Qwen4-Max-Preview review, pricing and limits
Alibaba's first Qwen 4 model to reach developers, offered through Token Plan subscriptions until Alibaba retires it.
- Best for
- Developers previewing Qwen 4
- Not for
- Production backends and scripts
Should you use Qwen4-Max-Preview?
Alibaba is running Qwen4-Max-Preview, the first model of its Qwen 4 line, as a two-week Token Plan preview that ends on 25 November 2026, when Alibaba says it will be retired. It suits developers who want to try the newest Qwen generation inside Claude Code or Cursor, not teams that need a stable model or a pay-as-you-go endpoint.
Best for
- Developers previewing Qwen 4 Alibaba aims it at coding, office work, frontend design and agent tasks, and it runs inside the coding tools the Token Plan supports.
- Teams comparing models on one quota The Token Plan serves Qwen, DeepSeek and GLM models on the same Credits, so the preview can be tested against them without a second subscription.
Not for
- Production backends and scripts Token Plan terms limit use to interactive coding and agent tools, and the model goes away when the preview ends.
- Buyers who need published scores Alibaba has released no benchmarks and no model card for it, and no independent board lists it yet.
Can you use Qwen4-Max-Preview?
- Alibaba Cloud Token Plan: preview (Personal Edition from $6 a month, and Team Edition; interactive use in coding and agent tools only; no pay-as-you-go). Preview ends 25 November 2026, then Alibaba retires the model.
- QwenWork: preview.
- Qoder: preview.
- Qwen app: preview.
- QwenStudio: preview.
- Free plan: not published.
- 25 Nov: Preview ends and Alibaba retires qwen4-max-preview.
Checked 13 Nov against Alibaba Cloud's own pages.
Qwen4-Max-Preview pricing
- Pricing model
- subscription
- Free tier
- No
| Tier | Rate |
|---|---|
| Token Plan Personal: Lite | $6 a month, or $3 for new users until 30 November 2026; 20,000 Credits a month; 1 to 2 concurrent agents |
| Token Plan Personal: Essential | $10 a month, or $5 for new users until 30 November 2026; 45,000 Credits a month; 2 to 3 concurrent agents |
| Token Plan Personal: Standard | $18 a month; 80,000 Credits a month; 3 to 4 concurrent agents |
| Token Plan Personal: Pro | $43 a month; 200,000 Credits a month; 6 to 8 concurrent agents |
Billed in Credits from a monthly Token Plan quota, not per token, and Alibaba publishes no per-token price for this model. Credits per call depend on the model type, token use, thinking mode and tool calls, and the console shows actual use. During the preview Alibaba cuts this model's Credits use to as little as 20% of the standard rate, with a further 60% off between 22:00 and 08:00 UTC+8 that brings it to 8%. When the quota runs out, an Extra Bundle adds 20,000 Credits for $15 a month; otherwise the service pauses until the next month.
What Qwen4-Max-Preview can do
Qwen4-Max-Preview is the first model of Alibaba's Qwen 4 flagship line, with a 1M-token context window and a 393,216-token output budget shared by thinking and the answer. It reads text, images and video, writes text, and runs only on Alibaba Cloud's Token Plan during a two-week preview.
What goes in and what comes out
Reads
- text
- image
- video
Writes
- text
Capabilities
- Builtin toolsSupported
- Function callingSupported
- Preserve thinking defaultSupported
- Reasoning enabled by defaultSupported
- Structured outputSupported
- Tool useSupported
- Video inputSupported
- VisionSupported
- Web searchSupported
- Reasoning modes
- low, medium, xhigh
Key features
Three reasoning levels
Thinking is on by default at xhigh; choose low, medium or xhigh with reasoning_effort, and do not send it together with thinking_budget.
Image and video input
Takes text, images and video for screenshot analysis, document understanding and video analysis, and answers in text.
Tools and structured output
Supports function calling, web search, built-in tools, and JSON Object and JSON Schema output.
Three API formats
Accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages requests.
Show all 6 features
Preserved thinking
Chat Completions keeps preserve_thinking on by default, so earlier thinking goes back in the reasoning_content field of assistant messages.
One shared output budget
Thinking and the final answer draw on a single output limit, which Alibaba sets above its previous flagship's.
Known limitations
- Only a Token Plan gets you in; there is no pay-as-you-go route.
- The reasoning_effort and thinking_budget settings cannot be sent together.
