Qwen4-Max is in preview until November 25. Where to try it, and what it costs

What Alibaba released

On November 11, 2026, the evening of November 10 in the US, Alibaba’s Qwen team posted: “Qwen4 is coming. Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities.” Then: “Qwen4-Max is now in preview on Alibaba’s Token Plan. You can also get early access to the new model on QwenWork, Qoder, Qwen APP, and QwenStudio.”

QwenCloud’s page for the model calls it “the next-generation flagship multimodal model in the Qwen family,” aimed at coding, office work, front-end design and complex agent tasks. What it gives you:

• Its ID: qwen4-max-preview. The alias qwen4.0-max-preview works too, and it’s the name your bill shows.
• Context: 1 million tokens in, and up to 393,216 tokens out, with thinking and the answer sharing that budget.
• Input: text, images and video.
• Thinking: on by default. “Choose low, medium, or xhigh (default).”
• Tools: function calling, web search, built-in tools and structured JSON output.
• APIs: OpenAI’s Chat Completions and Responses, and Anthropic’s Messages, which is how Claude Code, Codex and OpenCode connect to it.

Then the conditions. “Qwen4-Max-Preview is currently in preview and is available through Token Plan only. Pay-as-you-go access is not supported.” The preview “runs from 10:00 on November 11, 2026 to 24:00 on November 25, 2026 (UTC+8),” and Alibaba adds that “no changes will be made during this period, and the model will be retired after the preview ends.” In US time, it ends on November 25 at 10 a.m. Central, 8 a.m. Pacific, 11 a.m. Eastern.

Where you can use it

Route	What it costs	What you get
Token Plan Individual	$6 to $43 a month	A key for Claude Code, Codex, OpenCode, Qwen Code and other agents
Token Plan Team Edition	$20 to $180 a seat a month	The same for teams, with seats and usage reports
Qoder	Qoder’s credits, at 0.5× for this model	Alibaba’s own coding app
Qwen app, QwenStudio, QwenWork	Not published	Early access, in Qwen’s words, with no published limits
QwenCloud’s pay-as-you-go API	Not offered	Nothing, until Alibaba says otherwise

Qoder’s model selector lists Qwen4-Max-Preview at “0.5×” credit usage, without saying which Qoder plans include it. For the Qwen app, QwenStudio and QwenWork, Alibaba’s post says “early access” and nothing more: no plan, no limits, no price.

What Token Plan costs

Token Plan is a monthly subscription billed in credits. Here is the Individual edition, from Alibaba’s docs on November 12:

Plan	A month	New-user price	Credits a month	Agents at once
Lite	$6	$3	20,000	1–2
Essential	$10	$5	45,000	2–3
Standard	$18	None	80,000	3–4
Pro	$43	None	200,000	6–8

The new-user prices are for accounts that have never bought a Token Plan Individual subscription, until 23:59:59 on November 30, China time. The credits went up on November 9, when Pro also dropped from $68 to $43 (changelog); Lite had 11,500 credits before. A Credit Pack adds 20,000 credits for $15, up to five at a time, used after the plan’s own run out. Credits reset each subscription month, and unused ones don’t carry over. The Team Edition is $20, $75 or $180 a seat for 25,000, 100,000 or 250,000 credits.

For the preview, Qwen4-Max is discounted: “Credits consumption is as low as 20% of the standard rate (5x usage),” and “calls between 22:00 and 08:00 (UTC+8) enjoy an additional 60% off, that is, 8% of the standard rate.” China’s night is America’s workday: that window is 9 a.m. to 7 p.m. Eastern, or 6 a.m. to 4 p.m. Pacific.

The missing number is the standard rate itself. Alibaba doesn’t publish how many credits a Qwen4-Max token uses, so there’s no way to tell how much work 20,000 credits buys. Its FAQ says only that a pricier model uses more credits for the same tokens, and that rate limits “are adjusted dynamically based on overall load.”

Read the terms before you pay. “Token Plan Individual is non-refundable,” it can’t be canceled, only kept from renewing, and each account gets one Individual subscription, for its owner alone. Its only region is Singapore, but “the model inference will be performed globally,” and “the prompts and model outputs will involve cross-border data transfer.”

How to set it up

Get a Token Plan key from Alibaba’s console. It starts with sk-sp- and works only with the Token Plan address; a pay-as-you-go key there returns 401 invalid_api_key.

Claude Code. Alibaba’s guide sets these in a terminal:

export ANTHROPIC_AUTH_TOKEN="YOUR_API_KEY"
export ANTHROPIC_BASE_URL="https://token-plan.maas.qwencloudapi.com/apps/anthropic"
export ANTHROPIC_MODEL="auto"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="qwen3.6-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3.8-flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen4-max-preview"
export CLAUDE_CODE_SUBAGENT_MODEL="auto"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS="983616"

Then claude -p "Hello" checks it works. Claude Code started from that terminal sends everything to Alibaba, and your Claude plan isn’t used. Note where Qwen4-Max sits: in the Opus slot, with the main model on auto. Claude Code’s model docs say these variables control what its opus, sonnet and haiku aliases point to, so pick Opus with /model, or set ANTHROPIC_MODEL to qwen4-max-preview, to be sure you’re testing the new model. Alibaba also gives the same values for ~/.claude/settings.json, but that switches every Claude Code session you start, so the terminal is the better place to try it. If long answers get cut off, it suggests CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000.

