# Qwen API Pricing 2026: Every Qwen Model per Million Tokens on Alibaba Cloud, by Region, Free Quota, Cache, Batch and Cost per Conversation

> Qwen API pricing per million tokens as of 29 October 2026: Qwen3.8-Max at $2 input and $6 output on Alibaba Cloud Model Studio's Singapore region and $1.65 and $4.95 on Global-scope rows, Qwen3.7-Plus at $0.40 and $1.60 (20% off for now), Qwen3.8-Flash at $0.15 and $0.47, Qwen3.7-Flash from $0.03 and $0.13, 1 million free tokens per model, context tiers, cache and Batch rules, CNY prices, where open-weight Qwen runs cheaper, and what a 40-turn conversation costs next to DeepSeek, Groq, Gemini and Claude.

_Gaurav Dadhich · 2026-10-29_

\# Qwen API Pricing 2026: Every Qwen Model per Million Tokens on Alibaba Cloud, by Region, Free Quota, Cache, Batch and Cost per Conversation \*Published 29 October 2026. Every Alibaba Cloud price below was read from the \[Model Studio model pricing page\](https://www.alibabacloud.com/help/en/model-studio/model-pricing) and its \[Chinese counterpart\](https://help.aliyun.com/zh/model-studio/model-pricing) on 29 October 2026, and every price is as of that date.\* Qwen3.8-Max, Alibaba's current flagship, costs $2 per million input tokens and $6 per million output tokens on \[Alibaba Cloud Model Studio\](https://www.alibabacloud.com/help/en/model-studio/model-pricing) in the Singapore region, as of 29 October 2026, with one flat rate across its 1-million-token input range and the same price whether thinking is on or off. Qwen3.7-Plus lists at $0.40 input and $1.60 output for requests up to 256,000 tokens and currently sells at a limited-time 20% discount ($0.32 and $1.28), Qwen3.8-Flash costs $0.15 and $0.47, and Qwen3.7-Flash is the cheapest current text model at $0.03 and $0.13 for requests up to 32,000 tokens. Every eligible model comes with its own free quota of 1 million tokens for 90 days, Batch calls cost 50% of the real-time price, an explicit cache hit costs 10% of the input price after a write at 125%, and an implicit cache hit costs 20%. The same model does not have one price. Alibaba prices each model per region and per "service deployment scope", and on 29 October Qwen3.8-Max cost $1.65 input and $4.951 output on the Global-scope rows offered in Beijing, Hong Kong, Frankfurt, Virginia and Tokyo, which is 17.5% below Singapore's International rate on both lines. On Alibaba's China site the same model bills in renminbi at ¥12 and ¥36 in Beijing, and the free quota there applies only to Beijing, where the international site gives it only to Singapore. Several of the guides that rank for "qwen pricing" still lead with Qwen3-Max and Qwen-Turbo, and neither model appears anywhere on Alibaba's pricing page today. A 40-turn conversation, worked through further down, costs $1.04 on Qwen3.8-Max and $0.252 with explicit caching, $0.213 on Qwen3.7-Plus at list price ($0.051 cached) and $0.016 on Qwen3.7-Flash, which makes Qwen3.7-Flash the cheapest model in any of our pricing guides for a conversation of that length. ## Qwen price per million tokens, every current text model on Model Studio | Model (API ID) | Input tokens per request | Input | Output, non-thinking | Output, thinking | Context cache | Batch in Singapore | |---|---|---|---|---|---|---| | Qwen3.8-Max (\`qwen3.8-max\`) | up to 1M | $2.00 | $6.00 | $6.00 | Explicit and implicit | Image and video input only | | Qwen3.7-Max (\`qwen3.7-max\`) | up to 1M | $2.50 | $7.50 | $7.50 | Explicit and implicit | No | | Qwen-Max (\`qwen-max\`), non-thinking only | flat | $1.60 | $6.40 | not offered | Implicit | Yes | | Qwen3.7-Plus (\`qwen3.7-plus\`), 20% off for now | up to 256K | $0.40 | $1.60 | $1.60 | Explicit and implicit | Yes | | | 256K to 1M | $1.20 | $4.80 | $4.80 | | | | Qwen3.6-Plus (\`qwen3.6-plus\`) | up to 256K | $0.50 | $3.00 | $3.00 | Explicit | No | | | 256K to 1M | $2.00 | $6.00 | $6.00 | | | | Qwen3.5-Plus (\`qwen3.5-plus\`) | up to 256K | $0.40 | $2.40 | $2.40 | Explicit | No | | | 256K to 1M | $0.50 | $3.00 | $3.00 | | | | Qwen-Plus (\`qwen-plus\`) | up to 256K | $0.40 | $1.20 | $4.00 | Explicit and implicit | Yes | | | 256K to 1M | $1.20 | $3.60 | $12.00 | | | | Qwen3.8-Flash (\`qwen3.8-flash\`) | up to 1M | $0.15 | $0.47 | $0.47 | Explicit and implicit | No | | Qwen3.7-Flash (\`qwen3.7-flash\`) | up to 32K | $0.03 | $0.13 | $0.13 | Explicit and implicit | No | | | 32K to 256K | $0.10 | $0.40 | $0.40 | | | | | 256K to 1M | $0.20 | $0.80 | $0.80 | | | | Qwen3.6-Flash (\`qwen3.6-flash\`) | up to 256K | $0.25 | $1.50 | $1.50 | Explicit | No | | | 256K to 1M | $1.00 | $4.00 | $4.00 | | | | Qwen3.5-Flash (\`qwen3.5-flash\`) | up to 1M | $0.10 | $0.40 | $0.40 | Explicit | No | | Qwen-Flash (\`qwen-flash\`) | up to 256K | $0.05 | $0.40 | $0.40 | Explicit and implicit | Yes | | | 256K to 1M | $0.25 | $2.00 | $2.00 | | | | Qwen3-Coder-Flash (\`qwen3-coder-flash\`) | up to 32K | $0.30 | $1.50 | not listed | Explicit and implicit | No | | | 32K to 128K | $0.50 | $2.50 | | | | | | 128K to 256K | $0.80 | $4.00 | | | | | | 256K to 1M | $1.60 | $9.60 | | | | | Qwen3-VL-Plus (\`qwen3-vl-plus\`), vision | up to 32K | $0.20 | $1.60 | $1.60 | Explicit and implicit | No | | | 32K to 128K | $0.30 | $2.40 | $2.40 | | | | | 128K to 256K | $0.60 | $4.80 | $4.80 | | | US dollars per million tokens for the Singapore region (International deployment scope), from Alibaba Cloud's \[model pricing