Yes, Qwen4-Max is generally available as of November 16, 2026, under the production model ID qwen4-max with published pay-as-you-go API pricing, a 1,000,000-token context window, and live access on QwenCloud, OpenRouter, Vercel AI Gateway, and Novita AI. Singapore International list rates start at $1.50 per million input tokens and $4.50 per million output tokens.
If you saw early developer chatter around a November 11 preview without published unit rates, the picture changed on November 16. From Hangzhou, China, Alibaba formally unveiled Qwen4-Max as the first model in the Qwen4 family and the most capable Qwen release to date, moving the flagship from a Token Plan preview into full commercial API availability.
Is Qwen4-Max generally available or still in preview?
Alibaba made Qwen4-Max generally available for production workloads on November 16, 2026, under the model ID qwen4-max. On QwenCloud’s latest model documentation, qwen4-max is listed as the current production model and can be billed through Token Plan Personal, Token Plan Team (both consuming Credits), or standard pay-as-you-go API calls. Alibaba also rolled out access on QwenWork and QwenStudio, according to the official Alibaba Cloud press release.
Developers still need to watch out for the older preview identifier, qwen4-max-preview, which remains active on a separate schedule. According to the Alibaba Cloud Model Studio preview documentation updated on November 16, 2026:
- Production recommendation: The preview page now explicitly states, “The production version, Qwen4-Max, is available and recommended for use.”
- Preview billing restriction:
qwen4-max-previewis available through Token Plan only and does not support pay-as-you-go billing, which is why no pay-as-you-go unit rate was published when the preview opened on November 11. - Retirement window: The preview window runs from 10:00 on November 11, 2026 to 24:00 on November 25, 2026 (UTC+8), after which
qwen4-max-previewwill be retired. - Different defaults and caps: The preview uses
reasoning_effortvalues oflow,medium, orxhigh(default) and a combined maximum output of 393,216 tokens, both of which differ from the productionqwen4-maxconfiguration.
How much does the Qwen4-Max API cost?
Qwen4-Max pricing depends on the QwenCloud region, with Singapore International list rates set at $1.50 per million input tokens and $4.50 per million output tokens as of November 17, 2026. According to the QwenCloud model pricing page and the Qwen4-Max model card, Alibaba uses a single published tier covering requests from 0 to 1 million input tokens across both non-thinking and thinking modes, with chain-of-thought and final answer tokens billed together at the output rate.
| QwenCloud Pricing Item (as of Nov 17, 2026) | Qwen4-Max (Singapore / Int.) | Qwen4-Max (Beijing / Global) | Qwen3.8-Max (Singapore / Int.) |
|---|---|---|---|
| Input (per 1M tokens) | $1.50 | $1.413 | $2.00 |
| Output (per 1M tokens) | $4.50 | $4.24 | $6.00 |
| Implicit cache read (per 1M tokens) | $0.075 | Not separately listed here | $0.25 |
| Explicit cache write, 5 min (per 1M tokens) | $1.875 | Not separately listed here | $2.50 (cache creation) |
| Explicit cache write, 1 hour (per 1M tokens) | $3.00 | Not separately listed here | N/A |
| Explicit cache read (per 1M tokens) | $0.075 | Not separately listed here | $0.17 |
| Free pay-as-you-go quota | 1,000,000 tokens (Singapore only) | None | 1,000,000 tokens (Singapore only) |
Regional pricing matters if you deploy outside Southeast Asia. While Singapore carries the $1.50 / $4.50 International list rate and includes 1 million free pay-as-you-go tokens (separate from Token Plan credits), QwenCloud’s China North 2 (Beijing) and Global regions (Frankfurt, Hong Kong, Tokyo, and US Virginia) share lower standard list prices of $1.413 per million input tokens and $4.24 per million output tokens, with no free quota. Console discounts may also alter final bills.
For Qwen4-Max vs Qwen3.8-Max on the same International rate card, the generational cut is straightforward. On Qwen3.8-Max’s International model page and pricing documentation, list rates sit at $2.00 input and $6.00 output per million tokens, meaning Qwen4-Max reduces both base input and output list prices by 25%. Implicit cache reads drop even further, from $0.25 to $0.075 per million tokens, a 70% reduction. For a broader look at how token-based billing works across providers, see our AI API pricing comparison.
What are the Qwen4-Max architecture and 1M context window limits?
Qwen4-Max uses a 2.4-trillion-parameter sparse mixture-of-experts (MoE) architecture that activates 71 billion parameters per token and supports a 1,000,000-token context window. According to Alibaba’s announcement, the Qwen4 design updates sparse MoE routing across attention layers, residual connections, and embeddings, activating 25% fewer parameters per token than Qwen3.8-Max at the same 2.4T total parameter scale while incorporating native vision support.
In production on QwenCloud, the live token limits for qwen4-max differ meaningfully from the preview’s 393,216-token combined output cap:
- Context length: 1,000,000 tokens.
- Non-thinking limits: Up to 991,808 input tokens and 131,072 output tokens.
