Qwen pricing depends on the provider, region, deployment scope, exact model ID, input-token tier, output mode, and whether a request uses eligible caching or Batch inference. There is no single worldwide price for every service that offers a Qwen model.
This page separates QwenCloud PAYG, QwenCloud Token Plan, and Alibaba Cloud Model Studio regional pricing. It is a dated, independently checked reference—not a billing quote or a replacement for the official provider table and your account console.
Pricing verification
Last verified: August 23, 2026.
Currency and unit: US dollars per 1 million tokens for pay-as-you-go tables, unless a section says otherwise.
Route boundary: QwenCloud PAYG and Alibaba Cloud Model Studio regional prices are shown separately. QwenCloud Token Plan uses plan credits; credits are not converted here into an invented PAYG dollar-per-token rate.
Price type: public list prices in the official English documentation. Promotions, account credits, taxes, exchange rates, negotiated terms, and console-specific offers are excluded.
Independent-site notice: qwen-ai.chat is not QwenCloud, Alibaba Cloud, the Qwen team, or a billing provider. Confirm the route, model, region, plan, and effective price in the relevant console before production use.
Official Qwen3.7 output-limit discrepancy
Checked August 4, 2026: QwenCloud’s Marketplace cards for qwen3.7-max, qwen3.7-plus, and qwen3.7-flash displayed 131K maximum output. The QwenCloud text-model matrix displayed 64K for the corresponding rolling IDs, while the vision-model matrix displayed 64K for Plus, Flash, and the listed Max snapshot. Neither path exposed a visible revision date. Until QwenCloud reconciles them, budget and configure against a conservative output ceiling and verify the exact endpoint, protocol, account, deployment scope, and accepted output parameter. No authenticated request was made, and this site did not independently live test the maximum.
QwenCloud list prices versus promotions
Checked August 4, 2026: the QwenCloud PAYG pricing guide identifies its figures as list prices and directs readers to Marketplace for promotions. Its list-price baseline was $2.50 input / $7.50 output per 1M tokens for Qwen3.7 Max; Qwen3.7 Plus was $0.40 / $1.60 through 256K input and $1.20 / $4.80 above 256K through 1M. The model cards separately displayed a temporary 50% Max promotion ($1.25 / $3.75) and 20% off the Plus first band ($0.32 / $1.28). Promotions can change or end; keep the list price as the stable comparison baseline and label every observed discount with its date, route, and scope.
Contents
- Qwen pricing: quick answer
- Which Qwen service are you pricing?
- Model Studio PAYG prices by region
- Context tiers and token billing
- Worked cost examples
- Thinking-mode cost
- Cached-input pricing
- Batch pricing
- Legacy prices and retirement status
- Vision pricing
- Free quota and billing profile
- How to reduce Qwen API cost
- Frequently asked questions
Qwen pricing: quick answer
Qwen pricing starts with the service route. QwenCloud’s live Marketplace lists qwen3.8-max and qwen3.8-2.4t-a95b at $2 per 1M input tokens and $6 per 1M output tokens, and qwen3.8-27b at $0.50 input and $3 output. QwenCloud Token Plan is a separate Credits-based subscription route for interactive programming and agent tools. Alibaba Cloud Model Studio publishes region/workspace-specific pricing; do not copy a QwenCloud Marketplace amount into a Model Studio regional budget. See the Qwen3.8 Max guide for verified capabilities and lifecycle boundaries.
- Production Qwen3.8 Max PAYG: $2 input / $6 output per 1M tokens.
- Qwen3.8-2.4T-A95B on QwenCloud: $2 input / $6 output per 1M tokens; $0.25 implicit-cache input, $2.50 explicit-cache creation, and $0.17 explicit-cache read.
- Qwen3.8-27B on QwenCloud: $0.50 input / $3 output per 1M tokens; $0.10 implicit-cache input, $0.625 explicit-cache creation, and $0.05 explicit-cache read. QwenCloud documented this hosted route on August 19, 2026.
- Cache: $0.25 implicit-cache input, $2.50 explicit-cache creation, and $0.17 explicit-cache read per 1M tokens.
- Retired preview alias:
qwen3.8-max-previewhas ended its preview period. QwenCloud still accepts the old ID temporarily, but automatically routes requests to the productionqwen3.8-max; Credits and usage statistics are recorded asqwen3.8-max. Update configurations to the production ID. - Regional boundary: confirm provider, region, deployment scope, input tier, tools, and account terms before budgeting.
