Qwen4-Max is now a normal, pay-as-you-go model. Alibaba’s Qwen team published the launch post on 15 November 2026, and by early 16 November (UTC) the model ID qwen4-max was live on the QwenCloud API at $1.50 per million input tokens and $4.50 per million output tokens.
That is 25% cheaper than Qwen3.8-Max, with a 1M-token context, native image and video input, and benchmark charts that put it close to the top US models on agentic coding. The two-week, subscription-only preview I covered last week is now replaced by a model anyone can call.
This guide covers the confirmed specs, every price, the benchmarks (and which ones Alibaba made itself), how to call it from Python, curl and Claude Code, how it compares with GPT-6 Astra, Claude and Gemini, and the errors you are most likely to hit.
- What: Alibaba’s new flagship LLM, model ID
qwen4-max. Closed weights, API and apps only. - Price: $1.50 in / $4.50 out per 1M tokens (Singapore), same price with or without thinking. Cache reads $0.075.
- Free: 1M free tokens for new Singapore-region accounts, valid about 90 days.
- Specs: 2.4T total parameters, 71B active (MoE), 1M context, up to 262,144 output tokens with thinking.
- Benchmarks: 75.7 on DeepSWE v1.1 (top of Alibaba’s chart), but 48.5 on Terminal-Bench 4.0, well behind Claude Opus 5.5 and GPT-6 Astra. All vendor-reported.
- Preview users:
qwen4-max-previewstill retires on 25 November. Switch your model name toqwen4-max.
What Launched and When
Qwen4-Max is the top model in Alibaba’s Qwen4 family. Alibaba announced the family (Max, Plus, Flash and a 27B model) at its Apsara Conference on 29 October 2026, then opened a limited preview on 11 November.
The preview, qwen4-max-preview, was sold only through the Token Plan subscription. My Qwen4-Max Preview guide covers that setup in detail.
The full release adds the things the preview lacked: a per-token price, a benchmark table, an architecture description and access in several regions.
| Date (2026) | Event |
|---|---|
| 2 Sep | Qwen3.8-Max-0902 released (previous flagship) |
| 29 Oct | Qwen4 family announced at Apsara |
| 11 Nov, 02:00 UTC | Qwen4-Max-Preview opens on Token Plan only |
| 15 Nov | Qwen team blog post: Qwen4-Max officially launched |
| 16 Nov | qwen4-max listed with pay-as-you-go prices; Bloomberg reports the launch |
| 25 Nov, 16:00 UTC | Preview model retired |
Note: Hugging Face had no Qwen4 models at launch. Alibaba says smaller Qwen4 models will be open-sourced later but gave no date or sizes. Qwen4-Max itself stays closed.
Qwen4-Max Specs
Alibaba finally shared the architecture. It is a sparse mixture-of-experts model, so only a small part of its weights work on each token.
| Spec | Qwen4-Max | Qwen3.8-Max (for reference) |
|---|---|---|
| Model ID | qwen4-max |
qwen3.8-max, qwen3.8-max-0902 |
| Parameters | 2.4T total, 71B active per token, plus 328B N-gram embedding parameters | Not disclosed |
| Architecture | Sparse MoE; Gated DeltaNet mixed 3:1 with Qwen Sparse Attention (QSA) | MoE |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Max output | 131,072 (no thinking); 262,144 thinking + answer | 128K; 256K thinking budget |
| Input | Text, images, video | Text, images, video |
| Output | Text | Text |
| Tools | Function calling, web search, code execution, structured output | Same |
| API formats | OpenAI Chat Completions, OpenAI Responses, Anthropic Messages | Same |
| Images / video | Up to 16M pixels per image; 2,048 images by URL or 250 Base64; 64 videos, each up to 2 h / 2 GB | Same limits |
| Batch API | Not supported on the international model page | 50% batch discount |
| Fine-tuning | Not supported | Not supported |
| Weights | Closed | Closed |
Parameter and architecture details come from the Qwen blog post as reported by DataLearner and FourWeekMBA. Limits come from QwenCloud’s text generation model docs.
Heads up: some write-ups still quote 393,216 output tokens. That was the preview’s limit. The released qwen4-max lists 262,144 tokens for thinking plus answer.
Pricing, Free Tokens and Cache Costs
Pricing is flat across the whole 1M context, and thinking mode costs the same as non-thinking. That makes budgeting much easier than tiered models.
| Item (per 1M tokens) | Singapore (international) | Beijing / Global |
|---|---|---|
| Input | $1.50 | $1.413 |
| Output | $4.50 | $4.24 |
| Cache read | $0.075 | $0.071 |
| Cache write (5-minute) | $1.875 | $1.767 |
| Cache write (1-hour) | $3.00 | $2.827 |
| Free quota | 1M tokens for new accounts | None |
| Mainland China list price | ¥10 input / ¥30 output, ¥0.5 cache hit | |
Source: QwenCloud model prices page, read 16 November. The console may show temporary discounts on top of these list prices.
