In short (October 23, 2026): The Qwen API is sold through Alibaba Cloud Model Studio. On the International (Singapore) deployment, qwen3.8-max costs $2.00 input and $6.00 output per million tokens, one price across its whole 1M-token context and the same in thinking and non-thinking mode. qwen3.8-flash costs $0.15 and $0.47, and qwen3.8-27b $0.50 and $3.00. The Global deployment is cheaper for the same model names: $1.65 and $4.951 for Max, $0.113 and $0.382 for Flash. Only Singapore International comes with the free quota: 1M tokens per model, input and output counted together, valid 90 days, and billed automatically once used up unless you switch on Free Quota Only. The 90 days can start at a model's release rather than at your sign-up, so check each model's expiry date in the console.
Alibaba lists every Qwen model several times on its pricing page: once per region and deployment scope, sometimes at different prices, sometimes with a free quota and sometimes without. The model name in your request is the same everywhere, so the price you pay depends on which deployment you call. This page sets out the current Qwen3.8 rates, which deployment gets which price, what the free quota actually covers, and the cases where an older Qwen model is still the cheaper choice.
Qwen3.8 API prices as of October 2026
| Model | Released | International (Singapore), input / output | Global, input / output | Free quota |
|---|---|---|---|---|
qwen3.8-max | August 2, 2026 | $2.00 / $6.00 | $1.65 / $4.951 | 1M tokens |
qwen3.8-max-0902 (snapshot) | September 2, 2026 | $2.00 / $6.00 | $1.65 / $4.951 | 1M tokens |
qwen3.8-flash | August 26, 2026 | $0.15 / $0.47 | $0.113 / $0.382 | 1M tokens |
qwen3.8-27b | August 17, 2026 | $0.50 / $3.00 | Not listed (China, Beijing: $0.424 / $1.696) | 1M tokens |
USD per 1M tokens, from Alibaba Cloud Model Studio's model inference pricing page (last updated October 22, 2026). One price from the first token to 1M input tokens per request; thinking and non-thinking mode are billed the same. The free quota applies to the International (Singapore) deployment only.
qwen3.8-max is Alibaba's 2.4-trillion-parameter mixture-of-experts flagship with a native 1M-token context. qwen3.8-max-0902, also callable as qwen3.8-max-2026-09-02, is an upgraded snapshot of it with stronger coding and agent performance. qwen3.8-flash is the fast multimodal model, also with a 1M-token context, and qwen3.8-27b is a 27B dense vision-language model aimed at coding and office tasks.
Which deployment you call changes the price
- International (Singapore): the list prices above, and the only deployment with a free quota.
- Global (served from Hong Kong, Frankfurt, Tokyo or Virginia): 17.5% cheaper for qwen3.8-max on both input and output. For qwen3.8-flash, input is about 25% cheaper and output about 19% cheaper.
- US deployment scope (Virginia): qwen3.8-flash since October 3 and qwen3.8-max since October 8, at the International price of $0.15 / $0.47 and $2.00 / $6.00.
- China (Beijing): qwen3.8-max at $1.65 / $4.951 in the USD listing, the same as Global, and both Max and Flash carry the 50% batch discount there.
The trade-off is simple. The free quota only exists on Singapore International, so it is the place to test. For sustained traffic, the Global listing charges less for the same model name, and on qwen3.8-max that difference is a flat 17.5% of the bill.
The free quota: what you actually get
- 1M tokens per model, with input and output tokens deducted from one shared total, on the Singapore region with the International deployment scope only.
- 90 days from the date you activate Model Studio, the model is released, or your request is approved, whichever is later. The countdown keeps running if you do not use it, and registering a new account does not grant a second quota.
- Each model and each dated snapshot has its own quota. qwen3.8-max and qwen3.8-max-0902 each come with 1M tokens. Leftover quota never moves between models, and the API does not switch to a model that still has quota when one runs out: you change the model name yourself.
- Your Alibaba Cloud account and its RAM users share one quota per model.
- Real-time calls only. Batch calls, fine-tuning and deployed models are billed from the first token.
- Billing starts automatically when the quota is used up or expires. Free Quota Only, which Alibaba also calls worry-free mode, is off by default; with it on, calls fail with
AllocationQuota.FreeTierOnlyinstead of being billed. - Charges are taken in this order: free quota, resource plans, savings plans, then pay-as-you-go.
What the quota is worth: at $2.00 input and $6.00 output, 1M tokens of qwen3.8-max is between $2 and $6 of usage. That is enough to test a pipeline end to end, not to run one.
The start date matters more than the size. Because the 90 days run from the latest of activation, release and approval, an account opened before a model shipped starts that model's clock on its release date: August 2 for qwen3.8-max, August 17 for qwen3.8-27b, August 26 for qwen3.8-flash and September 2 for the 0902 snapshot. Alibaba also notes that its switch to a 90-day validity applied to first activations from September 8 and did not change the terms of accounts activated before that date. The Free Quota tab on the model usage page shows the remaining tokens and the expiry date for every model, and it is the number to trust.
