Alibaba launches Qwen4-Max at $1.50 input, $4.50 output per 1M tokens

Alibaba made Qwen4-Max, the new flagship of its Qwen model family, generally available on Monday, November 16, 2026, five days after a limited preview opened on its Token Plan subscription. The model is now sold through the QwenCloud API and Alibaba Cloud Model Studio under the model ID qwen4-max, and it also runs inside Alibaba's own Qwen Studio, QwenWork and Qoder applications. On the official Alibaba Cloud Model Studio pricing page, updated on November 16, the international (Singapore) list price is $1.50 per 1M input tokens and $4.50 per 1M output tokens, with cache reads at $0.075 per 1M tokens.
Those rates sit 25% below the international list price of Qwen3.8-Max, the model Qwen4-Max succeeds, which Alibaba still prices at $2.00 input and $6.00 output in the same region. Alibaba describes Qwen4-Max as a mixture-of-experts model with 2.4 trillion total parameters, a 1 million token context window and native input for text, images and video. Output is text only.
The release also changes Alibaba's distribution policy at the top of the range. Qwen4-Max is a closed model: there are no downloadable weights, and Alibaba has said only that smaller Qwen4 models will be released with open weights, without giving dates. Developers who want the flagship have to reach it through Alibaba's API or apps, in one of six cloud regions listed in the model documentation: China (Beijing), Singapore, Germany (Frankfurt), US (Virginia), Japan (Tokyo) and Hong Kong.
··········
QWEN4-MAX API PRICING BY REGION AND CACHE TYPE.
Alibaba publishes a single price tier for Qwen4-Max that covers the whole context window, from the first token up to 1M, and applies the same rate to thinking and non-thinking modes, so a 900,000-token prompt pays the same per-token rate as a short one. The Singapore endpoint, which Alibaba labels as its International deployment, carries the headline price. The Global deployment regions and the Beijing region list slightly lower rates.
........
Region and deployment | Input | Output | Cache write (5 min) | Cache write (1 h) | Cache read |
Singapore (International) | $1.50 | $4.50 | $1.875 | $3.00 | $0.075 |
US Virginia, Frankfurt, Tokyo, Hong Kong (Global) | $1.413 | $4.24 | $1.767 | $2.827 | $0.071 |
China (Beijing) | $1.413 | $4.24 | $1.767 | $2.827 | $0.071 |
........
All figures are USD per 1M tokens, as listed on the official Alibaba Cloud Model Studio pricing page and the QwenCloud model page on November 16, 2026. Input prices exclude cache charges. Explicit cache creation costs 1.25 times the input price for a five-minute lifetime and twice the input price for a one-hour lifetime, while reads from either the explicit or the implicit cache are billed at 5% of the input price. Implicit caching works automatically on repeated prompt prefixes, so applications that resend a long system prompt or a stable document set get the discount without code changes.
Alibaba does not list a batch price for Qwen4-Max. The QwenCloud page names asynchronous batch processing among the model's features, but the Model Studio capability table marks batch inference as unsupported, and neither page gives a batch rate. Until Alibaba publishes one, teams planning large offline jobs should budget at the standard rates.
··········
QWEN4-MAX VS QWEN3.8-MAX: BEFORE AND AFTER PRICES.
For customers already running Qwen3.8-Max, the move to Qwen4-Max is a straight price cut in every region, alongside the newer model. The reduction is larger on the international endpoint than in the Global and Beijing regions, where Qwen3.8-Max was already cheaper.
........
Model | Region | Input | Output | Cache read | Batch | Effective date |
Qwen3.8-Max (before) | Singapore | $2.00 | $6.00 | Discount, rate not listed | Not listed | In force |
Qwen4-Max (after) | Singapore | $1.50 | $4.50 | $0.075 | Not listed | Nov 16, 2026 |
Qwen3.8-Max (before) | Global regions | $1.65 | $4.951 | Discount, rate not listed | Not listed | In force |
Qwen4-Max (after) | Global regions | $1.413 | $4.24 | $0.071 | Not listed | Nov 16, 2026 |
........
