Qwen4-Max API Pricing: Cost per Token, Cache Math, and Token Plan Compared
Qwen4-Max pricing on the QwenCloud pay-per-token API is $1.50 per 1M input tokens and $4.50 per 1M output tokens in the Singapore region, with cached input at $0.075 per 1M. Thinking and non-thinking modes cost the same, and there are no input-length tiers. If you'd rather pay a flat monthly fee for coding tools, the same model is on Token Plan, starting at $6 a month.
Below: the rate card, real task costs, cache savings, Token Plan vs pay-per-token, and rival prices. New to the model? Start with what Qwen4-Max is.
Qwen4-Max API price per 1M tokens
Qwen4-Max is billed per million tokens, with separate rates for input, output, cache writes and cache reads. Output covers both reasoning ("thinking") tokens and the answer. Alibaba publishes a Singapore (International) column and a lower Beijing/Global column.
| Billing item (per 1M tokens) | Singapore (International) | Beijing / Global |
|---|---|---|
| Input | $1.50 | $1.413 |
| Output (thinking + answer) | $4.50 | $4.24 |
| Cache write, 5-minute TTL | $1.875 | $1.767 |
| Cache write, 1-hour TTL | $3.00 | $2.827 |
| Cache read (explicit or implicit hit) | $0.075 | $0.071 |
Source: the QwenCloud pricing docs and the Qwen4-Max model page, both checked on November 16, 2026. The docs list the Global deployment in Frankfurt, Hong Kong, Tokyo and US (Virginia).
A few rules shape the bill:
-
Thinking tokens are output. Turn on
enable_thinkingand every reasoning token is billed at $4.50 per 1M. - No long-context surcharge. A 900K-token request costs the same per token as a 9K one.
- Cache reads are 5% of the input price. $0.075 against $1.50 is a 95% discount on repeated prefixes.
- Limits: a 1M-token context, up to 991K input (983K with thinking), up to 393K output and 5,000,000 tokens per minute.
The model id is qwen4-max. If you haven't made a call yet, our Qwen4-Max API guide covers keys, base URLs and the first request.
What a real Qwen4-Max task costs
Per-token prices are hard to feel. Alibaba gives one concrete number in its launch post: on its Work360 agent benchmark, the average cost per task fell from $4.43 with Qwen3.8-Max to $1.67 with Qwen4-Max, 62% lower. Alibaba also reports that Qwen4-Max used about 24% fewer tokens at the Xhigh effort level and about 35% fewer at High on Work360. Those are vendor figures, so treat them as a direction, not a quote for your workload.
You can estimate your own cost with simple arithmetic. Here are three worked examples at Singapore list prices:
- Short chat call. 2,000 input tokens and 500 output tokens: 2,000 x $1.50/1M + 500 x $4.50/1M = $0.003 + $0.00225 = $0.0053. That's about $5.25 per 1,000 calls.
- Same call with thinking on. Say the model reasons for 6,000 tokens before the 500-token answer. Output becomes 6,500 tokens, so the call costs $0.003 + $0.02925 = $0.032, about six times more.
- Long-document question. 300,000 input tokens and 2,000 output tokens: $0.45 + $0.009 = $0.459 without cache, or about $0.032 if the 300K document is a cache hit.
Example 2 surprises people: thinking can dominate the bill even when the answer is short. For deeper numbers on quality per dollar, see the Qwen4-Max benchmarks breakdown.
How cache rates change the Qwen4-Max bill
Context caching is where Qwen4-Max pricing really moves. QwenCloud offers two modes for qwen4-max on the pay-per-token API, and on the Chat Completions, DashScope and Anthropic-compatible APIs they're mutually exclusive per request.
| Mode | How it works | Price | Hit guaranteed? |
|---|---|---|---|
| Implicit cache | Automatic, can't be disabled, reuses a common prefix of 1,024+ tokens | $0.075 per 1M on a hit | No. Validity is indeterminate |
| Explicit cache | You mark a breakpoint with cache_control
|
$1.875 (5 min) or $3 (1 h) to write, $0.075 to read | Yes, while the cache is valid |
The context cache docs say an explicit 5-minute cache resets its timer on each hit, so a busy agent loop keeps it warm.