Which Qwen service are you pricing?
| Service route | Current starting point | Billing boundary |
|---|---|---|
| QwenCloud PAYG | Production qwen3.8-max, hosted qwen3.8-2.4t-a95b, and hosted qwen3.8-27b; Qwen3.7 models remain available where documented |
International QwenCloud key and Marketplace/PAYG billing. Qwen3.8 Marketplace pricing is not a Model Studio regional quote. |
| QwenCloud Token Plan | The allowlist includes production qwen3.8-max; legacy calls using qwen3.8-max-preview are temporarily routed to it |
Plan Credits and a dedicated key; interactive compatible tools only, not automated scripts or application backends. |
| Alibaba Cloud Model Studio | The current matrix recommends qwen3.8-max, qwen3.7-plus, and qwen3.7-flash |
Regional workspace, deployment scope, key, and billing table. The exact Qwen3.8 Max model page now publishes region-specific list prices; use the matching workspace row. |
| Third-party provider | Provider-specific model resource | Use that provider’s catalog and pricing. |
| Self-hosted open weights | Exact public checkpoint and license | No hosted token fee, but infrastructure and operations remain costs. |
If you are comparing providers rather than Qwen routes, use Qwen vs ChatGPT to separate assistant subscriptions from API billing, and Qwen vs Claude for API, self-hosting, and workflow-cost trade-offs.
Permitted-use boundary: Token Plan is for interactive use in compatible programming and agent tools. It cannot be used for automated scripts, application backends, or non-interactive batch processing. QwenCloud recommends pay as you go for applications and API integrations.
Qwen3.8 Max production price and retired-preview routing
The model-specific QwenCloud Marketplace displayed the following public list prices for production qwen3.8-max on August 23, 2026. All amounts are US dollars per 1 million tokens:
| Billing item | Price |
|---|---|
| Input | $2.00 |
| Output | $6.00 |
| Implicit-cache input | $0.25 |
| Explicit-cache creation | $2.50 |
| Explicit-cache read | $0.17 |
qwen3.8-max-preview is retired. QwenCloud currently accepts the legacy ID only as a compatibility route to production qwen3.8-max, and records its Credits and usage statistics under the production model. These QwenCloud Marketplace amounts must not be copied into an Alibaba Cloud Model Studio regional row without exact regional evidence.
Qwen3.8 Max worked example
100,000 uncached input tokens and 20,000 output tokens cost (0.1 × $2) + (0.02 × $6) = $0.32, excluding tools, retrieved content, retries, taxes, promotions, and application infrastructure.
QwenCloud PAYG prices for current models
Public QwenCloud international list prices, in USD per 1 million tokens:
| Model | Input tokens per request | Input | Output |
|---|---|---|---|
qwen3.8-max |
Up to 991K input | $2.00 | $6.00 |
qwen3.8-2.4t-a95b |
Up to 991K input; 1M context | $2.00 | $6.00 |
qwen3.8-27b |
Up to 991K input; 1M context | $0.50 | $3.00 |
qwen3.7-max |
0–991K | $2.50 | $7.50 |
qwen3.7-plus |
Up to 256K | $0.40 | $1.60 |
qwen3.7-plus |
Over 256K–1M | $1.20 | $4.80 |
qwen3.7-flash |
Up to 32K | $0.03 | $0.13 |
qwen3.7-flash |
Over 32K–256K | $0.10 | $0.40 |
qwen3.7-flash |
Over 256K–1M | $0.20 | $0.80 |
The QwenCloud Marketplace also publishes model-specific cache rates per 1 million tokens:
| Model | Implicit-cache input | Explicit-cache creation | Explicit-cache read |
|---|---|---|---|
qwen3.8-max | $0.25 | $2.50 | $0.17 |
qwen3.8-2.4t-a95b | $0.25 | $2.50 | $0.17 |
qwen3.8-27b | $0.10 | $0.625 | $0.05 |
Do not combine these QwenCloud prices with the Alibaba Cloud Model Studio regional table below.
