Latest in AI
Qwen4-Max API Pricing: Cache Costs and the Move From Preview
The full release adds pay-as-you-go access, but the cheapest advertised input rate only applies when a request actually reads the cache.
Alibaba released Qwen4-Max on November 16 with pay-as-you-go API access and a one-million-token context window. In Singapore, standard pricing is $1.50 per million input tokens and $4.50 per million output tokens. The $0.075 cached-input rate is attractive, but it is not the price of every prompt.
The official announcement also reports lower token consumption than Qwen3.8-Max in internal evaluations. That makes this worth evaluating for agents that repeatedly read the same documents. It does not establish what your application will cost, because reasoning length, cache hits and retries all change the bill.
Qwen4-Max API prices, including cache writes
QwenCloud's pricing table lists the following Singapore rates in US dollars per million tokens, checked on November 16. Output billing includes reasoning as well as the final answer.
| Charge | Qwen4-Max, Singapore |
|---|---|
| Ordinary input | $1.50 |
| Output | $4.50 |
| Explicit cache write, five-minute lifetime | $1.875 |
| Explicit cache write, one-hour lifetime | $3.00 |
| Explicit or implicit cache read | $0.075 |
Qwen3.8-Max's listed Singapore input and output prices are $2 and $6. The new model is therefore 25% cheaper on those two unit rates. That is a clean comparison; a claim about total task savings needs actual token usage too.
Prices vary by region and deployment scope. The Beijing and Global tables list $1.413 input and $4.24 output for Qwen4-Max. Do not copy Singapore's rates into a forecast for a different endpoint, or change processing location just to save a few cents without checking your data requirements.
A worked example: repeated document questions
Suppose an application makes 100 requests. Each contains 20,000 input tokens and produces 2,000 output tokens, including any reasoning. With no cache, that is two million input tokens and 200,000 output tokens: $3.90 in token charges at Singapore rates.
Now assume 18,000 input tokens are an identical prefix. The first request writes that prefix into the five-minute cache, and the remaining 99 requests successfully reuse it. Cache writes cost $0.03375, cache reads cost $0.13365, the changing input costs $0.30, and output still costs $0.90. Total: about $1.37, roughly 65% less than the uncached example.
These are calculated scenarios, not measured results or a promise of savings. They exclude tools, taxes and other charges. If the prefix changes, expires or does not hit, the result changes. If the model produces 10,000 output tokens instead of 2,000 on every request, output alone costs $4.50. A cheap input path cannot compensate indefinitely for a long reasoning loop.
Make cache reuse deliberate
The Qwen4 explicit-cache guide distinguishes guaranteed matching-prefix reuse within the cache lifetime from implicit caching, where hits are not guaranteed. Explicit caching requires an identical prefix of at least 1,024 tokens. Put stable instructions, tool definitions and reference material first, then the changing question.
For example, a document assistant can keep the same policy manual before its cache breakpoint and append each user's question afterward. Placing a fresh timestamp or request ID before that manual makes the prefix different. This is an application-layout decision, not merely a provider setting.
The feature supports Chat Completions, Responses and Anthropic Messages, but not the DashScope API. Five minutes is the default lifetime; one hour costs more to write. Inspect returned cache-read and cache-write usage rather than assuming a successful response was a cache hit. Choose the longer lifetime only when the expected reuse justifies it.
The full model and the preview are separate choices
The current model guide identifies the release as qwen4-max, available through Token Plan or pay-as-you-go. The earlier qwen4-max-preview has separate preview benefits. Keep the model ID, credential type and endpoint together when changing a client; a subscription key is not a drop-in replacement for an ordinary API key.
The full model accepts text, images and video and returns text. Its documented maximum is 262,144 tokens for thinking plus the final answer, or 131,072 output tokens without thinking. Do not carry the preview's 393,216-token combined ceiling into the new model's configuration. Batch inference is not supported.
The Singapore OpenAI-compatible base URL is https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1. Use a key created in the matching region. Other regions have their own endpoints and availability, so an authentication problem is not automatically a model problem.
Set reasoning limits before comparing bills
QwenCloud's thinking documentation says Qwen4-Max enables thinking by default. The model-specific guide lists high as the default effort. max_completion_tokens covers reasoning and the answer together, which means a ceiling chosen only for the visible response can truncate useful work.
Start with a representative task and a clear acceptance check. Record total input, cached input, reasoning and answer tokens, elapsed time, and whether a retry was necessary. Compare completed tasks under a recorded configuration, not two screenshots with different effort settings.
This is a documentation-based cost guide, not firsthand model testing. The sensible adoption question is whether Qwen4-Max lowers the cost of an accepted result. A cheaper token rate is useful evidence, but a workflow that repeatedly fails its acceptance check is still expensive.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams explain model launches, pricing changes and developer workflows in content people can actually use.