Qwen API cost calculator

Alibaba Model Studio prices Qwen per token with separate input, output and cached-input rates. The prices here are the international (Singapore) list; the China region has its own list and model lineup. The context cache is implicit: a repeated prompt prefix is matched automatically and a hit is billed at 10.67%–20% of the input price. In our data Qwen3.8 Max, Qwen Max, Qwen3.7 Plus, Qwen Turbo list no cache-write fee, so a miss costs normal input, while Qwen3.8 Flash charge 1.33× the input price to write.

Some Qwen models have a Batch API at 50% off input and 50% off output for asynchronous jobs; models without a batch price are flagged in the table. Thinking mode is priced apart on Qwen Turbo: its thinking output costs 2.5× the standard output price, which the calculator applies when the output price switch is set to thinking. Long inputs cost more on tiered models: long-context rates apply above 256K for Qwen3.7 Plus and 2 more (input 1.25×–5×, output 1.25×–5×), to the whole request.

The table prices one workload on each model: 10,000 requests a day with 6,000 input and 500 output tokens per request. Columns: no caching; 70% of input read from the cache and 5% written to it; every request sent through the Batch API. For the same traffic Qwen3.8 Max costs 38× as much as Qwen Turbo.

Estimates from published per-token rates; check official pricing before budgeting. Data synced from the LiteLLM price list 2026-10-12 08:00 UTC.

Cost per 1,000 requests

For 1,000 requests of 6,000 input and 500 output tokens, Qwen Turbo costs $0.40 at standard rates, $0.23 with caching and $0.20 through the Batch API; Qwen3.8 Max costs $15.00, $7.65 and $15.00. Multiply by your own request count, or put your traffic into the calculator.

How to cut the bill, in order

1. Put the system prompt, tools and documents first so the implicit cache hits on every request; only the user turn should change.
2. Use Flash for routing and classification, Plus for chat and RAG, Max only for hard questions.
3. Turn thinking mode off where it is not needed: on models with a separate thinking price the output tokens cost more, and reasoning tokens are billed as output.
4. Keep requests under the tiered models' long-context threshold with retrieval; count prompts with the token counter.
5. Send offline work through the Batch API where the model has one, and check the caching calculator for your hit rate before budgeting in the calculator.

When does caching pay off for your traffic? Use the prompt caching calculator →