Alibaba Cloud Model Studio PAYG prices by region
Qwen3.8 Max regional list prices
Alibaba Cloud’s exact qwen3.8-max model page publishes the following original list prices, excluding promotions. Prices are USD per 1 million tokens and must be matched to the workspace region.
| Region and scope | Input | Output | Implicit-cache input | Explicit-cache creation | Explicit-cache read |
|---|---|---|---|---|---|
| Singapore / International | $2.00 | $6.00 | $0.25 | $2.50 | $0.17 |
| Beijing, Frankfurt, Virginia, Tokyo, or Hong Kong | $1.65 | $4.951 | $0.206 | $2.063 | $0.137 |
The comparison table below keeps the verified regional list prices for qwen3.7-max, qwen3.7-plus, and qwen3.6-flash. Every cell is input / output in USD per 1 million tokens. Do not combine Model Studio regional rows with QwenCloud Marketplace billing.
| Region and scope | qwen3.7-maxUp to 1M |
qwen3.7-plusUp to 256K |
qwen3.7-plusOver 256K–1M |
qwen3.6-flashUp to 256K |
qwen3.6-flashOver 256K–1M |
|---|---|---|---|---|---|
| Singapore International |
$2.50 / $7.50 | $0.40 / $1.60 | $1.20 / $4.80 | $0.25 / $1.50 | $1.00 / $4.00 |
| China (Beijing) Chinese mainland |
$1.65 / $4.951 | $0.276 / $1.101 | $0.826 / $3.301 | $0.165 / $0.99 | $0.66 / $3.961 |
| China (Hong Kong) Global |
$1.65 / $4.951 | $0.276 / $1.101 | $0.826 / $3.301 | $0.165 / $0.99 | $0.66 / $3.961 |
| Germany (Frankfurt) Global |
$1.65 / $4.951 | $0.276 / $1.101 | $0.826 / $3.301 | $0.165 / $0.99 | $0.66 / $3.961 |
| US (Virginia) Global aliases |
$1.65 / $4.951 | $0.276 / $1.101 | $0.826 / $3.301 | $0.165 / $0.99 | $0.66 / $3.961 |
| Japan (Tokyo) Global aliases |
$1.65 / $4.951 | $0.276 / $1.101 | $0.826 / $3.301 | $0.165 / $0.99 | $0.66 / $3.961 |
Important: these are list prices, not promotional effective prices. The official documentation displayed time-limited discounts for some floating aliases on the verification date. Because those offers can change independently of the underlying list price, they are intentionally excluded from this table.
US- and Japan-restricted scopes
Some regions also offer locality-restricted model IDs or scopes with different list prices:
| Model and scope | Input range | Input / output per 1M tokens |
|---|---|---|
qwen3.7-max-usUS scope |
Up to 1M | $2.50 / $7.50 |
qwen3.7-plus-usUS scope |
Up to 256K | $0.40 / $1.60 |
qwen3.7-plus-usUS scope |
Over 256K–1M | $1.20 / $4.80 |
qwen3.7-plusJapan scope |
Up to 256K | $0.40 / $1.60 |
qwen3.7-plusJapan scope |
Over 256K–1M | $1.20 / $4.80 |
A region can contain more than one deployment scope, and not every model is available in every scope. Select the required data-residency scope first, then verify the exact model ID and price row. Do not select a scope solely because another row appears less expensive.
How context tiers and token billing work
A model’s context window and its price bands are related but not identical. QwenCloud Qwen3.7 Plus has a 1M-token context and changes price above 256K input tokens; QwenCloud Qwen3.7 Flash changes bands above 32K and 256K. In the Alibaba Model Studio table below, the documented Plus and Flash starting points change at 256K.
The official tables select a band from the total input tokens in one request. Once a request enters a band, every input token in that request uses that band’s input rate, and its output tokens use the corresponding output rate. Always verify this rule for the exact provider and model because plan-based products can use different accounting.
- Input tokens can include system instructions, user messages, resent conversation history, tool definitions, tool results, and tokenized visual input where supported.
- Output tokens include generated answer tokens and, when Thinking is enabled, billable reasoning tokens.
- Cached input tokens can receive a lower rate only after an eligible cache hit under the applicable route’s rules.
- K and M mean 1,000 and 1,000,000 tokens in the official tables.
Pricing-tier warning: price bands are not progressive brackets. For example, a 260K-input Qwen3.7 Plus request uses the higher listed input rate for all input tokens in that request, not only the tokens above 256K.