What a typical job costs
- A 20K-token prompt with a 4K-token answer: about $0.03 + $0.018 = $0.048.
- A coding agent session that reads 2M tokens (mostly cached) and writes 300K: roughly $0.15 for cache reads, $0.30 for new input and $1.35 output, so under $2.
- Alibaba’s own figure: an average Work360 agent task fell from $4.43 on Qwen3.8-Max to $1.67 on Qwen4-Max, because it uses fewer tokens and costs less per token.
Tip: thinking tokens are billed as output. If you only need short answers, turning thinking off or lowering the effort is the biggest cost lever. See my reasoning effort and API cost guide for how this plays out across models.
How the free 1M tokens work
- Granted per model, so Qwen4-Max has its own 1M allowance. Input and output both count.
- Usually valid for 90 days from account activation. Unused tokens expire and are not reissued.
- Only for calls made with a Singapore API key to the Singapore endpoint.
- Does not cover built-in tool fees, such as web search.
- You must add payment details before the quota appears.
Benchmarks: Where It Wins and Where It Doesn’t
Alibaba published a 16-panel chart with the launch. Every number below is vendor-reported. Independent leaderboards had not listed the model when this was written.
| Benchmark | Qwen4-Max | Best rival in Alibaba’s chart | What it measures |
|---|---|---|---|
| DeepSWE v1.1 | 75.7 | GPT-6.1 Sol 75.2, Claude Opus 5.5 74.2, GPT-6 Astra 74.1 | Agentic software fixes |
| Terminal-Bench 4.0 | 48.5 | Kimi K3.1 67.2, Opus 5.5 59.6, Astra 59.1 | Working in a real terminal |
| Vibe Code Bench v1.1 | 86.66 | Not listed in reports | Building apps from prompts |
| BioMysteryBench | 85 | Reported as first place | Biology reasoning |
| Agents’ Last Exam v1.1 | 46.7 | Reported as first place | Long agent tasks |
| HLE (Humanity’s Last Exam) | 51.7 | Behind Opus 5.5 and GPT-6.1 Sol | Hard expert questions |
| Work360 (Alibaba’s own) | 52.69 | Qwen3.8-Max 47.74 | Office and agent work |
Sources: FourWeekMBA and DataLearner summaries of the Qwen blog charts, plus AlphaSignal for Work360.
The honest summary: Qwen4-Max ties the US frontier on code-fixing tasks, leads a few niche tests, and clearly trails on terminal work. FourWeekMBA counted it as top in only 4 of the 16 panels.
The bigger gain is efficiency. Alibaba says it used about 48% fewer tokens on Terminal-Bench 4.0 than Qwen3.8-Max (151K vs 289K per attempt) while scoring 48.5 instead of 37.9.
Be careful with Work360: Alibaba built that benchmark itself, so treat its 10% gain as a claim, not proof. Wait for Artificial Analysis and LMArena results before switching anything important.
How to Get an API Key and Make Your First Call
You need a QwenCloud account. Pay-as-you-go keys and Token Plan keys are different, so pick the right one.
- Create an account. Sign up at qwencloud.com and add a payment method. This also unlocks the free quota.
- Choose the region. Use Singapore (
ap-southeast-1) if you want the free 1M tokens. Keys only work in the region where you create them. - Create an API key. In the console, open API Keys and create a pay-as-you-go key. Store it as an environment variable, for example
DASHSCOPE_API_KEY. - Pick the base URL. OpenAI-style clients use
https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1. Anthropic-style clients usehttps://ap-southeast-1.maas.qwencloudapi.com/apps/anthropic. - Set the model to
qwen4-maxand send a test request (below).
Test with curl
curl https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1/chat/completions \
-H "Authorization: Bearer $DASHSCOPE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen4-max",
"messages": [{"role": "user", "content": "Explain MoE models in two lines."}],
"max_completion_tokens": 8192,
"stream": false
}'
Python with the OpenAI SDK
import os
from openai import OpenAI
client = OpenAI(
api_key=os.getenv("DASHSCOPE_API_KEY"),
base_url="https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1",
)
stream = client.chat.completions.create(
model="qwen4-max",
messages=[{"role": "user", "content": "Write a Python function that dedupes a list, keeping order."}],
extra_body={"enable_thinking": True},
max_completion_tokens=16384,
stream=True,
)
for chunk in stream:
if not chunk.choices:
continue
delta = chunk.choices[0].delta
if getattr(delta, "reasoning_content", None):
print(delta.reasoning_content, end="", flush=True) # thinking
if delta.content:
print(delta.content, end="", flush=True) # answer
Other regions use the same paths with a different host. Replace ap-southeast-1 with eu-central-1 (Frankfurt), us-east-1 (Virginia), ap-northeast-1 (Tokyo), cn-hongkong or cn-beijing. Check the console’s model list first, because not every model is in every region.