- Thinking mode limits: Up to 983,616 input tokens and 262,144 combined tokens (
max_completion_tokens) across thinking and the final answer. - Default reasoning settings: Thinking is enabled on
qwen4-maxby default withreasoning_effort=highandthinking_budget=131072. Production acceptslow,high, orxhighforreasoning_effort, and developers can passenable_thinking=falsein Chat Completions to disable thinking.
Supported capabilities and unsupported features
QwenCloud lists broad multimodal and agentic support for qwen4-max, though three enterprise operations remain unavailable at launch. Supported features include text generation, thinking, CoT Summary, function calling, parallel tool calls, structured output, image understanding, video understanding, PDF understanding, explicit and implicit caching, prefix completion, streaming, web search, web scraping, text-to-image and image-to-image search, MCP, and a code interpreter (which requires thinking to be enabled).
Compatibility spans OpenAI-compatible Chat Completions and Responses APIs, Anthropic Messages, and DashScope, though text-only DashScope requests must still target the DashScope multimodal endpoint. On the limitation side, QwenCloud explicitly marks batch inference, fine-tuning, and custom model deployment as unsupported for qwen4-max as of November 17, 2026.
How does Qwen4-Max compare on official claims and Artificial Analysis?
Alibaba and independent tracker Artificial Analysis both report gains for Qwen4-Max over Qwen3.8-Max as of November 17, 2026, though their metrics measure different aspects of model behavior. I have not run independent tests on Qwen4-Max at Lotook, so the figures below come strictly from Alibaba’s press materials and Artificial Analysis’s published evaluation logs.
On Alibaba’s proprietary Work360 benchmark covering 721 specialized fields, Alibaba states that Qwen4-Max outperforms Qwen3.8-Max overall while using 35% fewer tokens at High effort reasoning mode and cutting average cost per task by roughly 60%. Alibaba also states in the same press release that Qwen4-Max ranked 6th on Code Arena, and cites internal demonstrations such as 9% lower average model-call costs in data annotation, over 14x faster offline CPU rescheduling, and a 95% reduction in automotive workpiece programming time (from roughly 475 minutes to 24 minutes). Those are vendor demonstrations rather than third-party audits, and Alibaba did not publish numeric scores in the press release for Agents’ Last Exam, Vals Index, JobBench, or Chartography.
Independent numbers from Artificial Analysis’s Qwen4-Max evaluation, logged in the November 15, 2026 Artificial Analysis changelog and compared against the Alibaba creator leaderboard, give a concrete reference point:
- Intelligence Index: Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index, compared with 34 for Qwen3.8-Max (0902), a 9-point increase. That places Qwen4-Max at rank #38 out of 81 models in the displayed class.
- Verbosity and speed: Artificial Analysis records 127 million output tokens across the Intelligence Index suite (described by AA as very verbose relative to the 78.6 million median) and 32.1 minutes per task (compared with a 13.3-minute median), with an output speed around 54 tokens per second on the model summary page.
- Metadata differences: Artificial Analysis displays a 984k context window rather than QwenCloud’s official 1,000,000-token spec, and notes it has not listed a parameter count on its model page, whereas Alibaba’s press release discloses 2.4T total and 71B active parameters.
Where can developers call the Qwen4-Max API?
Developers can call Qwen4-Max directly through QwenCloud or through three verified third-party gateways as of November 17, 2026: OpenRouter, Vercel AI Gateway, and Novita AI. Each platform assigns its own model identifier, and reseller pricing or catalog limits can diverge from QwenCloud’s International list rates.
- QwenCloud (official): Use model ID
qwen4-max. For Singapore Chat Completions, the documented base URL ishttps://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1, with other regions substituting their regional host. - OpenRouter: Listed on OpenRouter’s Qwen4-Max page as
qwen/qwen4-maxwith a 1M context window. OpenRouter’s posted provider pricing for Alibaba Cloud International displays $1.425 input and $4.275 output per million tokens, $0.07125 for cache reads, and $1.78125 for 5-minute cache creation, which differs from QwenCloud’s $1.50 / $4.50 International list. - Vercel AI Gateway: Listed as
alibaba/qwen4-maxin the Vercel AI Gateway changelog and Vercel model catalog. Vercel’s catalog shows $1.50 input and $4.50 output per million tokens with a 1,000,000-token context window, though its changelog carried a November 15, 2026 gateway listing date ahead of Alibaba’s November 16 announcement and its model page displays 384,000 maximum output tokens rather than QwenCloud’s 131,072 / 262,144 production limits. - Novita AI: Available as
qwen/qwen4-maxvia OpenAI-compatible (https://api.novita.ai/openai) and Anthropic-compatible (https://api.novita.ai/anthropic) endpoints, per Novita AI’s launch post. Novita lists $1.413 input and $4.24 output per million tokens, $0.071 cached reads, $1.767 for 5-minute cache writes, and $2.827 for 1-hour cache writes, matching QwenCloud’s Beijing/Global base rates while displaying 977K context and 384K output on its model page alongside a 1,000,000 nominal context description.
How does Qwen4-Max pricing compare with Qwen3.8-Max, Claude, and Gemini?