page\](https://www.alibabacloud.com/help/en/model-studio/model-pricing) as of 29 October 2026. Cache support is from the \[context cache page\](https://www.alibabacloud.com/help/en/model-studio/context-cache) and Batch support from the \[Batch page\](https://www.alibabacloud.com/help/en/model-studio/batch-interfaces-compatible-with-openai). Each unsuffixed name is an alias for a dated snapshot (for example, \`qwen3.7-plus\` currently points at \`qwen3.7-plus-2026-05-26\`), and the snapshots carry the same prices. Three rows deserve a second look. Qwen3.8-Max costs less than Qwen3.7-Max on both lines, so there is no price reason to stay on the older flagship. Qwen-Plus is the only model in the table where turning thinking on more than triples the output price ($1.20 to $4.00); on the 3.5 to 3.8 generations, thinking and non-thinking output cost the same. And Alibaba's own \[Qwen-Coder page\](https://www.alibabacloud.com/help/en/model-studio/qwen-coder) now recommends "the latest general-purpose model instead of the Qwen-Coder model", which at these prices means Qwen3.8-Flash costs half of Qwen3-Coder-Flash on input and less than a third on output for short requests. ### Open-weight Qwen models on Model Studio | Model (API ID) | Input tokens per request | Input | Output | Context cache | |---|---|---|---|---| | Qwen3.8 2.4T-A95B (\`qwen3.8-2.4t-a95b\`) | up to 1M | $2.00 | $6.00 | Explicit and implicit | | Qwen3.8 27B (\`qwen3.8-27b\`) | up to 1M | $0.50 | $3.00 | Explicit and implicit | | Qwen3.6 35B-A3B (\`qwen3.6-35b-a3b\`) | up to 256K | $0.375 | $2.25 | not listed | | Qwen3.6 27B (\`qwen3.6-27b\`) | up to 256K | $0.60 | $3.60 | not listed | | Qwen3.5 397B-A17B (\`qwen3.5-397b-a17b\`) | up to 256K | $0.60 | $3.60 | not listed | | Qwen3.5 122B-A10B (\`qwen3.5-122b-a10b\`) | up to 256K | $0.40 | $3.20 | not listed | | Qwen3.5 27B (\`qwen3.5-27b\`) | up to 256K | $0.30 | $2.40 | not listed | | Qwen3.5 35B-A3B (\`qwen3.5-35b-a3b\`) | up to 256K | $0.25 | $2.00 | not listed | Singapore, International scope, as of 29 October 2026; thinking and non-thinking output are priced the same for all of these. \[OpenRouter's model listing\](https://openrouter.ai/qwen/qwen3.8-2.4t-a95b) describes Qwen3.8 2.4T-A95B as the open-weight variant of Qwen3.8-Max, and on Model Studio it costs exactly what Qwen3.8-Max costs. These are the models a third-party host can also serve, which is where the price comparison further down gets interesting. ## Why the same Qwen model costs different amounts by region Model Studio runs in six regions, and four of them (Hong Kong, Frankfurt, Virginia and Tokyo) let you pick a "service deployment scope" per workspace, according to Alibaba's \[regions page\](https://www.alibabacloud.com/help/en/model-studio/regions). The region decides where your request data is stored; the scope decides where inference runs. A Global scope schedules inference across Alibaba's worldwide pool, while a US, EU, Japan or Hong Kong scope keeps it inside that geography, and Singapore's International scope is the one the free quota applies to. Each region has its own endpoint, API key and model list, and keys do not work across regions. | Model | Singapore, International | Global scope (Beijing, HK, Frankfurt, Virginia, Tokyo) | US scope (Virginia) | EU scope (Frankfurt) | China site, Beijing (CNY) | |---|---|---|---|---|---| | Qwen3.8-Max | $2.00 / $6.00 | $1.65 / $4.951 | $2.00 / $6.00 | not offered | ¥12 / ¥36 | | Qwen3.7-Plus, up to 256K | $0.40 / $1.60 | $0.276 / $1.101 | $0.40 / $1.60 | not offered | ¥2 / ¥8 | | Qwen3.8-Flash | $0.15 / $0.47 | $0.113 / $0.382 | $0.15 / $0.47 | $0.15 / $0.47 | ¥0.8 / ¥2.7 | | Qwen3.7-Flash, up to 32K | $0.03 / $0.13 | $0.028 / $0.110 | $0.03 / $0.13 | not offered | ¥0.2 / ¥0.8 | | Qwen-Plus, non-thinking | $0.40 / $1.20 (up to 256K) | $0.115 / $0.287 (up to 128K) | $0.40 / $1.20 | $0.40 / $1.20 | ¥0.8 / ¥2 (up to 128K) | | Qwen3.8 27B | $0.50 / $3.00 | $0.424 / $1.696 (Beijing only) | not offered | not offered | ¥3 / ¥12 | Input / output per million tokens, as of 29 October 2026. Dollar prices from the \[English pricing page\](https://www.alibabacloud.com/help/en/model-studio/model-pricing), renminbi prices from the \[China site\](https://help.aliyun.com/zh/model-studio/model-pricing). Qwen-Plus on the Global scope has a different tier structure (up to 128K, 128K to 256K, 256K to 1M) from the International one. Read the table as two price levels. The Global-scope and Beijing rows sit at one level, the Singapore, US and EU rows at a higher one, and for Qwen-Plus the gap is wide: $0.115 input on Global against $0.40 in Singapore, about 3.5 times. A team that has no data-residency requirement and does not need the Singapore free quota pays less by creating a Global-scope workspace in Virginia or Frankfurt, and Alibaba's regions page says the Global scope also gives "a larger inference resource pool". A team that must keep inference in the US or EU pays the Singapore-level price for that guarantee. The renminbi list is its own price list. On the China site, Qwen3.8-Max costs ¥12 and ¥36 in Beijing and on the Global rows, and ¥14.988 and ¥44.965 in Singapore, so the same two price levels show up in both currencies. Budget in the currency and the site (alibabacloud.com or aliyun.com) you will be billed on. ### Limited-time and night discounts Two discounts were live on 29 October, and both are