Now take an agent session: a 200K-token system prompt plus tool definitions and repo context, reused across 50 turns, with 2K new input and 3K output per turn.
- No cache hits: 10.1M input tokens x $1.50 + 150K output x $4.50 = $15.83.
- Implicit cache hits from turn 2: the first 202K at full price ($0.30), 9.8M cached tokens at $0.075 ($0.74), 98K fresh tokens ($0.15) and output ($0.68) = $1.86.
- Explicit 5-minute cache: a $0.375 write instead of the first full-price prefix, the same reads and output = about $1.94, with a guaranteed hit.
The same session on Qwen3.8-Max, at its $2 input, $0.25 implicit-cache and $6 output rates, comes to $3.95. That's where Alibaba's claim of the "biggest cut on cached input" shows up: Qwen3.8-Max charges $0.25 for an implicit hit, Qwen4-Max charges $0.075.
Use explicit caching when you need a predictable hit and implicit caching for everything else. Check what you got in usage.prompt_tokens_details.cached_tokens and cache_creation_input_tokens. If you're weighing caching economics across vendors, our Claude Opus 5.5 prompt caching math uses the same method.
Is there a Qwen4-Max batch discount?
Yes, but only in one region. QwenCloud's Batch API normally bills input and output at 50% of the real-time price. For qwen4-max, the batch docs list it only under China North 2 (Beijing), at $0.707 input and $2.12 output per 1M tokens for Batch File. Batch Chat in Beijing is listed at $1.413 / $4.24.
Batch doesn't support context caching, doesn't stack with other discounts and isn't covered by the free quota. With high cache hit rates, the real-time API can beat batch.
Qwen4-Max pay-per-token vs Token Plan
Token Plan is QwenCloud's monthly subscription for coding and agent tools such as Claude Code, Cursor, Qwen Code, Codex and Qoder. It bills in Credits, not tokens, and uses its own sk-sp- API keys. qwen4-max is on the supported-model list. Individual plans, per the Token Plan docs on November 16, 2026:
| Plan | Price per month | Credits per month | Concurrent agents |
|---|---|---|---|
| Lite | $6 ($3 limited-time new-user price) | 20,000 | 1-2 |
| Essential | $10 ($5 limited-time new-user price) | 45,000 | 2-3 |
| Standard | $18 | 80,000 | 3-4 |
| Pro | $43 | 200,000 | 6-8 |
Extra Credit Packs cost $15 for 20,000 Credits (up to 5 at a time). When Credits run out, calls are blocked; they don't fall back to pay-per-token.
How to choose:
- Pick Token Plan if one person drives the model through a coding agent every day and wants a fixed monthly cost.
- Pick pay-per-token for products, backend jobs, batch, fine-tuning or spiky traffic. You pay for what you use and keep the 5,000,000 TPM limit.
Credits per Qwen4-Max request aren't a fixed token ratio, so check the console after a day of real use. We cover the plan in depth, including how far the Credits stretch, in Qwen4-Max Token Plan: what it costs and is it worth it.
Qwen4-Max pricing vs Qwen3.8-Max and rival flagships
Qwen4-Max is cheaper than its predecessor on input, output, cache reads and the 5-minute cache write (Qwen3.8-Max lists one $2.50 cache-write price; Qwen4-Max's 1-hour write is $3). Here's how it sits against other flagships at standard list prices, each read on the vendor's own page on November 16, 2026.
| Model | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| Qwen4-Max (Singapore) | $1.50 | $0.075 | $4.50 |
| Qwen3.8-Max | $2.00 | $0.25 (implicit), $0.17 (explicit read) | $6.00 |
| DeepSeek V4 Pro (peak / off-peak) | $1.32 / $0.66 | $0.044 / $0.022 | $3.96 / $1.98 |
| OpenAI gpt-5.6-sol | $4.00 | $0.40 | $20.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
Sources: Qwen3.8-Max model page, DeepSeek pricing, OpenAI API pricing and Claude pricing.