Flat vs tiered: when an older Qwen model is cheaper
The Qwen3.8 models have one price for any request size. Several older models are tiered, and the rule there is strict: the input size of a single request picks the tier, and every token in that request is billed at that tier's price. International (Singapore) prices per 1M tokens:
| Model | Input tokens per request | Input | Output |
|---|---|---|---|
qwen3.8-max | Up to 1M | $2.00 | $6.00 |
qwen3.7-max | Up to 1M | $2.50 | $7.50 |
qwen-max (non-thinking only) | No tiers | $1.60 | $6.40 |
qwen-plus | Up to 256K | $0.40 | $1.20 ($4.00 thinking) |
| 256K to 1M | $1.20 | $3.60 ($12.00 thinking) | |
qwen3.8-flash | Up to 1M | $0.15 | $0.47 |
qwen3.7-flash | Up to 32K | $0.03 | $0.13 |
| 32K to 256K | $0.10 | $0.40 | |
| 256K to 1M | $0.20 | $0.80 | |
qwen-flash | Up to 256K | $0.05 | $0.40 |
| 256K to 1M | $0.25 | $2.00 |
Three things stand out. qwen3.8-max is 20% cheaper than qwen3.7-max on both input and output, so moving up a generation lowers the bill. On the Flash line it is the other way round: for any request under 256K input tokens, qwen3.7-flash is cheaper than qwen3.8-flash on both input and output, and under 32K its input price is a fifth of the newer model's. qwen3.8-flash only wins on price above 256K tokens, where its flat $0.15 / $0.47 beats the older model's top tier. And on qwen-plus, thinking mode more than triples the output price, while the Qwen3.8 models charge nothing extra for it.
Caching and batch
Batch calls cost 50% of the real-time price, for input and output, on models that support them. On the International deployment that includes qwen3.8-flash, qwen3.7-flash, qwen-flash and qwen-max; qwen3.8-max shows the batch discount in the Beijing listing but not in the Singapore one. Batch and cache discounts never combine, batch calls with file input get no cache discount, and the free quota does not cover batch.
Context cache comes in two modes. The implicit cache is always on and bills the cached part of a prompt at, typically, 20% of the input price. The explicit cache is created with a cache_control marker, lasts 5 minutes and resets on each hit, needs at least 1,024 tokens, and is typically billed at 125% of the input price to create and 10% on a hit. The Qwen3.8 models are the exception: Model Studio's cache documentation says their cache-hit price is not the usual 20% or 10%, keeps cache creation at 125% for qwen3.8-max, and points to the console for the exact rates. For a reference point, OpenCode's gateway, which charges the same $2.00 / $6.00 and $0.15 / $0.47 for these two models, lists cache reads at $0.25 and $0.016 per million tokens and cache writes at $2.50 and $0.20, as covered in our OpenCode Zen and Go pricing breakdown. Cached reads are where most of an agent's input goes, so it is worth reading those rates in your own console before you budget, and tracking hits and writes separately, as explained in how prompt caching breaks cost tracking.
What a month costs: a worked example
Take 10,000 requests a month, each with 30,000 input tokens and 1,500 output tokens, no caching and no batch. At 30,000 tokens every request stays inside the first tier of the tiered models:
| Model and deployment | Input | Output | Total per month | Per request |
|---|---|---|---|---|
| qwen3.8-max, International | $600.00 | $90.00 | $690.00 | $0.069 |
| qwen3.8-max, Global | $495.00 | $74.27 | $569.27 | $0.057 |
| qwen3.8-27b, International | $150.00 | $45.00 | $195.00 | $0.020 |
| qwen3.8-flash, International | $45.00 | $7.05 | $52.05 | $0.0052 |
| qwen3.8-flash, Global | $33.90 | $5.73 | $39.63 | $0.0040 |
| qwen3.7-flash, International (up to 32K tier) | $9.00 | $1.95 | $10.95 | $0.0011 |
Three levers move this table more than quality differences between neighbours. The deployment: Global takes 17.5% off qwen3.8-max with no change to your code beyond the endpoint. Request size: the same workload costs about five times less on qwen3.7-flash while prompts stay under 32K, and the advantage disappears the moment a request crosses a tier boundary. And caching, which on an agent that resends a long, stable prefix can remove most of the input line. Whichever you pick, measure cost per completed task rather than per token, as covered in cost per task for workhorse models.
What changed in 2026
- August 2: qwen3.8-max released on the International and Global deployments.
- August 12: the open-weights flagship qwen3.8-2.4t-a95b released: 2.4 trillion parameters, about 95 billion active per step, 1M-token context.
- August 17: qwen3.8-27b released.
- August 26: qwen3.8-flash released.
- September 2: the qwen3.8-max-0902 snapshot released, with its own free quota.
- October 3 and 8: qwen3.8-flash and qwen3.8-max added to the US deployment scope in Virginia.
For how Qwen compares with DeepSeek's off-peak pricing, see DeepSeek API pricing; for choosing a cheap agent model on quality as well as price, see cheap agentic LLMs compared. Day-to-day price changes across providers are logged in our AI price tracker.
Sources
Read on October 23, 2026: Alibaba Cloud Model Studio's model inference pricing page (last updated October 22, 2026); Model Studio's Free quota for new users page, Context Cache page and newly released models list; OpenCode's Console models page for the cache rates quoted above.