For a normalized comparison across AWS, Azure, Google Cloud, and Alibaba routes, use the AI cloud pricing comparison; keep the exact model, region, deployment mode, context tier, and workload attached to every number.
Worked Qwen API cost examples
These examples use Qwen3.7 Plus standard list pricing in Singapore’s International scope. They exclude cache, Batch, tools, promotions, and taxes.
Example 1: request below 256K input tokens
A request contains 100,000 input tokens and generates 5,000 output tokens:
Input: 100,000 / 1,000,000 × $0.40 = $0.040
Output: 5,000 / 1,000,000 × $1.60 = $0.008
Total: $0.048Example 2: request above 256K input tokens
A request contains 400,000 input tokens and generates 10,000 output tokens. The whole request uses the second tier:
Input: 400,000 / 1,000,000 × $1.20 = $0.480
Output: 10,000 / 1,000,000 × $4.80 = $0.048
Total: $0.528Monthly estimate
For 50,000 requests per month averaging 2,000 input tokens and 500 output tokens, with every request in the lower Qwen3.7 Plus tier:
Per request:
(2,000 / 1,000,000 × $0.40) + (500 / 1,000,000 × $1.60)
= $0.0016
Monthly estimate:
50,000 × $0.0016 = $80.00Real bills can differ because prompt lengths, output lengths, cache hits, retries, tool use, model routes, and tier distribution vary. Calculate from measured production usage rather than one average request whenever possible.
Does Qwen Thinking cost more?
Thinking tokens are billed as output tokens. Qwen3.7 Max uses the same listed unit rate for non-thinking and thinking modes in the verified regional tables. Qwen3.7 Plus displays separate non-thinking and thinking output columns, but both columns have the same unit price for the scopes listed above.
Equal unit rates do not mean equal request cost. A thinking request can generate additional reasoning tokens before the final answer, increasing the total billable output-token count. Use non-thinking mode when a task does not benefit from multi-step reasoning, and monitor the returned usage object.
Do not generalize this rule to every legacy model. For example, the Singapore qwen-plus row lists different output rates for non-thinking and thinking modes.
Cached-input pricing
Context caching can lower the cost of repeated prompt prefixes for supported model-region combinations. Cache discounts affect input tokens only; generated output remains billed at the standard output rate.
| Cache event | Billing factor | Important condition |
|---|---|---|
| Regular uncached input | 100% of standard input price | Normal input billing |
| Explicit cache creation | 125% of standard input price | Minimum 1,024 cacheable tokens; five-minute validity resets after a hit |
| Explicit cache hit | 10% of standard input price | The repeated prefix must match the eligible cache block |
| Implicit cache hit | 20% of standard input price | Automatic matching; a hit is not guaranteed |
- Explicit and implicit cache modes are mutually exclusive for a request.
- Cache support varies by exact model, region, and deployment scope.
- The discount applies only to the tokens identified as cache hits in usage data.
- Cache and Batch discounts cannot be applied to the same invocation.
Cache creation is deliberately more expensive than ordinary input. It saves money only when the created block receives enough valid hits before expiration. Estimate creation and hit volume together instead of presenting “90% cheaper input” without the setup cost and eligibility conditions.
Batch inference pricing
For models and regions supported by the Model Studio Batch File API, successful input and output tokens cost 50% of the corresponding real-time inference rates. Batch jobs are asynchronous and are intended for workloads such as evaluation, classification, data processing, and other tasks that do not require an immediate response.
- Batch eligibility is model- and region-specific; it is not an automatic “bulk usage” discount.
- The new-user free quota does not cover Batch invocations.
- Batch and context-cache discounts cannot be combined.
- Successfully completed requests remain billable if a Batch task is later cancelled.
- Some Batch models have a lower maximum context than their real-time endpoint.
Always check the Batch supported-model list before budgeting. The pricing matrix and Batch guide can be updated on different dates, so this page does not promise Batch support for a model merely because another model in the same family supports it.