For agents, output is often the biggest line, and Qwen4-Max's $4.50 output sits well under the $20 that gpt-5.6-sol and Claude Opus 5.5 charge. DeepSeek V4 Pro is cheaper per token, especially off-peak. Cost per finished task depends on how many tokens each model spends on your workload. For the Qwen3.8-Max side in detail, see our Qwen 3.8 pricing guide and the Qwen4-Max vs Qwen 3.8 comparison.
Let your coding agent measure Qwen4-Max costs with Apidog CLI
The only reliable cost estimate is usage measured on your own prompts. Let your coding agent (Claude Code, Codex, Cursor or similar) build and run that check with Apidog CLI. Set it up once:
npm install -g apidog-cli@latest
apidog auth login --with-token <YOUR_APIDOG_TOKEN>
apidog skill install
Then give your agent one prompt:
In Apidog project <projectId>, create an environment "qwencloud-sg" with base URL
https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1 and a secret
variable for my DashScope API key. Add a POST /chat/completions endpoint for model
qwen4-max and two test cases: one with enable_thinking true, one false. Assert
status 200, a non-empty choices array and a usage object. Run both and report
prompt_tokens, completion_tokens and cached_tokens side by side.
The agent checks each payload shape against the CLI schema before it writes anything, then runs the cases:
apidog environment create qwencloud-sg --project <projectId> --base-url https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1
apidog cli-schema get endpoint-create
apidog cli-schema validate endpoint-create --file endpoint.json
apidog endpoint create --project <projectId> --file endpoint.json
apidog cli-schema get test-case-create
apidog cli-schema validate test-case-create --file case-thinking.json
apidog test-case create --project <projectId> --file case-thinking.json
apidog run --test-case <caseId> --project <projectId> -e <envId> -r cli,json
Compare the two completion_tokens numbers and multiply by $4.50 per 1M: that's your thinking surcharge, on your prompts. Add --upload-report to the apidog run command to view the run in the Apidog app, or put the same command in CI with -r junit as shown in running Apidog CLI in CI/CD. New to the CLI? The Apidog CLI installation guide and Apidog CLI in Codex get you running in minutes. Apidog's Free plan covers up to 4 users.
FAQ
How much does Qwen4-Max cost per 1M tokens? $1.50 for input and $4.50 for output in the Singapore region, with cache reads at $0.075. The Beijing/Global column is slightly lower at $1.413 and $4.24.
Does thinking mode cost more on Qwen4-Max? The rate is the same, but reasoning tokens are billed as output at $4.50 per 1M, so long reasoning multiplies the cost of a call.
Is Qwen4-Max cheaper than Qwen3.8-Max? Yes. Input drops from $2 to $1.50, output from $6 to $4.50, and an implicit cache hit from $0.25 to $0.075 per 1M tokens. Alibaba also reports Work360 cost per task falling from $4.43 to $1.67.
Can I use Qwen4-Max for free? QwenCloud gives new models a free quota in the Singapore region, typically 1M tokens valid for 90 days, after you add payment details. Check the Free Tier page in the console for qwen4-max, and see our Qwen4-Max free access guide for every $0 route.
Does Qwen4-Max support batch pricing? Only in the Beijing region: Batch File is $0.707 input and $2.12 output per 1M tokens, with no context caching.
Should I pick Token Plan or pay-per-token for Qwen4-Max? Token Plan suits one person using coding tools daily, from $6 a month. Pay-per-token suits apps, backend jobs and batch work.


Top comments (0)