labelled "limited-time" with no end date given. Qwen3.7-Plus is 20% off its list price in Singapore, on the US and Japan scopes and in Beijing, which is why it shows up at $0.32 and $1.28 on resellers. On the Global-scope rows in Hong Kong, Frankfurt, Virginia and Tokyo, Qwen3.7-Plus is 60% off at night (22:00 to 08:00 UTC+8, by billing time) and 20% off during the day, which puts a night-time request at roughly $0.110 input and $0.440 output per million tokens. That gives Qwen3.7-Plus a time-of-day price much like DeepSeek's peak and off-peak rates, and work that can wait until Beijing night time is the cheapest way to run it in real time. ## How Qwen context-tier pricing works Many Qwen models charge by tier, and the tier is set by the total input tokens in a single request. Alibaba's rule is that "all tokens in the request are billed at the unit price of the corresponding tier", and that applies to the output as well, because each tier carries its own output price. A 30,000-token request to Qwen3.7-Flash bills input at $0.03 and output at $0.13; a 33,000-token request bills every input token at $0.10 and every output token at $0.40, so crossing the line by 1,000 tokens more than triples the cost of the whole call. The tier is a property of each request, not of your monthly usage, and that matters for conversations. A chat that resends its history grows by a fixed amount every turn, so on a tiered model it crosses into the next tier at a predictable turn and stays there. In the conversation used further down, that happens at turn 61 on Qwen3.7-Flash, and it is why a 100-turn conversation on that model costs 12.8 times a 40-turn one, where the history alone would predict 5.3 times. Qwen3.8-Max, Qwen3.8-Flash and Qwen3.5-Flash have no tiers at all up to 1 million input tokens, and Qwen3.7-Plus has only one boundary, at 256,000 tokens. ## Qwen API free quota: 1 million tokens per model, 90 days, one region Alibaba gives new Model Studio users a free quota of typically 1,000,000 tokens per model, valid for 90 days from activation, the model's release or the approval of your model request, whichever is later, according to the \[free quota page\](https://www.alibabacloud.com/help/en/model-studio/new-free-quota). On the international site it applies only to models in the Singapore region with the International scope; on the China site, the pricing page says the free quota applies only in Beijing. The details that decide how far it goes: - The quota is per model and is not pooled. \`qwen3.7-max\` and a dated snapshot such as \`qwen3.7-max-2026-05-17\` are separate models with separate quotas, and when one runs out the API does not switch to another. - Input and output come out of one combined total. - It covers real-time inference only, not Batch, fine-tuning, deployments or storage. - An Alibaba Cloud account and all its RAM users share one quota, and re-registering does not grant a new one. - Since 15 September you must complete your account information before activating Model Studio, and the 90-day validity applies to users who activated after 8 September. Billing starts automatically when a model's quota runs out. To stop that, turn on Free Quota Only for each model in the console (it is off by default); the API then returns \`AllocationQuota.FreeTierOnly\` instead of charging you. Alibaba recommends against using it in production, warns that switching it takes time to take effect and that calls made in that window can be charged, and suggests a spending-limit alert as the other safeguard. A separate free allowance for OAuth sign-in still appears in the same FAQ, which describes "an independent free quota for model inference (2,000 calls per day)" that the console's quota page does not display. Several third-party guides report that a developer OAuth free tier ended on 15 April 2026; we could not reconcile the two, so confirm the OAuth allowance in your own console before planning around it. ## How Qwen context caching is billed Model Studio has two cache modes, and they are mutually exclusive, according to the \[context cache page\](https://www.alibabacloud.com/help/en/model-studio/context-cache): | | Explicit cache | Implicit cache | |---|---|---| | How you turn it on | Add \`"cache\_control": {"type": "ephemeral"}\` to a message, up to four markers per request | Automatic on supported models, cannot be switched off | | Writing to the cache | 125% of the input price, and only the new tokens beyond an existing cached prefix | No extra charge (billed as normal input) | | Reading from the cache | 10% of the input price | 20% of the input price | | Lifetime | 5 minutes, reset on every hit | Not fixed; old entries are cleared periodically | | Minimum prefix | 1,024 tokens | 1,024 tokens | | Hit guaranteed? | Yes, within the 5 minutes and a 20-block lookback | No | On Qwen3.7-Plus at list price, that means an explicit write costs $0.50, an explicit hit $0.04 and an implicit hit $0.08 per million tokens. Caching discounts input only, and Alibaba says the Batch and cache discounts "cannot apply simultaneously"; Batch calls get no cache discount at all. The newest models break the ratios. For \`qwen3.8-max\`, \`qwen3.8-max-0902\` and \`qwen3.8-2.4t-a95b\`, cache creation is 125% of input "but the cache hit price is not 10%", and for \`qwen3.8-flash\` neither ratio holds and both "vary by region"; Alibaba points to the