Legacy Qwen prices, aliases, and retirement status
A legacy alias can still have a price without being the right starting point for new development. A floating alias and a dated snapshot also have separate lifecycle questions.
| Legacy reference | Verified status | Migration or identity guide |
|---|---|---|
qwen-flash |
The checked Model Studio matrix maps the current legacy alias to qwen-flash-2025-07-28 |
Qwen Flash versions and migration |
qwen-turbo and qwen-turbo-realtime |
Alibaba Cloud’s Bailian / China Model Studio lifecycle table schedules both hosted IDs for decommissioning on October 10, 2026 and recommends the latest Qwen 3.7 and 3.6 models rather than one exact successor. QwenCloud international still lists qwen-turbo as Legacy, and its own deprecation table does not list Turbo; no current international listing was verified for qwen-turbo-realtime. |
Qwen Turbo status and migration |
qwen-max |
Separate non-thinking floating legacy alias; do not label it as the Qwen2.5-Max snapshot | Qwen2.5-Max historical identity |
| Qwen2.5-VL hosted/open variants | Historical visual family; hosted API aliases and downloadable checkpoints have different pricing and lifecycle | Qwen2.5-VL API, checkpoints, and license |
The following Singapore / International Model Studio list prices are dated maintenance references for exact IDs that remain callable on the documented route; a published price does not guarantee current availability or extend a lifecycle deadline. The qwen-turbo row is useful only for pre-decommissioning budgeting on the affected Bailian / China Model Studio route. Do not project that hosted ID’s inference charges beyond October 10, 2026, because the Bailian notice says calls fail after decommissioning. QwenCloud international status and pricing must be verified separately.
| Legacy model | Input tier | Input | Non-thinking output | Thinking output |
|---|---|---|---|---|
qwen-plus |
Up to 256K | $0.40 | $1.20 | $4.00 |
qwen-plus |
Over 256K–1M | $1.20 | $3.60 | $12.00 |
qwen-flash |
Up to 256K | $0.05 | $0.40 | Same published output column |
qwen-flash |
Over 256K–1M | $0.25 | $2.00 | Same published output column |
qwen-turbo |
No tier | $0.05 | $0.20 | $0.50 |
qwen-max |
No tier | $1.60 | $6.40 | Not supported in this legacy row |
Dated retirement notices to plan for
- The Alibaba Cloud Bailian / China Model Studio lifecycle table schedules
qwen-turboandqwen-turbo-realtimefor decommissioning on October 10, 2026. Its guidance names the latest Qwen 3.7 and 3.6 models, not one universal drop-in alias. - For the affected Bailian / China Model Studio route, the policy says the APIs remain functional before October 10, 2026, although QPM and TPM may be reduced during the notice period. From the official decommissioning time, inference calls to those IDs fail.
- QwenCloud international is separate: on August 4, 2026 its
qwen-turboMarketplace card and text-model matrix still listed the model as Legacy, while its deprecation table did not list Turbo. No current international listing was verified forqwen-turbo-realtime. - The Bailian / China Model Studio lifecycle table schedules
qwen3.6-max-preview,qwen3-max-preview, andqwen3-maxfor decommissioning on October 10, 2026 and directs affected users toqwen3.7-max. - The same Bailian / China Model Studio table gives October 10, 2026 for
qwen3-max-2026-01-23andqwen3-max-2025-09-23, withqwen3.7-maxas the replacement on that route. - The Bailian / China Model Studio lifecycle table schedules the hosted
qwen3-14bID for decommissioning on October 10, 2026. This hosted-service event does not remove separately published checkpoint files already obtained under their applicable license. - The Bailian / China Model Studio lifecycle table records
qwen-turbo-2024-09-19andqwen-turbo-2024-06-24as retired on January 30, 2026 and namesqwen-flash-2025-07-28as the historical replacement on that path.
“Priced,” “Legacy,” “scheduled for decommissioning,” and “offline” are different states. Confirm the provider, deployment scope, exact ID, current lifecycle table, and account console before using a row in a budget.
Qwen vision pricing
Alibaba publishes both input and output rates for its legacy Qwen-VL-Max row. For new visual-understanding workloads, start with the current qwen3.8-max route, then compare qwen3.7-plus or qwen3.7-flash when cost matters. QwenCloud also documents qwen3.8-27b for text, image, and video input with text output. Use the Qwen2.5-VL reference only when you specifically need that historical family or its separate open checkpoints. Qwen-Image-3.0 is an image-generation family with separate billing and should not be mixed with visual-understanding token prices.
| Model | Region and scope | Input per 1M tokens | Output per 1M tokens | Tier |
|---|---|---|---|---|
qwen-vl-max |
Singapore / International | $0.80 | $3.20 | No tiered pricing |
qwen-vl-max |
China (Beijing) / Chinese mainland | $0.23 | $0.574 | No tiered pricing |
Qwen-VL-Max is listed under legacy visual models rather than the recommended choices for new projects. The current official visual-model guide starts with Qwen3.8 Max and then recommends Qwen3.7 Plus or Qwen3.7 Flash as lower-cost alternatives. Verify exact model availability and pricing in the selected region.
