyone else's API.

What should be counted — the price per token or the volume of operations

A price in dollars per million tokens does not answer the question "what does it cost". A calculation on a specific workflow does.

Take the processing of incoming documents. The method: an average contract runs 20 pages, a page holds about 500 words, and a word in Russian text comes to roughly 1.5 tokens. That gives 15,000 input tokens per document. The model's answer is a structured summary of 1,500 output tokens. The exchange rate is RUB 80 to the dollar.

One document: 15,000 × $2/1,000,000 = $0.03 on input, 1,500 × $6/1,000,000 = $0.009 on output. Total $0.039, or about RUB 3.1.

A thousand documents a month comes to roughly RUB 3,100 at these rates. That is payment for model calls, and for nothing else. In an implementation budget it is the third line item, after development and support — what adds up to the total is broken down in the piece on the cost of implementing AI.

The benchmark rests on three assumptions: document length, answer length, exchange rate. Change any of them and the figure has to be recalculated. To sanity-check the order of magnitude on a real workflow, use the AI budget calculator.

The number looks small, and that is true only of the pricing line. The hours spent preparing data, checking output and reworking prompts are not in it.

Who does the arithmetic change for

Companies whose documents run past the old limit. Project documentation, multi-volume contracts, technical regulations, correspondence archives. These used to be cut into parts, with the loss of links between sections treated as the cost of doing business. What to do this week: take the three longest documents from a real workflow, count the tokens in them, compare against a million. If they fit, the slicing is no longer needed.

Companies that hit a price ceiling at high volume. At tens of thousands of operations a month, a difference in the rate stops being a rounding error. And the comparison is worth making against more than Western models: DeepSeek V4 Flash came out earlier at a different price point. What to do: export a month of statistics on text volume and recalculate it at both rates.

Companies that need their own perimeter. Healthcare, finance, work with personal data. Open weights mean that, once published, the model can be deployed on-premise so documents never leave the building. What to do: wait for the weights to be published and ask IT to size the hardware — a 2.4-trillion-parameter model does not waive its equipment requirements.

When is the older version still enough

A flagship is needed where the task is genuinely hard. Classifying enquiries into ten categories, extracting fields from a standard invoice, short answers from a knowledge base — on work like this, the gap between models is swallowed by the gap in data quality.

A million tokens of context is not a replacement for RAG. Pushing an entire document corpus into every request means paying full price on input with every call. A search over a vector database hands the model the three paragraphs it needs instead of a thousand pages, and at volume that is cheaper.

A separate word on security: a powerful model with open weights makes more than the useful scenarios easier. The same tool works for convincing phishing and for sifting through stolen data — a point made by CISOClub. Which means the bar for screening incoming email rises whether or not the model is adopted inside the company.

What is known about Russia

The model is stated to be available for testing to users worldwide. How the API can be paid for from Russia is not described in the sources — nothing on cards, on reseller partners, or on corporate entities. For now this is an open question, not a settled one.

Where data sits when the model is used through Alibaba's cloud is likewise not disclosed in the coverage. For personal data this is the decisive point: the processing requirements were covered in the piece on the FSTEC order.

Open weights close this question a different way — through local deployment. But the weights have not been published yet; they are promised within the week.

What is not yet known: real measurements of response speed, the licence on the weights, payment terms from Russia, and the data retention policy in Alibaba's cloud. Claims of comparison with Western flagships come from the vendor and have not been independently confirmed.