console for those prices. The only Qwen3.8 cache price printed on the pricing page is Qwen3.8-Omni-Flash's $0.016 cache-hit rate in Singapore. OpenRouter's endpoint data for Alibaba's own endpoints, read the same day, lists \`qwen3.8-max-0902\` at $0.25 per million for a cache read (12.5% of input) and $2.50 for a write, and \`qwen3.8-flash\` at $0.016 for a read and $0.20 for a write; we use those figures in the worked example and label them. Two practical rules come straight from Alibaba's guide. Keep tool definitions byte-identical between requests, including tool order and field order, because the \`tools\` array is cached as part of the system message. And in agent loops with parallel tool calls, merge the tool results into one tool message, because an explicit cache misses once more than 20 content blocks separate the marker from the cached content. ## Qwen Batch API: 50% off, Singapore and Beijing only The \[Batch File API\](https://www.alibabacloud.com/help/en/model-studio/batch-interfaces-compatible-with-openai) is OpenAI-compatible and charges "50% of the real-time inference price" on input and output for every request that succeeds; failed rows are not billed. A file holds up to 50,000 requests and 500 MB, the completion window runs from 24 hours to 336 hours (14 days), and an account can run two batch tasks in parallel. Batch is available only in Singapore and Beijing, according to the \[regions page\](https://www.alibabacloud.com/help/en/model-studio/regions), and Batch calls draw on neither the free quota nor any cache discount. Model coverage differs sharply by region. In Singapore, text Batch covers \`qwen-max\`, \`qwen3.7-plus\`, \`qwen-plus\` and \`qwen-flash\`, and \`qwen3.8-max\` only for image and video input; text-only Qwen3.8-Max batch is "not supported yet". Beijing covers the whole current line, including Qwen3.8-Max, Qwen3.8-Flash and Qwen3.7-Flash, which on Beijing's prices makes Qwen3.8-Max $0.825 input and $2.48 output in Batch. Every batch request is capped at 256,000 tokens of context for the 3.5 to 3.8 models, and Alibaba warns that those models have thinking switched on by default, so set \`enable\_thinking\` explicitly in each batch line or pay for reasoning you did not want. ## Thinking tokens: billed as output, sometimes twice Thinking content "is billed by Output Token", according to Alibaba's \[deep thinking guide\](https://www.alibabacloud.com/help/en/model-studio/deep-thinking), and on the 3.5 to 3.8 generations it is on by default. Alibaba's own FAQ says that on Qwen3.7-Plus reasoning accounts for "over 60% of the total Output Tokens" and that turning it off cuts total response time by 60% to 75%. \`enable\_thinking: false\` turns it off on hybrid models, and \`thinking\_budget\` caps how many tokens a reply can spend reasoning. A second switch decides whether you pay for reasoning again. With \`preserve\_thinking\` on, the previous turns' \`reasoning\_content\` is appended to the next request, and Alibaba says it is then "counted toward input Token count and billing". As an illustration, add 1,000 reasoning tokens to every reply in the 40-turn conversation below: on Qwen3.8-Max the uncached bill rises from $1.04 to $1.28, and to $2.84 if every earlier turn's reasoning is replayed as input. The 1,000-token figure is our assumption, not a measurement. ## What a 40-turn conversation costs on Qwen, DeepSeek, Mistral, Groq, Kimi, Gemini and Claude Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens; these are the same assumptions as our \[DeepSeek\](https://www.maximem.ai/blog/deepseek-api-pricing-2026), \[Groq\](https://www.maximem.ai/blog/groq-api-pricing-2026) and \[Kimi K3\](https://www.maximem.ai/blog/kimi-k3-api-pricing-2026) pricing guides, so the figures line up. Turn 1 sends 2,150 input tokens and turn 40 sends 21,650, because the API is stateless and every request carries the exchanges before it. The 40-turn conversation consumes 476,000 input tokens to produce 14,000 output tokens; a 100-turn conversation consumes 2,690,000 input tokens and its last request carries 51,650. Reasoning tokens are left out. | Model, price as of the date shown | 40 turns, no caching | 40 turns, cache warm | 100 turns, no caching | 100 turns, cache warm | |---|---|---|---|---| | Qwen3.8-Max, Singapore (29 Oct) | $1.04 | $0.252 | $5.59 | $0.999 | | Qwen3.8-Max, Global scope (29 Oct) | $0.855 | n/a | $4.61 | n/a | | Qwen3.7-Max, Singapore (29 Oct) | $1.30 | $0.286 | $6.99 | $1.08 | | Qwen3.8 27B, Singapore (29 Oct) | $0.280 | $0.078 | $1.45 | $0.269 | | Qwen3.7-Plus, Singapore list (29 Oct) | $0.213 | $0.051 | $1.13 | $0.187 | | Qwen3.7-Plus, Singapore with 20% off (29 Oct) | $0.170 | $0.041 | $0.906 | $0.150 | | Qwen3.7-Plus, Global scope list (29 Oct) | $0.147 | $0.035 | $0.781 | $0.129 | | Qwen3.8-Flash, Singapore (29 Oct) | $0.078 | $0.018 | $0.420 | $0.069 | | Qwen3.8-Flash, Global scope (29 Oct) | $0.059 | n/a | $0.317 | n/a | | Qwen3.7-Flash, Singapore (29 Oct) | $0.016 | $0.004 | $0.206 | $0.032 | | DeepSeek \`deepseek-flash\`, off-peak (27 Oct) | $0.080 | $0.013 | $0.425 | $0.037 | | Mistral Small 4 (28 Oct) | $0.080 | $0.019 | $0.425 | $0.069 | | Groq GPT-OSS 120B (27 Oct) | $0.080 | $0.046 | $0.425 | $0.227 | | Claude Haiku 5.5 (27 Oct) | $0.055 | $0.014 | $0.287 | $0.050 | | Gemini 3.8 Flash, through 31 December 2026 (27 Oct) | $0.410 | $0.111 | $2.15 | $0.376 | | Kimi K3.1 Pro (28 Oct) | $1.09 | $0.274 | $5.73 | $0.981 | | Claude Sonnet 5.5 (27 Oct) | $1.09 | $0.240 | $5.73 | $0.743 | Qwen rows are our arithmetic on Alibaba's rates read on 29 October 2026. The other rows are the figures published in our \[DeepSeek\](https://www.maximem.ai/blog/deepseek-api-pricing-2026), \[Mistral\](https://www.maximem.ai/blog/mistral-api-pricing-2026), \[Groq\](https://www.maximem.ai/blog/groq-api-pricing-2026), \[Gemini\](https://www.maximem.ai/blog/gemini-api-pricing-2026), \[Kimi K3\](https://www.maximem.ai/blog/kimi-k3-api-pricing-2026) and \[Claude\](https://www.maximem.ai/blog/claude-api-pricing-2026) guides on the dates shown, where the method for each is set out. "Cache warm" on Qwen uses the explicit cache: turn 1 writes its 2,150-token prompt at 125% of input, and every later turn reads the previous prompt at the hit rate and writes only the 500 new tokens at 125%, which is how Alibaba's guide describes a continuous multi-turn conversation with a marker on the last message. For Qwen3.8-Max and Qwen3.8-Flash, whose cache prices Alibaba does not publish as ratios, we used the Alibaba-endpoint rates on OpenRouter quoted above. The model assumes every reply comes within five minutes. With the implicit cache instead (hits at 20%, no write premium, no guarantee), Qwen3.7-Plus at list price costs $0.067 for 40 turns and $0.288 for 100. "n/a" marks Global-scope rows where no cache rate is published. Four things come out of the table: - \*\*Qwen3.7-Flash is the cheapest model we have priced for a 40-turn chat\*\*, at $0.016 uncached, a fifth of DeepSeek \`deepseek-flash\` off-peak and Mistral Small 4 ($0.080), and $0.004 with the cache warm. - \*\*Its tier boundary is the catch.\*\* At 100 turns Qwen3.7-Flash costs $0.206, almost half of DeepSeek's $0.425, because from turn 61 every request is over 32,000 tokens and bills all input and output at the second tier. Qwen3.8-Flash has no tier and grows 5.4 times over the same stretch. - \*\*Qwen3.8-Max prices against Western mid-tier models, not the cheap ones.\*\* At $1.04 uncached it sits just under Claude Sonnet 5.5 and Kimi K3.1 Pro ($1.09), and its warm-cache figure ($0.252) lands between them. - \*\*Region moves the bill by about a sixth on the flagship and more than a quarter on Plus.\*\* The Global-scope Qwen3.7-Plus conversation costs $0.147 against $0.213 in Singapore at list. At 10,000 forty-turn conversations a month, Qwen3.8-Max in Singapore costs $10,360 without caching and $2,517 with it, Qwen3.7-Plus at today's discounted Singapore price costs $1,702 and $411, Qwen3.8-Flash costs $780 and $182, and Qwen3.7-Flash costs $161 and $40. Going from 40 to 100 turns multiplies turns by 2.5 and the uncached Qwen3.8-Max bill by 5.4, because every turn carries every turn before it, and on a tiered model the history eventually pushes each request into a pricier tier as well. Caching cuts the price of the repeated tokens; it does not stop them being sent, and a five-minute cache does nothing for a user who comes back tomorrow. The other lever is to stop resending the whole transcript. We build \[Maximem Synap\](https://www.maximem.ai/synap), a memory layer that returns a ranked set of facts about the user, trimmed to a token budget you set, in place of the full history. When that beats caching, and when it does not, is worked through in \[how to reduce LLM token costs in long conversations\](https://www.maximem.ai/blog/reduce-llm-token-costs-long-conversations). ## Where open-weight Qwen models are cheaper than Alibaba Alibaba's pricing page files Qwen3.8 27B, Qwen3.8 2.4T-A95B and the Qwen3.6 and Qwen3.5 models under "Qwen (open source)", and other inference providers serve the same weights at their own prices. For the mid-size models, several of those providers undercut Model Studio's Singapore price by a wide margin; for the 2.4-trillion-parameter model, nobody does. | Model | Host | Input | Cached input | Output | Notes | |---|---|---|---|---|---| | Qwen3.8 27B | Model Studio, Singapore | $0.50 | explicit hit at 10% | $3.00 | 1M input range | | | Model Studio, Beijing | $0.424 | | $1.696 | Chinese mainland | | | \[Groq\](https://console.groq.com/docs/model/qwen/qwen3.8-27b) | $0.80 | not offered | $4.00 | Preview model, ~450+ tokens/s by Groq's figure, 131,072 context, 16,384 max output | | | \[DeepInfra\](https://deepinfra.com/Qwen/Qwen3.8-27B) | $0.15 | $0.038 | $1.875 | Promotional 25% off ($0.20, $0.05 and $2.50 list); 262,144 context | | | \[OpenRouter\](https://openrouter.ai/qwen/qwen3.8-27b), cheapest endpoint | $0.039 | $0.0039 | $1.25 | Chutes, fp8 weights; 21 providers listed | | | OpenRouter, Cerebras endpoint | $0.99 | $0.99 | $1.49 | fp16, 65,536 context | | Qwen3.8 2.4T-A95B | Model Studio, Singapore | $2.00 | | $6.00 | | | | \[Together AI\](https://www.together.ai/pricing) | $2.00 | $0.25 | $6.00 | | | | OpenRouter, all five providers | $2.00 | $0.25 | $6.00 | Alibaba, Together, Novita, SiliconFlow, Venice | | Qwen3.5 397B-A17B | Model Studio, Singapore | $0.60 | | $3.60 | up to 256K | | | Model Studio, Global scope | $0.172 | | $1.032 | up to 128K | | | Together AI | $0.60 | | $3.60 | | | | OpenRouter, DeepInfra endpoint | $0.45 | $0.22 | $3.00 | fp8 | | Qwen3.6 35B-A3B | Model Studio, Singapore | $0.375 | | $2.25 | | | | Model Studio, Global scope | $0.248 | | $1.485 | | | | OpenRouter, DeepInfra, Parasail and CoreWeave endpoints | $0.10 | from $0.025 | $0.90 to $0.95 | fp8 | | Qwen3.6 27B | Model Studio, Singapore | $0.60 | | $3.60 | | | | OpenRouter, Chutes endpoint | $0.25 | $0.025 | $2.00 | fp8 | US dollars per million tokens, read from each provider's page on 29 October 2026; OpenRouter figures from its public models and endpoints API on the same day. For Qwen3.8 27B the spread between hosts is wider than for any other model in the table. The 40-turn conversation costs $0.280 on Model Studio in Singapore, $0.437 on Groq, $0.098 on DeepInfra at its promotional price and $0.036 on the cheapest OpenRouter endpoint. Groq is the only host in the table that costs more than Alibaba, and what it sells for that premium is speed; it also caps output at 16,384 tokens and has no cache discount for this model. The cheapest listings serve quantized weights (fp8 or fp4), which OpenRouter's data shows per endpoint, so compare outputs on your own prompts before moving production traffic, and pin a conversation to one provider if you route through OpenRouter, or the cache rebuilds on every hop. The proprietary models can be bought elsewhere too, and here a reseller sometimes beats the source. Together AI lists Qwen3.7-Max at $1.50 input, $0.30 cached and $4.50 output, below Model Studio's $2.50 and $7.50 in Singapore and its $1.65 and $4.951 on the Global rows, and OpenRouter's Alibaba endpoint for Qwen3.7-Max showed $1.475 and $4.425. Together's Qwen3.7-Plus ($0.32 and $1.28) and Qwen3.8 Flash ($0.15 and $0.47) match Alibaba's current Singapore prices. If the free quota does not matter to you and data residency does not either, check the reseller price before signing up with Alibaba directly. ## Qwen API rate limits Model Studio limits each model separately, counted across the whole Alibaba Cloud account, which includes every RAM user, workspace and API key, according to the \[rate limit page\](https://www.alibabacloud.com/help/en/model-studio/rate-limit). Limits also apply per second (RPM/60 and TPM/60), so a burst can be rejected inside a minute that stays under the per-minute numbers. | Model | Region and scope | Requests per minute | Tokens per minute | |---|---|---|---| | \`qwen3.8-max\`, \`qwen3.8-flash\` | Singapore | Dynamic | Dynamic | | \`qwen3.8-max\` | Virginia, Global and US | 30,000 | 5,000,000 | | \`qwen3.7-max\` | Singapore | 600 | 1,000,000 | | \`qwen3.7-plus\`, \`qwen3.7-flash\`, \`qwen3.6-plus\` | Singapore | 15,000 | 5,000,000 | | \`qwen3.7-plus\` | Virginia, Global | 30,000 | 5,000,000 | | \`qwen-plus\`, \`qwen-max\` | Singapore | 600 | 1,000,000 | | \`qwen-flash\` | Singapore | 600 | 5,000,000 | | Dated snapshots (for example \`qwen3.7-plus-2026-05-26\`) | Singapore | 60 | 1,000,000 | Alibaba listed these limits on 29 October 2026. Exceeding a limit returns HTTP 429 and usually clears within a minute. Two points catch teams out. Topping up your balance "does not change the model's default RPM and TPM", so higher limits mean a request to your account manager rather than a bigger payment. And dated snapshots get far lower limits than the aliases, 60 requests a minute on several of them, which is a reason to pin a snapshot in evaluation and use the alias in production only after you have tested it. Batch calls on \`qwen-max\`, \`qwen-plus\` and \`qwen-flash\` are not subject to these limits. ## Vision, Omni, OCR and translation prices | Model | Input | Output | Notes | |---|---|---|---| | Qwen3.8-Omni-Flash (\`qwen3.8-omni-flash\`) | $0.15 | $0.47 | Cache hit $0.016; Global scope $0.113, $0.014 and $0.382 | | Qwen3.5-Omni-Plus | $1.40 text, image or video; $11 audio | $8.30 text; $44 text plus audio | | | Qwen3.5-Omni-Flash | $0.40 text, image or video; $3 audio | $2.20 text; $11.90 text plus audio | | | Qwen3.8-Omni-Flash-Realtime | $0.23 text, image or video; $0.93 audio | $0.70 text; $1.87 audio | Speech output bills both audio and its text | | Qwen-VL-OCR (\`qwen-vl-ocr\`) | $0.07 | $0.16 | Global scope $0.043 and $0.072 | | Qwen-MT-Plus, Flash, Lite (translation) | $2.46 / $0.16 / $0.12 | $7.37 / $0.49 / $0.36 | Global scope $0.259 / $0.101 / $0.086 input | Singapore prices per million tokens unless stated, as of 29 October 2026, from the \[pricing page\](https://www.alibabacloud.com/help/en/model-studio/model-pricing). Qwen-MT-Plus shows the widest regional gap in the catalogue: $2.46 input in Singapore against $0.259 on the Global scope, about 9.5 times. ## Data location, endpoints and SLA On Model Studio, the region you call stores the request data and the result, and inference runs within the scope you chose, with transmission encrypted between the two; Alibaba's \[regions page\](https://www.alibabacloud.com/help/en/model-studio/regions) says "your static data always remains in the selected region". For a team with an EU or US residency requirement, that means a Frankfurt workspace on the EU scope or a Virginia workspace on the US scope, at the Singapore-level price and with a shorter model list: the EU scope offered only \`qwen-plus\`, \`qwen3.8-flash\`, \`qwen3.5-flash\` and \`qwen3-vl-plus\` among the text and vision models on 29 October. Alibaba recommends the workspace-dedicated endpoint (\`{WorkspaceId}.