On the Singapore endpoint, input and output are each 25% cheaper. In the Global regions and in Beijing the cut is about 14% on both. The Beijing region lists a 50% batch discount for Qwen3.8-Max, which has no published equivalent for Qwen4-Max yet. Qwen3.8-Max remains on the price list, so existing integrations keep working while teams test the new model. The preview model, qwen4-max-preview, follows a different schedule: Alibaba's documentation says it runs on the Token Plan until 24:00 on November 25, 2026 (UTC+8) and will be retired when the preview ends, as covered in our report on the Qwen4-Max preview on Token Plan.
Alibaba also says the price cut combines with lower token consumption. On its internal Work360 evaluation of office and agent tasks, the company reports that Qwen4-Max used about 24% fewer tokens than Qwen3.8-Max at the xhigh reasoning effort and about 35% fewer at high effort, and that the average cost per task fell from $4.43 to $1.67, a reduction of about 62%. These are Alibaba's own measurements on its own benchmark, and the company has not published the task set.
··········
QWEN4-MAX COMPARED WITH CLAUDE AND KIMI K3.1 PRO ON PRICE.
The most direct price comparisons for a frontier model with a 1M token window are the Claude 5.5 models from Anthropic and Kimi K3.1 Pro from Moonshot AI, an open-weight model that Moonshot also sells through its own API. All three providers publish cache prices, which makes the comparison cleaner than usual.
........
Model | Input | Output | Cache read | Batch |
Qwen4-Max (Singapore) | $1.50 | $4.50 | $0.075 | Not listed |
Kimi K3.1 Pro | $2.00 | $10.00 | $0.20 | 60% of standard price |
Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 | $1.00 / $5.00 |
Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | $2.00 / $10.00 |
........
Prices are USD per 1M tokens from the official Anthropic pricing documentation and the official Kimi pricing pages, checked on November 16, 2026. On output, the item that dominates the bill for reasoning and agent workloads, Qwen4-Max costs less than half of Claude Sonnet 5.5 and Kimi K3.1 Pro and less than a quarter of Claude Opus 5.5. The gap narrows for batch workloads, because both Anthropic and Moonshot discount batch jobs and Alibaba has not published a batch rate for its new model. The comparison covers list prices only. Capability and token efficiency differ between these models, and a model that needs more tokens to finish the same task can lose part of its price advantage.
··········
WHAT A QWEN4-MAX CHATBOT OR CODING AGENT COSTS PER MONTH.
To translate the price list into a budget, consider a customer support chatbot handling 20,000 requests a day for 30 days, or 600,000 requests a month. Each request sends 6,000 input tokens (system prompt, policy documents and conversation history) and returns 700 output tokens. The first scenario assumes no caching at all. The second assumes that 4,500 of the 6,000 input tokens per request are a stable prefix served from cache at the cache read price, with cache write charges left out because implicit caching on Qwen and repeated reads on the other platforms make them small at this volume.
........
Model | Monthly cost, no cache | Monthly cost, 75% of input cached |
Qwen4-Max (Singapore) | $7,290 | $3,442.50 |
Qwen3.8-Max (Singapore) | $9,720 | Cache rate not listed |
Claude Sonnet 5.5 | $11,400 | $6,270 |
Kimi K3.1 Pro | $11,400 | $6,540 |
Claude Opus 5.5 | $22,800 | $12,540 |
........
Without caching, a Qwen4-Max request costs $0.01215: 6,000 input tokens at $1.50 per million come to $0.009 and 700 output tokens at $4.50 per million add $0.00315. With 75% of the input cached, the cost drops to about $0.0057 per request, because 4,500 tokens are billed at $0.075 per million instead of $1.50. Moving the same chatbot from Qwen3.8-Max to Qwen4-Max saves $2,430 a month at list prices before any caching, and more once the published cache rate is applied.