- The Personal Edition is available only in the Singapore region.
Who it is for
- Developers evaluating Qwen 4 in coding tools
- AI engineers comparing models on one Token Plan quota
- Analysts tracking Alibaba's Qwen releases
Strengths and limits
Strengths
- Larger output budget than the previous flagship: Alibaba's pages list 393,216 tokens for the preview against 131,072 for Qwen3.8-Max.
- Alibaba says it will make no changes to the model during the preview, so results from two weeks of testing stay comparable.
- One Token Plan quota also covers other vendors' models, so the preview can be tried against Qwen3.7-Max, DeepSeek and GLM models without a second subscription.
Limits
- No independent evidence exists yet: Alibaba has published no benchmarks or model card, Artificial Analysis had no page for it, and the Arena text board listed no Qwen 4 model, all read on 13 November 2026.
- Short life: Alibaba says it will be retired when the preview ends, so anything built on it needs another model within two weeks of launch.
- Data terms are thin on the Personal Edition: Alibaba states global inference with cross-border transfer but no training or retention policy, unlike the Team Edition's no-training statement.
Qwen4-Max-Preview alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Qwen4-Max-Preview This page | Alibaba Cloud | 1M | -- | text, image, video → text |
| Qwen3.8-MaxPick Qwen3.8-Max if you need a model you can call from your own backend or keep using after the preview. Pick Qwen4-Max-Preview only to test the newest Qwen generation in a coding tool before Alibaba retires it. | Alibaba Cloud | 1M | $2 / $6 | text, image, video → text |
 Gemini 4 HydrogenGoogle DeepMind's November 2026 Gemini 4 model for coding agents, with four thinking levels from low to xhigh and a free tier in AI Studio. | 1M | $1.50 / $7.50 | text, image, audio, video, pdf → text | |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
Qwen4-Max-Preview specs
- Context window
- 1.0M tokens
- Max output
- 393K tokens
- Architecture
- Not disclosed
- Parameters
- Not disclosed
Regions
- Singapore
Prompting patterns
Show 3 prompting patterns
Leave room for thinking when you set an output limit
Set max_tokens well above the length of the answer you expect.
Alibaba's notes say thinking uses the same budget as the final answer.
Choose reasoning_effort or thinking_budget, not both
reasoning_effort: "medium"
Alibaba's guide says the two settings must not be sent together.
Return earlier thinking in the reasoning_content field on multi-turn calls
Put each assistant turn's thinking in reasoning_content.
Alibaba keeps preserve_thinking on for Chat Completions and says history belongs in reasoning_content, not content.
Known quirks
Show 3 known quirks
max_tokens now covers thinking
For this model, max_tokens equals max_completion_tokens in Chat Completions, while existing Qwen models keep the old behavior.
Workaround: Allow room for both thinking and the final answer when choosing the limit.
Two model IDs, one billing name
The alias qwen4.0-max-preview works as well as qwen4-max-preview, but bills show only qwen4.0-max-preview.
Workaround: Match usage rows by the alias when you reconcile a bill.
Thinking starts at xhigh
Thinking runs unless you change it, and the starting effort is xhigh, the deepest of the three levels.
Workaround: Set reasoning_effort to low or medium for quick edits.
Qwen4-Max-Preview facts
- Deployment
- API
- Released
- 11 Nov 2026
- Current version
- qwen4-max-preview
- Family
- Qwen4
- Variant
- Multimodal
- Licence
- Not published (access under Alibaba Cloud Token Plan terms) · Proprietary
- Data retention
- Token Plan Team Edition: Alibaba states that conversation data stays out of model training. Token Plan Personal Edition: Alibaba's page states Singapore-region service with global inference and cross-border transfer of prompts and outputs, and states no training or retention policy. No retention period was read for either edition.
Content restrictions
- Token Plan use is limited to interactive work inside coding and agent tools; automation scripts, custom application backends and batch calls are not allowed.
Version history
qwen4-max-preview · 11 Nov 2026
Preview opens to Token Plan subscribers; Alibaba says the model is retired when the preview ends on 25 November 2026.
qwen3.8-max-0902 · 02 Sep 2026
Snapshot of the previous flagship, Qwen3.8-Max, listed on Alibaba's page, where it is sold pay-as-you-go.
What changed
Questions about Qwen4-Max-Preview
How much does Qwen4-Max-Preview cost?
