Alibaba’s Qwen team released Qwen4-Max to the public API today, November 16, 2026, 5 days after a preview that ran on its Token Plan alone. The QwenCloud pricing page lists the model at $1.50 per million input tokens and $4.50 per million output tokens in its International scope. Converted at the Bank of Canada rate of November 13, 2026, those figures read $2.14 CAD and $6.42 CAD. The model it replaces on the same page, Qwen3.8-Max, lists at $2.85 CAD per million input tokens and $8.55 CAD per million output tokens. A new flagship that undercuts its own predecessor by a quarter is a routing story before it is a quality story.
The direct answer: Qwen4-Max costs $2.14 CAD per million input tokens and $6.42 CAD per million output tokens on QwenCloud, with cached input at $0.11 CAD per million tokens and a free quota of 1 million tokens. A Canadian firm should test it on low-risk, high-volume steps, summaries, classification, first-pass extraction, and keep client data and regulated records on the models and regions it has already cleared. The price earns a test. The price alone does not earn a migration.
- What the price table actually says
- Which steps to route, and which to keep home
- The test to run before any client work moves
What the price table actually says
Read the table with its conditions attached, the way a buyer should. The $2.14 CAD input price holds for requests up to 1M tokens, in thinking and non-thinking modes alike. Cached input reads at $0.11 CAD per million tokens, which is where a steady workload with repeated context saves real money. Cache writes cost more, at $2.67 CAD per million tokens for a 5-minute cache and $4.28 CAD for a 1-hour cache, so a workflow that writes a cache and reads it once loses the discount it chased. The free quota is 1 million tokens, enough for a serious routing test and nothing more.
The comparison that matters sits one row up the same table. Qwen3.8-Max still lists at $2 and $6 in US dollars. Alibaba cut the flagship price while raising the model generation, which tells a buyer where the contest now runs: the large-model price floor drops every few months, and any routing plan built on today’s prices needs a review date. This site priced the same pressure from the other side last week. Gemini 4 Hydrogen costs $2.14 CAD. Speed is the real change put Google’s flat rate at the same input figure, and Gemini 4 Carbon is out. The token price is the small number argued that cost per finished task outranks cost per token. Both points hold here. Qwen4-Max wins the token table. The task table is the test a buyer still has to run.
Which steps to route, and which to keep home
Route by risk alone. The steps that suit a cheaper model share 3 traits: the input holds no client secrets, the output gets checked by a person or a second pass, and a wrong answer costs a redo rather than a client. Meeting summaries from your own team’s notes, ticket classification, first-pass extraction from public documents and draft internal summaries all fit. Claude for Manufacturing drafts the quote. Your estimator still sets the price drew the same line inside a plant office last week: the model drafts, a named person decides.
The steps that stay home are the ones carrying other people’s information. QwenCloud prices by deployment scope and region, and the International scope is the one quoted here. A Canadian firm handling client records, health information or legal files has to answer, in writing, which region serves its calls, where the data rests, and which privacy law the arrangement answers to, before the first client document moves. PIPEDA governs most private-sector handling in Canada, and several provinces add their own health and public-sector rules. A $6.42 CAD output price does not answer any of those questions. Gemini Decision 1 answers 1,000 questions in 1 pass showed the volume a routed step can reach. Volume is exactly why the data question comes first, because a routing rule set at scale is hard to unwind.
The test to run before any client work moves
The recap is short. The flagship price fell by a quarter in 1 generation, cached input fell further, and none of that settles where your data goes.
Run this test before any client work moves: take 20 finished tasks from your own workflow, run each on Qwen4-Max and on your current model, and score the outputs blind against the version a person approved. Count accepted outputs, total tokens and the region each call touched. Put 1 question to the vendor in writing alongside the test: which region serves a Canadian account on the International scope, and where do prompts rest? No hands-on test was run for this piece. Your 20-task count is the evidence that counts, and a routing rule written after it will survive the next price cut.