Latest in AI
Qwen4-Max Preview: Token Plan Pricing, Access and API Limits
The preview has an official model ID and published context limits, but its subscription rules make it a different proposition from a production API.
Alibaba’s Qwen4-Max-Preview opened on November 11 through QwenCloud Token Plan. Developers can select qwen4-max-preview, but the preview does not support pay-as-you-go access. Before changing a coding tool’s configuration, check the subscription’s permitted use, region and the model’s scheduled retirement.
The official model guide specifies a preview from November 11 at 10:00 to November 25 at 24:00, UTC+8. It says the model will be retired when that period ends. That makes this an evaluation opportunity with an expiry date, not a stable dependency to build a service around.
What Qwen4-Max-Preview supports
The model guide lists a one-million-token context window and a maximum output of 393,216 tokens shared between thinking and the final answer. Inputs include text, images and video. Supported interfaces include OpenAI Chat Completions, Responses and Anthropic Messages, with tool calling and structured output available.
Those are documented capabilities, not evidence that every client exposes them equally well. A coding tool that accepts a custom endpoint may still handle images, tool results or reasoning history differently. Start with a small task whose expected result you know before trusting a long agent run.
This guide is based on the published documentation, not firsthand performance testing. It does not rank Qwen4-Max against another frontier model.
Token Plan pricing is credits-based
The Personal Token Plan documentation lists these standard monthly prices as of November 11:
| Personal plan | Monthly list price | Monthly credits |
|---|---|---|
| Lite | $6 | 20,000 |
| Essential | $10 | 45,000 |
| Standard | $18 | 80,000 |
| Pro | $43 | 200,000 |
Credit use depends on the model, token usage, thinking and tools, so this table cannot be converted into a universal price per million tokens or a fixed number of coding tasks. Check the checkout terms and actual usage dashboard rather than comparing subscriptions by headline dollars alone.
Personal plans are limited to interactive programming and agent tools. Their terms prohibit automation scripts, custom application backends and non-interactive batch calls. Subscription credits do not turn the preview into a general-purpose backend API.
The Personal documentation says exhausted plan and credit-pack balances pause service. Do not assume automatic paid overage.
Fix the region and key before troubleshooting the client
The Token Plan overview says purchase and use are available in the Singapore region. The plan uses its own API keys, prefixed sk-sp-, and its dedicated endpoint. Ordinary pay-as-you-go keys are not interchangeable with those plan keys.
If the purchase entry is missing in another console region, the documentation directs users to Singapore. That is an account configuration requirement, not evidence that the model has disappeared. Check your organization’s approved processing arrangements before changing regions or sending data.
For setup, follow the integration guide for your specific tool, use the plan-specific credential and select qwen4-max-preview. The dedicated base host is https://token-plan.maas.qwencloudapi.com; the full endpoint depends on the protocol. Keep credentials out of source control and shared screenshots.
An authentication failure should first trigger a check of the key type and endpoint pairing. Repeatedly changing the model name will not repair a request sent with the wrong credential family.
Why an old output limit can truncate answers
QwenCloud’s API compatibility notice states that, starting with qwen4-max-preview, max_tokens is equivalent to max_completion_tokens. Both constrain thinking and the final answer together. Older models retain their previous behavior.
Suppose an existing request sets a 2,000-token ceiling because the desired response is roughly 2,000 tokens. Under the preview’s semantics, reasoning spends from that same allowance. A shorter final answer does not necessarily mean the prompt asked for less; the request may have exhausted its combined budget.
Give reasoning and the answer room within the output ceiling, then measure the resulting usage. Increasing the ceiling blindly can make a slow or expensive run worse, so compare completed task quality and consumption together. Preserve your old model configuration so the evaluation is reversible.
Evaluate the preview without making it indispensable
A useful trial has one representative coding task, a known acceptance check and a recorded configuration. Track whether the output passes that check, how long it takes and how many credits it consumes. Repeat with the same inputs on your current model rather than judging a polished demo in isolation.
Keep the retirement date in the evaluation plan. Save prompts, task definitions and results in a provider-independent form, and decide what happens when this preview ID stops working. The immediate value is learning whether the model solves your work better, without creating a dependency you already know has an end date.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn model releases and API changes into clear stories and practical developer content.