Qwen4-Max enters the market at $1.50 per million input tokens and $4.50 per million output tokens on QwenCloud International, undercutting Qwen3.8-Max and current flagship Claude and Gemini tiers on list price as of November 17, 2026. The table below compares published standard list prices only; it is not a quality ranking, and choosing between providers also depends on ecosystem fit, rate limits, and how you match a model to your workload.
| Model | Provider / Tier | Input ($ / 1M tokens) | Output ($ / 1M tokens) | Context / Pricing Notes (as of Nov 17, 2026) |
|---|---|---|---|---|
Qwen4-Max (qwen4-max) |
Alibaba QwenCloud (Int.) | $1.50 | $4.50 | 1M context; $1.413 / $4.24 in Global & Beijing |
Qwen3.8-Max (qwen3.8-max) |
Alibaba QwenCloud (Int.) | $2.00 | $6.00 | Predecessor model on QwenCloud |
| Claude Sonnet 5.5 | Anthropic API | $2.00 | $10.00 | Official Claude pricing |
| Claude Opus 5.5 | Anthropic API | $4.00 | $20.00 | Official Claude pricing |
Gemini 4 Hydrogen (gemini-4-hydrogen) |
Google AI Studio (Paid) | $1.50 | $7.50 | Unified 1M context; output includes thinking; $0.075 implicit cache |
Gemini 4 Carbon (gemini-4-carbon) |
Google AI Studio (Paid) | $2.00 (then $4.00) | $10.00 (then $20.00) | $2 / $10 through Nov 30, 2026; $4 / $20 from Dec 1, 2026; 1M long-decode mention |
On Google’s Gemini API pricing page, Gemini 4 Hydrogen matches Qwen4-Max’s $1.50 input and $0.075 implicit cache read rate while charging $7.50 per million output tokens. Gemini 4 Carbon, which is excluded from Google’s free tier, sits at $2.00 / $10.00 through November 30, 2026 before doubling to $4.00 / $20.00 on December 1, 2026, matching Claude Opus 5.5’s list price.
One final gap for self-hosters is licensing: Qwen4-Max is a proprietary API model rather than an open-weight release. Alibaba stated in its press release that smaller-scale, open-weight Qwen4 models will follow soon, whereas teams looking for an open-weight model with a 1M context window today can already deploy Moonshot’s Kimi K3.1 Pro open-weights release from November 4, 2026.
FAQ
What is the difference between qwen4-max and qwen4-max-preview?
The production model ID qwen4-max launched on November 16, 2026, with pay-as-you-go and Token Plan billing, a 262,144-token combined thinking and answer cap, and high reasoning effort by default. The earlier qwen4-max-preview ID remains Token Plan only until its preview window closes on November 25, 2026 at 24:00 UTC+8, uses a 393,216-token combined output cap, and defaults to xhigh effort.
Does Qwen4-Max have a 1-million-token context window in production?
Yes, QwenCloud documentation specifies a 1,000,000-token context length for qwen4-max in production as of November 17, 2026. Maximum input reaches 991,808 tokens in non-thinking mode or 983,616 tokens with thinking enabled. Some third-party listings display slightly different figures, such as 984k on Artificial Analysis and 977K on Novita AI’s model page.
Are Qwen4-Max prices the same in every QwenCloud region?
No, QwenCloud separates Singapore International pricing from Global and Beijing pricing as of November 17, 2026. Singapore charges $1.50 per million input tokens and $4.50 per million output tokens, with 1 million free pay-as-you-go tokens. Beijing and Global regions (Frankfurt, Hong Kong, Tokyo, and US Virginia) share $1.413 input and $4.24 output pricing without free quota.
Will Alibaba release open-weight Qwen4 models?
Alibaba has not released open weights for Qwen4-Max, which remains a proprietary API model. However, the November 16, 2026 press release states that smaller-scale, open-weight models from the Qwen4 series will be released to developers globally soon. Developers needing open weights with a 1M context window today can look at Moonshot’s Kimi K3.1 Pro.
Sources
The specifications, pricing, and availability details in this article come from the following primary sources checked on November 17, 2026:
- Alibaba Cloud Press Room: Alibaba Unveils Qwen4-Max
- Qwen Official Blog: Qwen4-Max
- Qwen on X: Official Qwen4-Max Launch Post
- QwenCloud Docs: Latest Model (Qwen4-Max)
- QwenCloud Docs: Model Prices
- QwenCloud Model Card: qwen4-max
- Alibaba Cloud Model Studio: qwen4-max-preview
- QwenCloud Model Card: qwen3.8-max
- QwenCloud Docs: Qwen3.8-Max
- Artificial Analysis: Qwen4 Max
- Artificial Analysis: Changelog
- Artificial Analysis: Alibaba Models
- OpenRouter: qwen/qwen4-max
- Vercel AI Gateway Changelog: Qwen4-Max
- Vercel AI Gateway Model Page: alibaba/qwen4-max
- Novita AI Blog: Qwen4-Max Launch Pricing
- Anthropic: Claude API Pricing
- Google AI Studio: Gemini API Pricing