Image and video requests are converted into billable input tokens according to the selected model and request. Inspect the API’s usage data; do not estimate multimodal cost from file size alone. Visual understanding also returns text—it is not image-generation billing.
Qwen free quota and billing-profile rules
The Model Studio free quota is not a universal six-month trial. Alibaba documents these rules:
- Only eligible models in the Singapore region with the International deployment scope receive the new-user quota.
- The general free-quota guide states a validity period of 30–90 days. The recommended-model pricing rows checked for this page show 1 million tokens per model, valid for 90 days after Model Studio activation.
- The quota covers eligible real-time inference only. It does not cover Batch, fine-tuning, model deployment, or custom models.
- An Alibaba Cloud root account and its RAM users share the same model quota.
- Quota amount, remaining tokens, and expiration should be checked in the console for the exact account and model.
What happens when the quota ends?
For a new user whose billing profile is incomplete, Model Studio stops eligible calls after the quota is exhausted and can return AllocationQuota.FreeTierOnly. To continue with pay-as-you-go billing, the user must complete the required account information.
For an account with completed billing information, usage after quota exhaustion can roll into pay-as-you-go charges unless the Free Quota Only control is enabled. That control stops further eligible calls when the free allocation is exhausted.
Use Alibaba Cloud’s official Model Studio activation guide instead of an unofficial registration link.
Payment-method accuracy: this page does not claim that every user in every country can activate or continue using Model Studio without a payment card. Account verification and available payment methods depend on the account and contracting entity. Follow the requirements displayed in your Alibaba Cloud account.
Costs not included in the basic token table
- Built-in tools: web search or another provider-run tool can have a separate invocation charge, and retrieved content can add input tokens.
- Fine-tuning and deployment: training, custom-model hosting, and dedicated resources use separate billing rules.
- Provisioned Throughput Units: reserved-capacity billing is different from ordinary pay-as-you-go inference.
- Coding Plan and Token Plan: these plans use their own supported-model lists, credentials, limits, and subscription or token rules.
- Application infrastructure: your database, vector store, observability, network, storage, and gateway can add costs outside Model Studio.
How to reduce Qwen API cost safely
- Choose by measured task quality. Test Qwen3.6 Flash, Qwen3.7 Plus, and Qwen3.7 Max on a representative evaluation set instead of selecting only by unit price.
- Control the 256K boundary. Remove duplicated history and irrelevant retrieved passages before a Plus or Flash request crosses into a higher tier.
- Cap unnecessary output. Give a clear format and length requirement, and use non-thinking mode for simple tasks.
- Cache stable prefixes. Use eligible context caching for repeated system instructions, long documents, or stable tool definitions after measuring cache-hit frequency.
- Use Batch only for eligible asynchronous work. Verify the exact model and region instead of assuming every bulk request receives 50% off.
- Record usage by request. Store input, output, reasoning, and cache-token counts with the model ID, tier, scope, and application feature.
- Set safeguards. Review the Free Quota Only control, billing alerts, application quotas, API-key access, and retry limits.
For request examples, authentication, regional base URLs, streaming, errors, and retries, read the Qwen API guide.
Frequently asked questions
What does Qwen3.8 Max cost?
The QwenCloud Marketplace list price checked August 23, 2026 was $2 per 1M input tokens and $6 per 1M output tokens for production qwen3.8-max. It also listed $0.25 implicit-cache input, $2.50 explicit-cache creation, and $0.17 explicit-cache read per 1M tokens. Confirm the live page and account console. The retired qwen3.8-max-preview ID is temporarily routed to qwen3.8-max; Token Plan calls still use Credits, while PAYG calls use the applicable production pricing route.
What does Qwen3.8-27B cost on QwenCloud?