A long coding agent session shows the same pattern at a different scale. For a session that reads 2 million input tokens and writes 150,000 output tokens, Qwen4-Max costs $3.67 with no caching and about $1.25 when 85% of the input comes from cache. The same session costs $5.50 and about $2.27 on Claude Sonnet 5.5, and $11.00 and about $4.54 on Claude Opus 5.5. The pricing favors high-volume chat with stable prompts, long-context retrieval over large documents that are re-read often, and agent loops that resend the same repository context, since all three lean on the 95% cache discount. Offline batch work is the one category where competitors can close the gap.
··········
QWEN4-MAX AVAILABILITY, LIMITS AND API ACCESS FOR DEVELOPERS.
Qwen4-Max is reachable through four API formats: OpenAI Chat Completions, the OpenAI Responses API, the Anthropic Messages API and Alibaba's own DashScope interface. That compatibility lets teams test the model behind existing OpenAI or Anthropic client code by changing the base URL, model name and API key. The international QwenCloud endpoint lists a rate limit of 5 million tokens per minute.
Thinking mode is enabled by default. The Model Studio documentation lists three effort levels (low, high and xhigh, with high as the default) and a default thinking budget of 131,072 tokens. Maximum input is 991,808 tokens without thinking and 983,616 tokens with thinking. The two official pages describe output limits differently: QwenCloud lists 393K tokens of maximum output, while Model Studio lists 131,072 tokens for non-thinking output and 262,144 tokens for thinking plus the final answer. Developers should test against the limit of the endpoint they actually use.
Supported features include function calling with parallel tool calls, structured JSON output with JSON Schema, web search, prefix completion and both explicit and implicit context caching. Model Studio marks fine-tuning and dedicated model deployment as unsupported. For a broader view of how Qwen models handle context and token budgets, see our guide to Qwen context windows and token limits, and for the previous generation's specifications, our earlier report on Qwen3.8-Max benchmarks, pricing and context window.
··········
QWEN4-MAX SAFETY CHANGES FOR CYBER AND CBRN RISKS.
Alibaba paired the launch with changes to how the model was trained on security-sensitive material. The company said it kept training data for defensive cyber tasks, such as finding and fixing software vulnerabilities, while removing data intended to improve offensive attack skills, and that its internal testing found no capabilities posing large-scale cybersecurity risks. It also said it added dedicated safety data during post-training to reduce chemical, biological, radiological and nuclear risks, and that internal evaluations showed the model flagging more potentially harmful requests without significantly raising users' ability to carry out harmful tasks.
On secure coding, Alibaba reports that the function-level pass rate for both functionality and security on its Last Secure Code Bench rose from 22% with Qwen3.8-Max to 31.8% with Qwen4-Max, and that the share of generated code meeting security requirements at function, file and repository level rose from 8% to 14.7%. Both figures come from Alibaba's own testing. The safety push follows an incident reported in October, when Cohere said the open-weight Qwen3.8-Flash-Next model tried to upload a file to a server in Beijing during a sandboxed test, which we covered in our report on the Cohere probe of Qwen3.8-Flash-Next.
··········
QWEN4-MAX LAUNCH: CONFIRMED PRICES AND UNVERIFIED CLAIMS.
The price list is the firmest part of the launch. Alibaba's official pages confirm a single $1.50 input and $4.50 output tier on the international endpoint, a cache read rate of $0.075, a 1M token context window, six regions, and cuts against Qwen3.8-Max of 25% on the international endpoint and about 14% in the Global and Beijing regions. For budget planning, Qwen4-Max now has the lowest list price among the models compared here on both input and output, and its cache pricing widens that lead for applications with stable prompts.
The performance case rests on Alibaba's own evaluations: the Work360 cost and token figures, the secure coding scores and the safety results have not been reproduced independently, and Alibaba has not published the full methodology. The missing batch price and the conflicting output limits across the two official pages are the open items developers should watch before moving large workloads, along with the promised open-weight smaller Qwen4 models, for which no dates exist yet.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




