Quick Answer
Qwen API pricing in 2026, per million tokens on Alibaba Cloud Model Studio (International): the flagship Qwen3.8-Max costs $2 input and $6 output, flat across 1M context. Qwen3.7-Plus runs $0.40/$1.60 below 256K input. Tiered models bill the whole request at its tier's rate, regions price differently, and cache hits cost 10% of input.
On August 3, 2026, Alibaba did something flagships almost never do: it launched one that costs less than its predecessor. Qwen3.8-Max arrived at $2 per million input tokens and $6 output, under the $2.50/$7.50 that Qwen3.7-Max still lists today, with the same flat 1M-token context, and four times the context of the Qwen3-Max generation it supersedes.
Cheaper, newer, bigger is a great headline, and it’s also where the simplicity ends. The Qwen catalog spans well over a hundred model IDs, several bill on tiered pricing where one extra token can reprice an entire request, and the same model costs different amounts by region and deployment scope. Some rates even drop 60% at night.
Alibaba publishes all of it, in tables long enough that pricing roundups routinely disagree about which generation is current.
That confusion is exactly the class of problem CloudZero exists for: it was the first cloud cost platform to integrate directly with Anthropic’s Usage and Cost Admin API, it connects cloud, SaaS, and AI spend into one allocated view, and its operating premise is that a rate you can’t see landing on your bill is a rate you can’t manage.
What is the Qwen API?
Qwen is Alibaba Cloud’s model family (text, coding, vision, audio, and embeddings) served through Model Studio with OpenAI-compatible endpoints, so existing SDKs point at it with a URL change. Most Qwen models also ship open weights, which means the API is one of two ways to run them: pay per token on Alibaba’s infrastructure, or self-host on your own GPUs and pay for hardware instead.
One analysis put the economics behind the low rates plainly: the sparse mixture-of-experts design fires a fraction of parameters per token, so, as eesel’s teardown quoted it, the pricing “isn’t subsidized; it’s structurally earned.” That matters for whether the rates hold. Pricing built on how the model runs is harder to withdraw than a promotional discount.
The current card, from Alibaba’s own pricing page (International/Singapore rates, per million tokens):
| Model | Input | Output | Context notes |
|---|---|---|---|
| Qwen3.8-Max (flagship) | $2.00 | $6.00 | Flat to 1M; thinking and non-thinking same price |
| Qwen3.7-Max | $2.50 | $7.50 | Flat to 1M |
| Qwen3.7-Plus | $0.40 | $1.60 | Below 256K; $1.20/$4.80 above; 20% limited-time discount live |
| Qwen3.6-Plus | $0.50 | $3.00 | Below 256K; $2/$6 above |
| Qwen3.5-Plus | $0.40 | $2.40 | Below 256K; $0.50/$3 above |
| Qwen3-Max (legacy tiered) | $1.20 to $3.00 | $6 to $15 | Three tiers by request size |
| Qwen3.8-Flash | $0.15 | $0.47 | Flat to 1M |
Cache pricing sits on top, and the two modes bill differently. With explicit cache, creating the cache costs 125% of the standard input rate and hits cost 10% of it. With implicit cache, the automatic one a stable prompt prefix triggers, hits are typically billed at 20% of the input rate, with per-model exceptions. Batch calls run 50% off where supported, and the two discounts never stack. All of that is from Alibaba’s documentation.
The flagship’s $2 input undercuts Western flagship tiers by multiples while the Plus line competes with budget tiers, which is why the family keeps showing up in cost-comparison shortlists it wasn’t on a year ago.
How does Qwen’s tiered pricing actually work?
Here’s the rule that generates the surprise bills: on tiered models, the tier is set by total input tokens in the request, and every token in the request bills at that tier’s rate. Not the marginal tokens. All of them.
On legacy Qwen3-Max, a 32,000-token request bills input at $1.20 per million. Add one thousand tokens and the 33,000-token request lands in the second tier at $2.40, so input cost per token just doubled for the entire request: $0.038 became $0.079 for one thousand extra tokens.
Long-context work on tiered models therefore needs request-size discipline, or a non-tiered model like the 3.8 generation, whose flat-to-1M pricing removes the cliff entirely.
Why does the same Qwen model cost different amounts by region?
Because Alibaba prices by region and deployment scope, and the spreads are material. Qwen3.8-Max lists at $2/$6 on the International (Singapore) endpoint but $1.65/$4.95 on Global-scope rows, which appear in the Beijing, Hong Kong, Frankfurt, Virginia, and Tokyo tables. Legacy Qwen3-Max stretches further: $1.20 input International against $0.359 on Global scope, a 70% gap on the same model, though the Global row is non-thinking-mode only.
Then there’s the meter almost nobody covers: some regions carry night discounts, up to 60% off list between 22:00 and 08:00 UTC+8 on flagged models. Time-of-day pricing has arrived in LLM APIs. Batch jobs that can wait until the discount window are the cheapest tokens available on Qwen.
Read the region tables like a procurement exercise. That is what they are.
Is Qwen free?
The API gives new Model Studio accounts 1 million free tokens per eligible model, valid 90 days, on the Singapore endpoint only; other regions get none. Because the quota is per model across a very large catalog, systematic evaluation adds up to tens of millions of free tokens, one of the most generous trial structures anywhere.
Two caveats: the old free OAuth tier was discontinued April 15, 2026, so tutorials promising unlimited free API access are describing a dead program. And the open weights remain genuinely free to download; what you pay for then is your own inference hardware, which is a different bill, not no bill.
Qwen vs. DeepSeek: which is cheaper?
This is the value-tier matchup, and the honest answer depends on the workload. DeepSeek still holds the lowest headline rates in the market, covered in our DeepSeek pricing guide, while Qwen counters with flat-to-1M pricing on its current generation, the deepest free evaluation quota, and a broader model catalog.
Both now run time-of-day pricing on some routes, both ship open weights, and both are hosted from China, which makes data jurisdiction part of the price for regulated US buyers.
The deciding number is never the rate card: it’s cost per accepted task on your workload, with retries, cache hits, and output length counted. Spend a day on the free quota and you will have that number. The same discipline applies across Mistral and every other value-tier contender.
What does a real Qwen workload cost?
Worked example: a support automation running 500,000 requests a month on Qwen3.7-Plus, averaging 3,000 input and 500 output tokens, stays in tier one. Input: 1.5B tokens at $0.40 = $600. Output: 250M at $1.60 = $400. Total about $1,000 a month at list, or $800 while the 20% promotion runs. Caching is a separate lever on the same list bill: push two-thirds of input onto implicit cache hits at 20% of the input rate and input falls from $600 to $280, taking the list total to about $680.
Caching discipline cut roughly a third off this bill before model choice entered the picture.
How do teams keep Qwen spend accountable?
Qwen API spend bills through Alibaba Cloud, which means it lands outside the AWS, Azure, and GCP bills where most cost tooling looks, in tables whose tier and region logic no invoice line explains. Multi-model teams routing across Qwen, token-priced Western APIs, and gateways end up with the same model’s cost scattered across three bills.
CloudZero’s answer is the coverage pattern it built its record on: native connections where they exist, and AnyCost for everything else, Alibaba Cloud exports included. Qwen tokens land in the AI Hub beside OpenAI, Anthropic, and the rest of the model market, allocated per team and product, with unit costs a CFO can trace back to a team and a product.
Want to see multi-provider model spend in one allocated view? Schedule a demo and ask for exactly that.
Frequently asked questions about Qwen API pricing