{region}.maas.aliyuncs.com\`), which carries a 99.9% SLA, a 3,600-second request timeout and the newer transport protocols. The older DashScope domain also has a 99.9% SLA and a 600-second timeout, but "will no longer support new features after September 30, 2026". Batch and fine-tuning run only in Singapore and Beijing, and the advanced monitoring and alerting features are limited to Singapore, Virginia (monitoring only) and Beijing. ## Which older Qwen models are gone, and how retirement works Qwen3-Max, Qwen-Turbo, Qwen3-Coder-Plus and the Qwen3-235B line, which several ranking guides still price, do not appear on Alibaba's pricing page as of 29 October 2026, and its \[release notes\](https://www.alibabacloud.com/help/en/model-studio/model-release-notes) list retirements of legacy mainline models and snapshots in April and June 2026 and a sunset notice for legacy models in July. Some of those models are still served by third parties at their own prices, but none can be bought from Model Studio on a new account. Alibaba's \[retirement policy\](https://www.alibabacloud.com/help/en/model-studio/model-depreciation) gives stable models (the alias, the dated version and the \`-latest\` name) at least three months between the retirement notice and end of service, with new purchases stopping seven days after the notice; preview models get 14 days. New users lose access, and default rate limits may fall, from the day of the notice. ## Which Qwen model to pick on cost For a new chat or agent workload on Model Studio, start with Qwen3.8-Flash. It has no context tiers up to 1 million tokens, it supports both cache modes, and at $0.15 and $0.47 it costs about the same as the cheap Western and Chinese models in the comparison table. Move to Qwen3.7-Flash if most of your requests stay under 32,000 tokens, where it is the cheapest current model Alibaba sells, and watch the tier boundary as conversations grow. Use Qwen3.7-Plus at its discounted price for work that needs a stronger model with thinking, and set \`enable\_thinking: false\` on turns that do not. Qwen3.8-Max is the pick when you need the flagship. It costs less than Qwen3.7-Max on every line, so there is no cost reason to stay on 3.7, and on a Global-scope workspace it costs 17.5% less than in Singapore. If the free quota is what you want, sign up in Singapore, evaluate against the quota on the aliases you will actually use, and create a Global-scope workspace in Virginia or Frankfurt for production if you have no residency constraint. For the open-weight models, the cheapest route is usually not Alibaba. Qwen3.8 27B and the Qwen3.6 and Qwen3.5 mid-size models cost a fraction of Model Studio's Singapore price on DeepInfra and the low-cost OpenRouter endpoints, at the price of quantized weights and a provider you have to evaluate yourself. Qwen3.8 2.4T-A95B costs $2 and $6 wherever you buy it. ## How to estimate your monthly Qwen bill, and cut it A real-time request costs (cache-hit tokens times the hit rate) plus (cache-write tokens times the write rate) plus (other input tokens times the input rate) plus (output and reasoning tokens times the output rate), all at the tier that the request's total input falls into. For a conversation, the input on each turn is the system prompt plus every earlier turn, including any replayed \`reasoning\_content\`, so estimate per conversation as in the table above and multiply by conversations per month. Batch requests cost half and get no cache discount. The levers that move a Qwen bill, roughly in order of size: 1. \*\*Pick the region and scope deliberately.\*\* A Global-scope workspace costs 17.5% less on Qwen3.8-Max and about 31% less on Qwen3.7-Plus than Singapore, if residency allows it. 2. \*\*Turn thinking off where the answer is simple.\*\* On Qwen3.7-Plus reasoning is more than 60% of output tokens by Alibaba's own measure, and \`preserve\_thinking\` bills it again as input. 3. \*\*Use the explicit cache on long, stable prefixes.\*\* A hit costs 10% of input against 20% for the implicit cache, and the write premium repays itself on the second hit. 4. \*\*Keep requests under the tier boundary\*\* on tiered models, or use a model without tiers once prompts get long. 5. \*\*Send work that can wait through Batch\*\* in Singapore or Beijing, or schedule real-time Qwen3.7-Plus calls into the Global-scope night window. 6. \*\*Check the reseller price\*\* for the model you want, and the open-weight hosts for Qwen3.8 27B and the 3.6 and 3.5 models. 7. \*\*Stop resending the whole conversation\*\* once chats run long, by summarising or retrieving only what the next turn needs. Set a spending-limit alert before any of this; Alibaba's usage records update within minutes, and monitoring data is "for reference only" and not a billing basis. ## Frequently asked questions ### How much does the Qwen API cost per million tokens? As of 29 October 2026, on Alibaba Cloud Model Studio's Singapore region, Qwen3.8-Max costs $2 input and $6 output per million tokens, Qwen3.7-Plus $0.40 and $1.60 for requests up to 256,000 tokens (20% off for a limited time), Qwen3.8-Flash $0.15 and $0.47, and Qwen3.7-Flash $0.03 and $0.13 for requests up to 32,000 tokens. Global-scope regions charge less, for example $1.65 and $4.951 for Qwen3.8-Max. ### Is the Qwen API free? New Model Studio users get a free quota of typically 1 million tokens per model, input and output combined, valid for 90 days. On alibabacloud.com it applies only in the Singapore region with the International scope, and on the China site only in Beijing. It does not cover Batch calls, and billing