Short answer: Qwen's cheapest useful text model, qwen3.8-flash, is $0.15 per million input tokens and $0.47 output in Singapore, with a 1M context and a $0.016 cache-hit input rate. qwen3.7-plus is $0.40/$1.60 up to 256K input tokens, then $1.20/$4.80 for the whole request above that line. qwen4-max is $1.50/$4.50. Region matters: the same qwen3.8-flash costs $0.113/$0.382 under the Global deployment scope in Frankfurt, the US, Japan and Hong Kong.
Qwen prices per million tokens
| Model (Singapore, International) | Input | Output | Cache-hit input | Context |
|---|---|---|---|---|
qwen4-max | $1.50 | $4.50 | $0.075 | 1,000,000 |
qwen3.7-plus, input ≤ 256K | $0.40 | $1.60 | $0.08 | 1,000,000 |
qwen3.7-plus, 256K < input ≤ 1M | $1.20 | $4.80 | $0.24 | 1,000,000 |
qwen3.8-flash | $0.15 | $0.47 | $0.016 | 1,000,000 |
qwen3.7-flash, input ≤ 32K | $0.030 | $0.130 | $0.006 | 1,000,000 |
qwen3.7-flash, 32K–256K | $0.100 | $0.400 | $0.020 | 1,000,000 |
qwen3.7-flash, 256K–1M | $0.200 | $0.800 | $0.040 | 1,000,000 |
USD per 1 million tokens, retrieved 19 November 2026 from Alibaba Cloud Model Studio's official model pages and the Model Invocation Pricing page. Cache-hit figures are the published implicit-cache input rates. Explicit cache creation is billed at a premium: on qwen4-max it is $1.875 for a 5-minute cache and $3.00 for a 1-hour cache, with reads at $0.075. Prices vary by region and by deployment scope; Singapore International is shown because it is the free-quota region.
The 256K cliff, explained
Model Studio's tiered billing is not marginal. The unit price is chosen by the total number of input tokens in a single request, and then every token in that request is billed at that tier's price. The documentation is explicit: an input of 100K tokens in a model with 0–32K and 32K–128K ranges is billed entirely at the second-tier price.
That creates cliffs that are easy to miss in a capacity plan. On qwen3.7-plus, a 250K-token request costs 0.25M × $0.40 = $0.10 in input. A 300K-token request costs 0.3M × $1.20 = $0.36. You sent 20% more input and paid 3.6× as much. On qwen3.7-flash the same pattern repeats at 32K and 256K. If your RAG pipeline sometimes retrieves 270K tokens, chunk it below 256K or accept a step change, not a gradual slope.
The practical rule: choose the tier first, then the model. A 300K-token request on qwen3.8-flash ($0.15 flat) is cheaper than the same request on qwen3.7-plus at the 256K+ tier ($1.20). Model quality may justify the jump, but do not pay the tier premium by accident.
The Qwen free quota
Model Studio gives new users a free quota of about 1 million tokens per eligible model, valid for 90 days from activation, model release or approval, whichever is later. It is available only in the Singapore region for models whose deployment scope is International, and it does not cover batch calls, fine-tuning, deployment or storage. Input and output tokens draw from the same combined allowance, and each model has an independent quota that cannot be transferred.
Two operational details matter. First, the countdown does not pause: unused quota still expires at 90 days. Second, pay-as-you-go billing starts automatically when the quota runs out unless you enable the “Free Quota Only” switch, which returns AllocationQuota.FreeTierOnly instead of charging. Alibaba removed free quotas for several models on 15 November 2026, including deepseek-v4.1-flash-1017, qwen3.8-max-0902 and kimi-k3, so re-check the model page rather than trusting an old tutorial. The free LLM API tracker records the dated comparison.
What changes the Qwen bill
Region and deployment scope
This is the largest single lever and the least obvious. qwen3.8-flash is $0.15/$0.47 in Singapore International but $0.113/$0.382 under Global scope in Frankfurt, Virginia, Tokyo and Hong Kong, with cache hits at $0.014 instead of $0.016. qwen3.7-plus follows the same pattern: $0.40/$1.60 versus $0.276/$1.101 below 256K. qwen4-max is $1.50/$4.50 versus $1.413/$4.24. The discount is roughly 20–30% on the Flash and Plus input rates, though only about 6% on qwen4-max, and it comes with a different data-residency footprint.
You cannot take both the Singapore free quota and the Global price on the same request. Decide whether the 90-day trial or the long-run rate matters more, then pin the endpoint and region in configuration so a client library default does not move your traffic.
Caching: implicit is cheap, explicit is a prepayment
Implicit cache hits are the best line on the page. qwen3.8-flash drops from $0.15 to $0.016 per million input tokens, an 89% discount; qwen3.7-plus drops from $0.40 to $0.08. The discount applies only to input, and only to the cached prefix. Output never caches.
Explicit cache is different. You pay 125% of the standard input rate to create a 5-minute cache on qwen4-max ($1.875 vs $1.50) and 200% for a 1-hour cache ($3.00). Reads then cost $0.075. Explicit cache pays off when a large prefix is reused many times inside the TTL, not when it is reused twice. Batch and cache discounts cannot be combined: Model Studio states that the two cannot take effect at the same time.
