Qwen4-Max-Preview: Token Plan Access, Costs and Coding

Use model ID qwen4-max-preview through a Token Plan Individual or Team account. The documented alias is qwen4.0-max-preview; the billing view uses that alias. The preview runs from November 11, 2026 at 02:00Z until November 25 at 16:00Z, when the documentation says the model will be retired.

The official route uses a dedicated Token Plan key and Base URL, with OpenAI Chat Completions, Responses and Anthropic Messages compatibility. Token Plan is available only in the Singapore region. A regular pay-as-you-go endpoint is not the documented access route for this preview. Configure the client for the protocol you intend to use rather than changing only the model name.

The advertised capabilities include web search, function calling and structured output. Built-in model tools require the Responses API; a Chat Completions client does not automatically invoke those tools. Those help a client assemble an agent workflow; actual file access, browser actions and shell commands depend on the tools and permissions of that client.

The Individual plan lists Lite at $6 per month for 20,000 Credits, Essential at $10 for 45,000, Standard at $18 for 80,000 and Pro at $43 for 200,000. For accounts that have never purchased Token Plan Individual, limited new-user prices list Lite at $3 and Essential at $5, with eligibility ending November 30, 2026 at 15:59:59Z. These are subscription prices shared across the plan’s eligible models.

Qwen4-Max-Preview’s limited promotion consumes as little as 20% of the standard Credits rate. The additional night promotion brings that to 8% from 14:00Z through 00:00Z the following day. Because consumption depends on the model, tokens, thinking and tool calls, neither figure is a public dollar-per-million-token quote.

Each $15 monthly Credit Pack supplies 20,000 additional Credits. An active subscription is required; the account can hold up to five packs at once. Usage consumes the monthly quota first, then available pack Credits. Service pauses when both are exhausted, and unused monthly quota does not carry over.

For a practical comparison, record the Credits before and after several equivalent accepted tasks. Divide the measured Credits per accepted task into your remaining quota. For example, if your measured task uses 500 Credits, a 20,000-Credit quota covers 40 such tasks before other usage.

Obtain the dedicated key from your Token Plan account and use the protocol path in its guide. This Chat Completions example uses a small acceptance-oriented coding task. It does not install tools or run generated code. Setting an output budget limits the combined thinking and final answer.

The context limit is 1M tokens and the maximum output is 393,216 tokens, shared by thinking and the final answer. Thinking is enabled by default. The guide lists low, medium and xhigh effort, with xhigh as the default; do not send reasoning_effort and thinking_budget together.

For this model, max_tokens has the same meaning as max_completion_tokens. An old client budget sized only for the visible answer can therefore be used up by thinking. A truncated response calls for inspecting the combined budget, rather than assuming the model cannot complete the task.

Chat Completions preserves thinking by default with preserve_thinking. Return historical reasoning in the assistant message’s reasoning_content field, not concatenated into ordinary content. When changing clients, verify that its message-history handling follows this contract.

Large context is useful for related evidence, but repeatedly sending a whole repository can consume Credits without improving the patch. Start with the relevant files, a defined change and acceptance checks, then expand context when the task requires it.

Choose a task with a finish condition: implement one feature, pass its existing checks and stop with a reviewable diff. Record the model ID, effort, client, initial repository state, elapsed time and Credits consumed. A screenshot of a finished page answers a different question from a working change with passing behavior checks.

In a November 12 Reddit post, Dear_Lock_5280 shared a Qwen4-Max-Preview interactive sketchbook demo, a comparison demo attributed to GPT Sol 6.1 Max and the original prompt. The author preferred different aspects of each result. The post does not establish matched effort, task cost or repeated-run reliability, so it is useful as a design task to inspect rather than a general model ranking.

Include failure cases in your trial: truncated output, unsupported tool calls, quota exhaustion and changes outside the requested files. Set the client’s execution permissions and ask for a stop after the agreed artifact is complete. Review the diff and actual behavior before accepting the result.

Qwen4-Max-Preview fits an evaluation inside its published preview window when Token Plan is already a useful purchase. For a service that must continue after the window, keep the model configurable and test the replacement on the same acceptance set. Do not assume a later Qwen4 release will inherit this endpoint or quota rules.

Qwen3.8-27B provides downloadable weights and a different local-deployment decision. Gemini 4 Carbon is a hosted alternative with token-based task budgeting. Compare accepted work, access lifetime and total spend rather than treating plan Credits as interchangeable with API tokens.