Alibaba lists no per-token price for it, and pay-as-you-go access is not supported, so you pay for a Token Plan. Personal Edition plans cost $6 a month for Lite with 20,000 Credits, $10 for Essential with 45,000, $18 for Standard with 80,000 and $43 for Pro with 200,000, and Lite and Essential drop to $3 and $5 for new users until 30 November 2026. During the preview each call uses as little as 20% of the standard Credits, and 8% between 22:00 and 08:00 UTC+8.
Can you use Qwen4-Max-Preview for free, and are the weights open?
Alibaba's pages list no free tier or trial for it, so the cheapest way in is the Lite Token Plan. Its documentation lists no weights download, and AlphaSignal's launch coverage reports that open weights for the Max model are unconfirmed. Treat it as a hosted, closed model unless Alibaba says otherwise.
When does the Qwen4-Max-Preview preview start and end?
It opened at 10:00 on 11 November 2026 (UTC+8) and runs until 24:00 on 25 November 2026. Alibaba says nothing will change during that window and that the model is retired afterwards. Alibaba's Model Studio list names no other Qwen 4 model, so a Qwen 4 27B or Flash model is not on offer there.
Qwen4-Max-Preview vs Qwen3.8-Max: which should you use?
Alibaba's pages give both a million-token context window, and the preview a larger output budget. Qwen3.8-Max is the one Alibaba sells per token in six regions, according to its Model Studio page, while the preview is subscription-only and retired when it ends.
Can you call Qwen4-Max-Preview from your own app or backend?
Not on the Token Plan terms Alibaba publishes: the plan is for interactive use inside coding and agent tools, and Alibaba says calls from scripts, application backends or batch jobs can get a subscription suspended or an API key banned. Each verified identity can hold one subscription per edition, and the Personal Edition is for its owner's exclusive use. The API accepts OpenAI Chat Completions, Responses and Anthropic Messages requests at a dedicated Token Plan base URL.
About Qwen4-Max-Preview
Qwen4-Max-Preview is the first model from Alibaba Cloud's Qwen 4 line to reach developers, and Alibaba's documentation calls it the flagship multimodal model of the newest Qwen generation. The preview opened on 11 November 2026 (UTC+8), the day the Qwen team announced it on X, and Alibaba says the model will be retired when the preview ends. It follows Qwen3.8-Max, the 2.4-trillion-parameter mixture-of-experts flagship that Alibaba still sells.
Read the full background
Alibaba says it is designed for coding, office productivity, frontend design and complex agentic tasks, and the Qwen team's announcement promises deeper knowledge, richer world understanding and stronger coding and agentic ability. Neither the announcement nor the documentation backs those claims with a benchmark or a model card, and AlphaSignal's launch coverage reports the same gap.
The only route in is a Token Plan, a subscription billed in Credits and meant for coding and agent tools such as Claude Code, Cursor, Qwen Code, Qoder and OpenClaw. The announcement also lists early access in QwenWork, Qoder, the Qwen app and QwenStudio. The same plan serves Qwen3.7-Max, DeepSeek-V4 Pro and GLM-5.3, so one subscription can set the preview beside them on your own tasks.
It has a million-token context window, and thinking and the visible answer share one output budget that is larger than Qwen3.8-Max's. Alibaba changed how that budget works for this model: max_tokens now equals max_completion_tokens, so a limit sized for the answer alone can cut a thinking model short, while existing Qwen models keep their old behavior. Thinking is on by default, and the default effort is xhigh.
The Personal Edition of the Token Plan is sold only in the Singapore region and runs inference globally, so prompts and outputs cross borders and Alibaba leaves legal compliance to the user. The Team Edition page says conversation data is not used for model training; the Personal Edition page states no training policy, and no retention period appears on either page.
Per Alibaba's Qwen3.8-Max model page, read on 13 November 2026, Qwen3.8-Max is sold pay-as-you-go through the Model Studio API in six regions and carries no retirement notice, while the preview is subscription-only, limited to interactive tools and retired when its window closes. No published scores exist either: Alibaba has released no benchmarks, and Artificial Analysis and Arena had no entry for it when read on 13 November 2026.
Sources and guides
Official sources
- Qwen4-Max-Preview model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Alibaba Cloud's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Lists that include Qwen4-Max-Preview
The provider
More AI Models on Hokai
- Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens.
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
- Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.