A generated title card reading Qwen4-Max leaves preview, with the subtitle The metered price card, and the output budget that got smaller, and four chips reading $1.50 in / $4.50 out per 1M, 1,000,000-token context, 262,144 shared budget and Six regions.
Guides & Insights

Qwen4-Max Leaves Preview: The Metered Price Card, and the Output Budget That Got Smaller

Author

jinhao.song

Date Published

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

Qwen4-Max went from a token-plan preview to a metered model on 16 November 2026, and the number to read first is not a benchmark — it is $1.50 in and $4.50 out per million tokens in Singapore, against $2.00 and $6.00 for Qwen3.8-Max. That is a straight 25% cut on both sides of the meter. But the same switch that published a real price card also shrank the shared thinking-plus-answer budget from 393,216 tokens on the preview to 262,144 at general availability, moved the model off a credit plan and onto per-token billing in six regions, and left batch inference available in exactly one of them. If you were running the preview and treating 16 November as a formality, the spec deltas are the part that will bite, not the discount.

What the 16 November switch actually changed

The preview was a good deal and a narrow one: Token Plan credits only, Singapore only, and a documented retirement at 24:00 on 25 November 2026 (UTC+8). General availability replaces all three of those constraints and adds two that nobody asked for. The model identifier is now plain qwen4-max; the preview strings qwen4-max-preview and qwen4.0-max-preview are the ones with an expiry date, so a config still pointing at either is on a clock that ends in nine days.

• Billing — per-token metering in six regions, versus Token Plan credits in one. The 1-million-token free quota on the Singapore card is valid for 90 days from activation, model release, or approval, whichever is later.

• Regions — Beijing (Chinese mainland), Singapore (International), and Frankfurt, US Virginia, Tokyo and Hong Kong all under the Global scope. The preview was Singapore alone.

• Shared output budget — 262,144 tokens combining thinking and the final answer at GA, against 393,216 on the preview card. A third of the headroom is gone, and it is gone from the part that long agent runs spend fastest.

• Non-thinking output — capped at 131,072 tokens, with a 991,808-token input ceiling. In thinking mode the input ceiling drops to 983,616, because the reasoning and the answer come out of the same 262,144-token bucket.

• Batch inference — supported in Beijing only, and not supported in the other five regions. The preview had no batch path at all, so this is new capability rather than a regression, but it is regional in a way the marketing does not lead with.

• Protocol surface — OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope, all four, with structured output, parallel tool calling, prefix completion, web search, and both explicit and implicit caching. Input accepts text, image and video; output is text only.

The price card, laid out the way the bill arrives

Alibaba publishes two different cards for the same model, and the difference is not small enough to ignore when you are choosing a deployment scope.

• Singapore, input and output — $1.50 / $4.50 per million tokens, flat across the full 1,000,000-token window with no tiering at 32K or 128K.

• Everywhere else, input and output — $1.413 / $4.24 per million tokens, flat across the same window. Same model, same limits, 5.8% cheaper than Singapore per token.

• Cache writes, Singapore — $1.875 per million for a 5-minute write, $3.00 for a 1-hour write. Elsewhere: $1.767 and $2.827.

• Cache reads, Singapore — $0.075 per million. Elsewhere: $0.071. That is 5% of the Singapore input rate, and it is the single largest lever on a long-context agent workload.

• Batch, Beijing only — Batch File at $0.707 / $2.12 per million, Batch Chat at $1.413 / $4.24. Only Batch File carries the 50% batch discount; Batch Chat is listed at the real-time rate, so "batch" in the endpoint name is not a price guarantee.

• Against Qwen3.8-Max, same page — $2.00 / $6.00 in Singapore, $1.65 / $4.951 in the Global regions. Qwen4-Max is 25% cheaper in Singapore and 14.4% cheaper elsewhere, because the older model's non-Singapore card is itself discounted.

The 14.4% figure is the one that matters if you already run in Frankfurt or Virginia: the generation-over-generation cut is smaller there than the Singapore headline suggests, and the vendor's 62% average-cost-per-task claim is doing a lot of work that the per-token card alone does not.

A generated two-column price scoreboard for Qwen4-Max and Qwen3.8-Max: Singapore input and output $1.50 and $4.50 against $2.00 and $6.00 per 1M, Global input $1.413 against $1.65, cache read $0.075 against a listed discount, both at 1,000,000 context, shared output budget 262,144 against not published.

A worked example, so the cache line stops being abstract

Take a 400,000-token document prefix reviewed in three passes, with 8,000 output tokens per pass, non-thinking, Singapore rates. This is arithmetic on the published list prices, not a measured run.

• Pass one — the prefix is written to cache at $1.875 per million, $0.75, plus 8,000 output tokens at $4.50 per million, $0.036. Total $0.786.

• Passes two and three — the prefix is read at $0.075 per million, $0.03 per pass, plus the same $0.036 of output. $0.066 each.

• Three passes with caching — $0.918. The same three passes paying full input each time — $1.908. The prefix is read twice for less than the price of one fresh pass.

• A fourth pass — caching wins by a wider margin every time you re-read, which is exactly the shape of a long agent run that keeps a fixed system prompt, tool schema and document set in front of a moving question.

The break-even is close to immediate: the cache-write premium over the normal input rate is $0.375 per million in Singapore, so a cached prefix pays for itself on the first re-read. The place this stops working is a prefix that changes every turn — you pay the write premium and never collect, and the 5-minute default write window is short enough that a slow tool call can expire it.

Three defaults that decide the invoice before your prompt does

Qwen4-Max ships with reasoning switched on and set high. That is a reasonable default for a flagship and an expensive one for a classifier.