starts automatically when the quota runs out unless you switch on Free Quota Only. ### Why are Qwen prices different in different regions? Alibaba prices each model per region and per service deployment scope. Global-scope rows in Beijing, Hong Kong, Frankfurt, Virginia and Tokyo are cheaper than Singapore's International scope and the US and EU scopes, which keep inference inside that geography; Qwen3.8-Max costs $1.65 and $4.951 on Global against $2 and $6 in Singapore. ### How does Qwen tiered pricing work? On tiered models, the tier is chosen by the total input tokens in one request, and every input and output token in that request is billed at that tier's price. A Qwen3.7-Flash request just over 32,000 tokens costs $0.10 input and $0.40 output per million on all its tokens, against $0.03 and $0.13 just under the line. ### How much does Qwen context caching cost? With the explicit cache, writing costs 125% of the input price and a hit costs 10%, with a 5-minute lifetime that resets on each hit; the automatic implicit cache bills hits at 20% with no write charge. Qwen3.8-Max and Qwen3.8-Flash use different ratios that Alibaba publishes only in the console. ### Does Qwen have a Batch API discount? Yes. Batch costs 50% of the real-time price on input and output, with a 24-hour to 14-day completion window and up to 50,000 requests per file. It runs only in Singapore and Beijing, does not stack with cache discounts and does not use the free quota. ### Are Qwen thinking tokens charged? Yes, thinking tokens bill at the output price, and on the 3.5 to 3.8 models thinking is on by default. If \`preserve\_thinking\` is on, earlier reasoning is also billed again as input on later turns. Among the current text models in the main table, Qwen-Plus is the only one whose thinking output costs more than its non-thinking output ($4.00 against $1.20 in Singapore). ### What is the cheapest Qwen model? Qwen3.7-Flash is the cheapest current text model on Model Studio at $0.03 input and $0.13 output per million tokens in Singapore for requests up to 32,000 tokens, or $0.028 and $0.110 on the Global scope. A 40-turn chat costs about $0.016 on it without caching. ### Where is Qwen3.8 27B cheapest? As of 29 October 2026, the cheapest OpenRouter endpoint for Qwen3.8 27B charged $0.039 input and $1.25 output, and DeepInfra $0.15 and $1.875 at a promotional price, against $0.50 and $3.00 on Model Studio in Singapore and $0.80 and $4.00 on Groq. The cheapest endpoints serve quantized weights. ### What are Qwen API prices in CNY? On Alibaba's China site, Qwen3.8-Max costs ¥12 input and ¥36 output per million tokens in Beijing, Qwen3.7-Plus ¥2 and ¥8 up to 256,000 tokens, Qwen3.8-Flash ¥0.8 and ¥2.7, and Qwen3.7-Flash ¥0.2 and ¥0.8 up to 32,000 tokens. ### Is Qwen cheaper than DeepSeek, Claude or Gemini? It depends on the model. A 40-turn conversation costs $0.016 on Qwen3.7-Flash and $0.078 on Qwen3.8-Flash, against $0.080 on DeepSeek \`deepseek-flash\` off-peak and $0.410 on Gemini 3.8 Flash. Qwen3.8-Max costs $1.04 for the same conversation, close to Claude Sonnet 5.5 at $1.09. ## Sources Alibaba Cloud and host sources retrieved 29 October 2026; comparison figures from our guides as dated in the table. - Alibaba Cloud Model Studio: \[model pricing\](https://www.alibabacloud.com/help/en/model-studio/model-pricing), \[model pricing in CNY (China site)\](https://help.aliyun.com/zh/model-studio/model-pricing), \[context cache\](https://www.alibabacloud.com/help/en/model-studio/context-cache), \[new user free quota\](https://www.alibabacloud.com/help/en/model-studio/new-free-quota), \[Batch File API\](https://www.alibabacloud.com/help/en/model-studio/batch-interfaces-compatible-with-openai), \[rate limits\](https://www.alibabacloud.com/help/en/model-studio/rate-limit), \[regions and endpoints\](https://www.alibabacloud.com/help/en/model-studio/regions), \[deep thinking\](https://www.alibabacloud.com/help/en/model-studio/deep-thinking), \[Qwen-Coder\](https://www.alibabacloud.com/help/en/model-studio/qwen-coder), \[model retirement policy\](https://www.alibabacloud.com/help/en/model-studio/model-depreciation) and \[release notes\](https://www.alibabacloud.com/help/en/model-studio/model-release-notes) - Hosts: \[Groq Qwen 3.8 27B\](https://console.groq.com/docs/model/qwen/qwen3.8-27b), \[DeepInfra Qwen3.8-27B\](https://deepinfra.com/Qwen/Qwen3.8-27B), \[Together AI pricing\](https://www.together.ai/pricing), and OpenRouter's public models and endpoints API for \[Qwen3.8 27B\](https://openrouter.ai/qwen/qwen3.8-27b), \[Qwen3.8 2.4T-A95B\](https://openrouter.ai/qwen/qwen3.8-2.4t-a95b), Qwen3.8-Max-0902, Qwen3.8-Flash, Qwen3.7-Max, Qwen3.7-Plus, Qwen3.7-Flash, Qwen3.6 35B-A3B, Qwen3.6 27B and Qwen3.5 397B-A17B - Maximem: \[DeepSeek API pricing\](https://www.maximem.ai/blog/deepseek-api-pricing-2026), \[Mistral API pricing\](https://www.maximem.ai/blog/mistral-api-pricing-2026), \[Groq API pricing\](https://www.maximem.ai/blog/groq-api-pricing-2026), \[Gemini API pricing\](https://www.maximem.ai/blog/gemini-api-pricing-2026), \[Kimi K3 API pricing\](https://www.maximem.ai/blog/kimi-k3-api-pricing-2026) and \[Claude API pricing\](https://www.maximem.ai/blog/claude-api-pricing-2026) Worked examples are our own arithmetic on those rates.

---

Source: [https://www.maximem.ai/blog/qwen-api-pricing-2026](https://www.maximem.ai/blog/qwen-api-pricing-2026)