Worked examples: estimate your own workload

Model Studio: 1,000 short requests

Assume 1,000 requests, each using 2,000 input and 500 billed output tokens, with no cache hits: 2 million input and 0.5 million output tokens in total. This is a cost calculation, not a performance test.

The tier is determined per request. The 2 million input tokens across this entire workload do not move these individual 2,000-token requests into a long-input tier.

Model Studio: one long request

A Qwen3.7 Plus request with 300,000 input tokens and 10,000 billed output tokens falls in the higher tier: (0.3 × $1.20) + (0.01 × $4.80) = $0.408 before other charges or discounts. Using the short-input rate would underestimate the bill.

QwenCloud: 30,000 monthly requests

Suppose an application makes 30,000 requests a month, each with 2,000 input and 500 billed output tokens. That is 60 million input plus 15 million output tokens:

• Max: (60 × $2) + (15 × $6) = $210.
• Flash: (60 × $0.15) + (15 × $0.47) = $16.05.

These are hypothetical QwenCloud list-price estimates with no cache, free quota, tools or tax adjustment. They are not a quality comparison or a guarantee that 500 output tokens will finish each task. If thinking is enabled, include its billed output as well as the visible answer, without adding the same tokens twice.

Build your own estimate

Model cost = (input tokens × input rate + billed output tokens × output rate) / 1,000,000

Record usage for representative tasks, including history and retries. Separate any cache categories at their own rates, then add tool and infrastructure costs. Use returned usage and the provider’s billing records to check the estimate. An output-token limit does not cap input costs, request count or monthly spending; a spending alert is not an automatic hard stop.

Compare performance on the same workload before choosing a cheaper model.

Our measured QwenCloud usage: 146 tokens

On 25 September 2026, our Qwen API test sent three first-attempt requests to qwen3.8-max: ordinary text, the same date-conversion prompt streamed, and strict JSON extraction. Thinking was disabled and no retries or tools were used.

The $0.000452 figure was not a paid charge. Free quota only was enabled for the tested model, and the console’s consumed amount rose from 0 to 146. Our account showed a one-million-token allowance expiring 25 December 2026; other accounts can have different eligibility and expiry.

All three responses reported zero cached tokens. These tiny tasks do not predict normal conversation cost, monthly capacity or model quality. The full test and downloadable evidence preserve the actual answers, settings and usage.

Is there a free Qwen API allowance?

QwenCloud

QwenCloud documents payment verification for quota activation. Allowances are per model, combining input and output; dated snapshots have separate quotas. They generally expire after 90 days from activation or a later model release. Your console supplies the applicable amount and expiry.

1. Open Benefits / Free Tier and find the exact model ID used by your application.
2. Check its active status, remaining allowance and expiry.
3. Enable Free quota only for that model before testing. It is off by default; enabling it for one model does not protect every other model.
4. Reconcile returned usage with the console, allowing for reporting delays.

With free-only enabled, the provider blocks calls when the allowance ends. Otherwise, a payment-verified account can continue with paid usage. Free quota covers eligible real-time inference, not batch calls, built-in tool fees, fine-tuning, deployment or custom models. It is not an account-wide spending cap.

Source: QwenCloud free-quota rules. Do not disable protection just to work around a quota error.

Alibaba Cloud Model Studio

Alibaba documents a new-user allowance for eligible models in Singapore with International deployment. It is not a universal free allowance for every region or product. Under the current new-activation rules, the 90-day period starts from the latest applicable date: service activation, model release or approval of your application for model access. Older accounts can follow earlier rules.

Check the exact model’s remaining allowance and expiry in your console. Input and output consume the same per-model quota. If your account is eligible for paid billing, usage beyond the allowance can become chargeable.

For a limited trial, check whether the model supports Free Quota Only and enable it before the allowance runs out if you want calls to stop instead of continuing as paid usage. Do not assume it is enabled by default.

Source: Alibaba Cloud free-quota rules. Separate coding/token plans and their dedicated keys follow their own billing arrangements.

How caching and batch affect the bill

QwenCloud: use the exact cache rate

For qwen3.8-max, the published cache prices are $0.25 per million implicit-cache input tokens, $2.50 per million explicit-cache creation tokens and $0.17 per million explicit-cache read tokens. Ordinary input remains $2 and output $6. These are different billing categories, not discounts to stack on every input token.

An illustrative request with 8,000 implicit-cache-hit tokens, 2,000 ordinary input tokens and 500 output tokens costs (8,000 × $0.25 + 2,000 × $2 + 500 × $6) / 1,000,000 = $0.009, versus $0.023 without a cache hit. We did not test this cache scenario. A repeated prompt does not guarantee a hit.

QwenCloud’s billing guide lists a 50% real-time rate for supported Batch API calls and says batch and cache discounts cannot be combined. Check model support; batch is outside the free quota.

Model Studio: context caching

On supported models, cached input can cost less than ordinary input. Alibaba documents typical implicit-cache hits at 20% of the standard input rate. Explicit caching has different economics: creation is typically 125% of the normal input rate, while later hits are typically 10%.

Model-specific exceptions: qwen3.8-max and qwen3.8-flash do not use those generic 20% implicit-hit and 10% explicit-hit rates. The dated qwen3.8-max-0902 is also an exception. Check their current cache-hit prices in the Model Studio console before estimating savings.

Model support and cache rules matter. A repeated prompt does not guarantee an implicit hit. Use the actual cache counts and the model-specific rate when calculating savings; output is still billed separately. Read the official context-cache rules.

Model Studio: batch processing

Eligible Batch File calls are priced at 50% of real-time inference, but support differs by model and region. A model appearing in the ordinary API table does not prove that the same model supports batch in Singapore. Check the supported-model list before applying that discount. Batch File calls do not receive context-cache discounts and are not covered by the new-user free quota. Read the Batch File documentation.

Before you compare two Qwen prices

1. Identify the billing product and the account or plan attached to your key.
2. Match the exact API model ID.
3. Match the service region and deployment scope.
4. Use the input tier for a single request.
5. Separate input, output, cache and tool charges.
6. Check whether the quote is a list price, promotion or different service plan.

The page separates official rates, hypothetical calculations and a dated account observation. Use your own usage and billing records to check the estimate.

Use the Qwen model guide to understand the available families, then follow the Qwen API guide to match your key, region and workspace.

Continue with the Qwen overview, compare Kimi pricing or DeepSeek pricing, and read our research approach. Report a pricing correction.