The live QwenCloud Marketplace lists qwen3.8-27b at $0.50 per 1 million input tokens and $3.00 per 1 million output tokens. It also lists $0.10 implicit-cache input, $0.625 explicit-cache creation, and $0.05 explicit-cache read. This hosted 1M-context service accepts text, image and video input with text output and is separate from the downloadable Qwen/Qwen3.8-27B repository.
How much does the Qwen API cost?
There is no single worldwide amount. The cost depends on the provider, region, deployment scope, model ID, input tier, output volume, Thinking usage, cache hits, Batch eligibility, tools, and account terms. For example, Singapore list pricing for Qwen3.7 Plus starts at $0.40 input and $1.60 output per 1 million tokens for requests with up to 256K input tokens.
Is the Qwen API free?
Eligible first-time Model Studio users can receive model-specific free quotas in Singapore’s International scope. The quota is limited by model, tokens, expiration, account, and request type. It is not an unlimited free API and does not cover Batch, fine-tuning, or deployment.
Does the Qwen free quota last 180 days?
No general 180-day rule is supported by the verified documentation. The general guide states 30–90 days, while the eligible recommended-model rows display 90 days after activation. Check the expiration attached to your quota in the console.
Does Qwen require a credit card?
Alibaba’s documentation does not provide a universal card-free promise for every country and account type. Eligible new users with incomplete billing information can use qualifying quota and are stopped when it ends; continuing with pay-as-you-go requires completion of the applicable account and billing information.
Why did a long Qwen request cost more per token?
Some models use tiered prices based on total input tokens in one request. For Qwen3.7 Plus and Qwen3.6 Flash, crossing 256K input tokens moves the whole request to the higher listed tier.
Are cached Qwen tokens cheaper?
For an eligible cache hit, yes. Explicit cache hits are billed at 10% of the standard input rate, while implicit hits are generally billed at 20%. Explicit cache creation costs 125%, output is not discounted, and support depends on model and region.
Is Qwen Batch always 50% cheaper?
Successful eligible Batch input and output are billed at 50% of real-time rates. The discount does not apply to ordinary high-volume real-time calls, and Batch support varies by model and region. Free quota and cache discounts do not apply to Batch.
Is Qwen-VL-Max output pricing published?
Yes. The verified official table lists Qwen-VL-Max output at $3.20 per 1 million tokens in Singapore’s International scope and $0.574 in Beijing’s Chinese-mainland scope. The model is listed as a legacy visual option.
Should a new application use Qwen Turbo?
Do not start a new deployment on these IDs. Alibaba Cloud’s Bailian / China Model Studio lifecycle table schedules qwen-turbo and qwen-turbo-realtime for decommissioning on October 10, 2026 and recommends the latest Qwen 3.7 and 3.6 models rather than one exact successor. QwenCloud international still documented qwen-turbo as Legacy on August 4, 2026 and did not list Turbo in its deprecation table, so verify that route separately. Test the replacement on the exact provider, deployment scope, region, protocol, workload, and account.
Official documentation
- QwenCloud PAYG pricing
- QwenCloud Token Plan overview
- QwenCloud text models and access routes
- QwenCloud model changelog
- QwenCloud model deprecations
- QwenCloud qwen3.8-max Marketplace card
- QwenCloud qwen3.8-2.4t-a95b Marketplace card
- QwenCloud qwen3.8-27b Marketplace card
- QwenCloud qwen-turbo Marketplace card
- QwenCloud qwen3.7-max Marketplace card
- QwenCloud qwen3.7-plus Marketplace card
- QwenCloud qwen3.7-flash Marketplace card
- QwenCloud vision-model matrix and output limits
- Alibaba Cloud official Qwen Turbo decommissioning notice
- Alibaba Cloud exact qwen3.8-max regional model page
- Alibaba Cloud Model Studio model inference pricing
- Model Studio free-quota rules
- Qwen Context Cache billing and supported models
- Model Studio Batch inference
- Regions, deployment scopes, and access domains
- Qwen text-generation model matrix
- Qwen visual-model matrix
- Qwen Thinking billing behavior
- Alibaba Cloud Bailian / China Model Studio lifecycle and decommissioning policy
- Billing queries and cost management
Prices, promotions, model aliases, regional availability, free quotas, and billing rules can change. Verify the exact regional price row and your account console before deployment.