Codex. For Codex 0.94.0 and later, Alibaba’s guide adds this to ~/.codex/config.toml and puts your key in OPENAI_API_KEY:

model_provider = "Token_Plan"
model = "auto"

[model_providers.Token_Plan]
name = "Token_Plan"
base_url = "https://token-plan.maas.qwencloudapi.com/compatible-mode/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"

Change model to qwen4-max-preview to use it. Alibaba says to keep wire_api = "responses" for this model, and that a warning about missing model metadata doesn’t affect calls.

OpenCode. Alibaba’s guide has you run opencode, type /connect, choose “Alibaba Token Plan (China)”, the name it gives for both editions, and paste your key. Then /models lists Qwen4 Max Preview; if it doesn’t, update OpenCode. OpenCode caps this model’s output at 32,000 tokens unless you raise it with OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX.

One setting catches tools that set a small output cap. On this model, max_tokens and max_completion_tokens “limit the combined tokens for thinking and the final answer,” and thinking starts at xhigh, so a tight cap can come back empty because the thinking used all of it. Raise the cap, or turn the effort down.

How good it is

Alibaba hasn’t said, beyond its post’s “stronger coding and agentic capabilities.” There are no benchmarks, no model card and no parameter count on its pages. When we read Artificial Analysis’s leaderboard on November 12, it had no Qwen4 entry. Its predecessor, Qwen3.8 Max (0902), scores 34 on the leaderboard’s Intelligence Index.

With a two-week window, independent results may arrive close to the day the model retires. Qwen’s own post says “Put it to the test on a real problem,” and for now that is the only test there is.

The rules that come with Token Plan

Token Plan is for working with an agent yourself, not for anything that runs on its own. Alibaba’s terms say: “Use is limited to interactive use within programming and agent tools (such as Claude Code, Cursor, Qwen Code, Qoder, OpenClaw, etc.).” And: “It must not be used for automation scripts, custom application backends, or any non-interactive batch call scenarios.” The FAQ adds that “background scheduled tasks” aren’t allowed either, and the overview says calls outside that scope “may result in subscription suspension or API Key banning.” You can set the same key up “on multiple devices you own,” but sharing it with others “may be considered a violation.”

On training, the Team Edition says “Token Plan Team Edition does not use your conversation data for model training.” The Individual pages we read make no such statement either way.

Qwen4-Max in Enjoy

Enjoy runs OpenCode, and its Model column lists the models OpenCode offers from your providers (how to set it up). Add the Token Plan to OpenCode as above, reopen the agent settings, and choose Qwen4 Max Preview there. Your Claude Code conversations stay on your Claude plan, and you can hand a conversation from Claude Code to OpenCode and back, so the same task gets a second answer from Qwen4-Max without starting over.

Two limits. Don’t put Qwen4-Max in a scheduled workflow: Token Plan’s terms forbid background scheduled tasks, so a local model is the one for the night shift. And Enjoy doesn’t add Alibaba’s models to any other lab’s account; you still need the Token Plan (Enjoy’s pricing).

What to do before November 25

1. Decide if two weeks are worth a month. A new account’s Lite plan is $3, can’t be refunded, and keeps the other Token Plan models after Qwen4-Max retires. If you buy, buy early: every day you wait is a day less of the preview.
2. Test it on a task you know. Keep your usual setup in another terminal, give both the same job, and compare the diffs, not the prose.
3. Use the cheap hours. From 9 a.m. to 7 p.m. Eastern, calls cost as little as 8% of the standard rate instead of 20%.
4. Don’t build on it. It retires on November 25, and Alibaba hasn’t named its successor. Qwen3.8-Max stays, on Token Plan and in the pay-as-you-go API at $2 in and $6 out per million tokens in the international region (prices).

What Alibaba hasn’t said

• How many credits a Qwen4-Max call uses. The preview discount is a share of a standard rate Alibaba doesn’t publish.
• How good it is. No benchmarks, no model card, no parameter count, no license.
• Whether weights will follow. Its September Apsara post said only that Qwen 4 “is currently in training,” with no dates and nothing about open models.
• What comes after November 25. No date for the full Qwen4-Max or the rest of the family, and no word on a pay-as-you-go API.
• The limits in the Qwen app, QwenStudio and QwenWork.
• Whether Individual subscribers’ data trains Alibaba’s models.

We’ll update this memo when the preview ends or Alibaba fills these in.