• Effort defaults to high — reasoning_effort accepts low, high and xhigh, and omitting it gives you high. thinking_budget likewise defaults to 131,072, and max_completion_tokens to 262,144 in thinking mode. A request that sets nothing has reserved more output budget than most jobs will use.

• There is no medium — Chat Completions and DashScope map minimal to low, medium to high, and max to xhigh. On the Anthropic Messages path the parameter is output_config.effort and maps the same way. Asking for medium costs what high costs.

• preserve_thinking defaults to true — prior reasoning is carried into context on multi-turn calls. That inflates the input side of every follow-up and simultaneously raises the chance the prefix still matches the cache. On a long agent loop the two effects partly cancel; on a short chat they do not, and you are paying to re-read thoughts you will never use.

• Sampling defaults are thinking 1.0 / 0.95 / 20 and non-thinking 0.7 / 0.8 / 20 for temperature, top_p and top_k, with presence_penalty at 1.5 outside thinking mode. Setting temperature explicitly without touching effort changes the shape of the output but not the bill.

What the vendor says improved, and what nobody has checked

Every number in this section is Alibaba's own and has not been reproduced outside the lab. As of 16 November there is no independent index entry for Qwen4-Max, so treat the comparison as a claim rather than a measurement.

• Work360 — 52.69 against 47.74 for Qwen3.8-Max, on a taxonomy the vendor built: 26 Tier-1 domains, 182 Tier-2, 721 Tier-3.

• Token efficiency — roughly 24% fewer tokens at xhigh effort and 35% at high on Work360; roughly 48% and 49% on Terminal-Bench 4.0. This is the claim that compounds with the price cut, and it is the one to test on your own traces first, because token savings are workload-shaped.

• Average cost per task — $4.43 to $1.67 on Work360, a 62% reduction. That figure multiplies a price cut by a token-saving claim; neither factor is independently verified.

• Base-model quality — 5% to 13% lower bits per byte than Qwen3.8-Max-Base while activating 25% fewer parameters, measured in-house across knowledge, code, QA and STEM. The vendor notes that two of the compared checkpoints are post-trained, which weakens the base-model framing.

• Secure coding — Last Secure Code Bench function-level pass rate up from 22.0% to 31.8%, and code that is secure at function, file and repository granularity simultaneously up from 8% to 14.7%.

The architecture behind the efficiency claim is a 2.4-trillion-parameter sparse mixture-of-experts backbone with 71 billion parameters activated per token, plus 328 billion more in N-gram embedding parameters that are looked up rather than multiplied. Three Gated DeltaNet layers compress history into fixed-size state for each Qwen Sparse Attention layer, which selects context at micro-block granularity through a lightweight indexer, and four parallel residual branches with dynamic read and write gates stabilise the deep stack. It is a plausible account of why long-context cost falls. It is still an account.

A screenshot of the Alibaba Cloud Model Studio page for qwen4-max, captured 16 November 2026 with an English interface, showing the qwen4-max heading, a 11-16 changelog entry, the model description, and the documentation sidebar including Model inference pricing and Batch inference.

Trying it without moving a production path onto it

The awkward part of this launch is that the model is genuinely new, the benchmarks are unaudited, and the price card is attractive enough that someone will want it in production this week. Those three facts point at a staging strategy rather than a migration.

Qwen4-Max is reachable through Alibaba's own Model Studio API and several third-party platforms; it is not in the OrcaRouter catalogue as of 16 November. Qwen3.8-Max is, at $2.00 in and $6.00 out per million tokens in Singapore — the vendor's list price passed through with 0% markup, so the 25% gap to Qwen4-Max is a real gap and not a routing margin. That makes the honest test cheap: keep the incumbent on the key you already have, send a shadow slice of traffic at the new model through the vendor, and compare cost per completed task rather than cost per token. The vendor's own framing — $4.43 to $1.67 — is a cost-per-task number, and cost per task is the only comparison where a token-efficiency claim and a price cut land on the same axis.

For teams already routing through OrcaRouter, the switching cost that normally blocks this kind of evaluation is the second contract and the second SDK. One API for 200-plus models, automatic failover across providers, and a routing DSL that composes several models into a single call means the mixed state — incumbent handling the volume, candidate handling a slice — is a config change rather than an integration. For an unproven model that is the difference between a two-week proof and a two-day one, and failover is what stops the proof from becoming an outage when the candidate has a bad afternoon.

A screenshot of the OrcaRouter model page for qwen/qwen3.8-max, captured 16 November 2026 with an English interface, showing the model name, the qwen/qwen3.8-max identifier, text, image and video input capabilities, a 1,000,000-token context, and a pricing panel.

What to do before 25 November

If you are on the preview, the deadline is not a soft one: qwen4-max-preview is documented for retirement at 24:00 on 25 November 2026 UTC+8, and the GA identifier is qwen4-max with different limits. Migrate the string, then re-check two things that will not throw an error when they are wrong — any code that assumed a 393,216-token shared budget now has 262,144, and any job that relied on Token Plan credits is now metered per token.

If you are choosing whether to adopt, the case for waiting is thinner than it was a week ago, because the metered card and the cache pricing are now published and the price is genuinely below the previous generation in every region. The case for care is the shared output budget, the Beijing-only batch path, and a benchmark suite that the vendor wrote. Run the shadow slice, price it per completed task, and let the next two weeks of your own traces decide — that evidence will be worth more than a 52.69 on a taxonomy none of us can audit.

One more thing is coming that changes the maths again: Alibaba says the smaller Qwen4 models will be open-sourced on the same lineage, which historically means the price floor under this tier moves within a quarter. Anything you commit to at today's rates should be a commitment you can walk away from.