Batch: 50% off, model and region permitting
Where batch inference is supported, Model Studio bills input and output at 50% of the real-time price. Support is not universal. On the Singapore International scope, qwen3.7-plus lists batch inference as supported, while qwen4-max, qwen3.8-flash and qwen3.7-flash list it as unsupported. Batch also cannot be combined with context caching, so a cache-heavy workload should usually stay synchronous. Check the capability table on the exact model page before designing a nightly batch job.
Thinking mode and output length
qwen4-max prices output as “chain-of-thought + response” in one line. In thinking mode it can produce up to 262,144 output tokens in total, including chain-of-thought, all billed at $4.50 per million. qwen3.7-plus and the Flash models also support thinking, with output capped at 131,072 tokens and chain-of-thought ceilings that differ by model. A reasoning-heavy request can therefore cost several times a non-thinking request with the same visible answer. Set reasoning effort deliberately and measure total output tokens, not answer length.
Worked example: 10M input, 2M output on qwen3.8-flash
| Scenario | Input cost | Output cost | Total |
|---|---|---|---|
| Singapore, no cache | 10M × $0.15 = $1.50 | 2M × $0.47 = $0.94 | $2.44 |
| Singapore, 70% cache hit | 3M × $0.15 + 7M × $0.016 = $0.562 | $0.94 | $1.502 |
| Global scope, no cache | 10M × $0.113 = $1.13 | 2M × $0.382 = $0.764 | $1.894 |
| Global scope, 70% cache hit | 3M × $0.113 + 7M × $0.014 = $0.437 | $0.764 | $1.201 |
The spread from $2.44 to $1.20 is a 51% reduction with no model change, achieved by moving region and structuring the prompt for cache hits. No model upgrade was involved; the entire swing came from region and cache structure. That is why “which Qwen model” is the wrong first question for this workload. “Which region and what cache-hit rate” comes first.
When a flat plan wins
A flat plan wins when your requests stay inside its context ceiling and your monthly token count is steady. qwen3.8-flash at $0.15/$0.47 is cheap per token, but 500 million input and 80 million output tokens still costs $75 + $37.60 = $112.60 in Singapore, or about $87 under Global scope. At that volume, a $60 unlimited-fair-usage plan is cheaper and removes the region and tier arithmetic.
The exception is long context. Plugsky's Builder plan caps input at 256K and Scale at 512K, while Qwen offers 1M on every model in this table. If your workload genuinely needs 600K–1M tokens in one request, no flat plan in this comparison fits, and Qwen's per-token price is the price you pay. Similarly, if you need explicit cache control with a 1-hour TTL for a large shared document, Model Studio exposes a lever that a flat plan may not.
Between those ends, the decision is ordinary: below roughly 300 million blended tokens per month, per-token Qwen is usually cheaper; above it, a flat plan is usually cheaper; above 256K input per request, Qwen is the only option. Plugsky plans start at $8/month, with Builder at $60 and Scale at $120. The cost calculator handles the crossover arithmetic.
How to verify these numbers
- Open the Model Invocation Pricing page and the individual model page; the model page is more current for tier boundaries.
- Confirm the region and deployment scope, not just the model ID. Singapore International is not the same price as Global scope.
- Note the tier boundary before you benchmark. A 260K-token test on qwen3.7-plus is billed at the higher tier for all tokens.
- Check whether batch is supported for that model in that region; the capability table differs by scope.
- Re-check free-quota eligibility, because Alibaba removes quotas from models without changing the model itself.
Use the rate-limit calculator for the RPM/TPM side, the free API tracker for quota comparisons, and the cheapest LLM API guide for a workload-based ranking. The model catalogue lists the Qwen-family options available without a Model Studio account.
FAQ
How much does the Qwen API cost?
In Singapore on 19 November 2026, qwen4-max listed $1.50 input and $4.50 output per million tokens, qwen3.7-plus $0.40/$1.60 up to 256K input tokens and $1.20/$4.80 above that, qwen3.8-flash $0.15/$0.47, and qwen3.7-flash from $0.030/$0.130 at the smallest tier.
Is there a free Qwen API?
Model Studio grants a new-user free quota of about 1 million tokens per eligible model, valid for 90 days. It applies only to the Singapore region with International deployment scope, does not cover batch calls, and combines input and output tokens into one allowance. There is no permanently free production tier.
Why did my Qwen cost jump at 256K tokens?
Model Studio uses tiered billing. All input tokens in a request are billed at the rate of the tier the request falls into, not just the tokens above the threshold. A 300K-token request on qwen3.7-plus is billed entirely at $1.20 per million instead of $0.40, so a 20% larger prompt can cost 3.6 times as much.
Do Qwen prices differ by region?
Yes. Singapore International pricing is higher than the Global deployment scope available in Frankfurt, the United States, Japan and Hong Kong for several models. For example qwen3.8-flash is $0.15/$0.47 in Singapore and $0.113/$0.382 under Global scope. The free quota is tied to Singapore.
Need Qwen without a Model Studio account?
Plugsky is OpenAI-compatible, starts free with no card, and serves 39 models on one endpoint. Flat plans when your volume is predictable.
Start Free → See plans