Try it without code
Alibaba lists Qwen4-Max in Qwen Studio, QwenWork, Qwen Office and the Qoder coding app. Qwen Chat (chat.qwen.ai) carried the preview. Usage limits for these apps were not published.
Use Qwen4-Max in Claude Code and Other Agents
Because QwenCloud speaks the Anthropic Messages format, Claude Code can use Qwen4-Max as its main model. This is the most popular way people tested the preview.
- Open your settings file:
~/.claude/settings.jsonon macOS/Linux orC:\Users\<you>\.claude\settings.jsonon Windows. - Add the env block below with your pay-as-you-go key.
- Skip the Anthropic login by setting
"hasCompletedOnboarding": truein~/.claude.json. - Open a new terminal, run
claude, then type/statusto check the base URL.
{
"env": {
"ANTHROPIC_AUTH_TOKEN": "YOUR_QWENCLOUD_KEY",
"ANTHROPIC_BASE_URL": "https://ap-southeast-1.maas.qwencloudapi.com/apps/anthropic",
"ANTHROPIC_MODEL": "qwen4-max",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen4-max",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen3.8-flash",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen3.8-flash",
"CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000"
}
}
The Sonnet and Haiku slots point at qwen3.8-flash ($0.15 / $0.47) so quick background tasks stay cheap. If you are on the Token Plan instead, keep the token-plan.maas.qwencloudapi.com/apps/anthropic URL and your sk-sp- key from the preview setup.
To switch between Claude, Qwen and other providers without editing JSON each time, use profiles as shown in my Claude Code Router guide.
Cline, Cursor, OpenCode and Codex
- Provider type: OpenAI-compatible.
- Base URL:
https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1(no extra/chat/completions). - Model:
qwen4-max. - Context setting: up to 1,000,000; many tools default lower, so raise it if you load big repos.
Thinking and Output Settings
Most surprises with Qwen4-Max come from thinking settings. Here is what the docs and early write-ups say.
| Setting | What to know |
|---|---|
enable_thinking |
Hybrid model: thinking is on by default. Pass false (via extra_body in the OpenAI SDK) for fast, cheap replies. |
reasoning_effort |
Write-ups on the release report a default of high, adjustable to low or xhigh. The preview used low/medium/xhigh with xhigh as default. Check the API reference if a value is rejected. |
thinking_budget |
Works on Chat Completions, not Responses. Do not send it together with reasoning_effort. |
max_tokens |
With thinking on, it must be between 1 and 32,768. Use max_completion_tokens for longer outputs. |
| Multi-turn | Send earlier reasoning back in the assistant message’s reasoning_content field, not inside content. |
| Streaming | Recommended for anything long, to avoid HTTP timeouts. |
Tip: start with thinking off for chat and simple edits, low or default effort for normal coding, and xhigh only for hard debugging or planning.
Qwen4-Max vs GPT-6 Astra, Claude and Gemini
The headline is price. Qwen4-Max costs a fraction of the US flagships, while scoring close to them on some coding tests.
| Model | Input / output per 1M | DeepSWE v1.1 | Terminal-Bench 4.0 | Open weights |
|---|---|---|---|---|
| Qwen4-Max | $1.50 / $4.50 | 75.7 | 48.5 | No |
| Qwen3.8-Max | $2 / $6 | Not in chart | 37.9 | No |
| GPT-6 Astra | $10 / $50 | 74.1 | 59.1 | No |
| GPT-6.1 Sol | Output $10 | 75.2 | Not reported | No |
| Claude Opus 5.5 | $4 / $20 | 74.2 | 59.6 | No |
| Claude Fable 5.2 | $8 / $40 | Not in chart | Not in chart | No |
| Gemini 4 Carbon | Output $10 (rises to $20 on 1 Dec) | Not in chart | Not in chart | No |
| Kimi K3.1 | See Moonshot pricing | Not reported | 67.2 | Yes |
Prices: QwenCloud, Anthropic, OpenAI and Google list prices as reported on 16 November (FourWeekMBA and our own pricing posts). Scores: Alibaba’s launch chart. For full US price sheets, see my OpenAI API pricing breakdown and Claude Fable 5.2 guide.
Who should pick which
- Pick Qwen4-Max for high-volume coding agents, long-document work and image/video understanding where cost matters most.
- Stay with Claude Opus 5.5 or GPT-6 Astra for terminal-heavy agents and tasks where the last few points of reliability matter.
- Pick Kimi K3.1 if you need open weights and top terminal scores. My Kimi K3.1 Pro guide covers hardware needs.
- Run Qwen locally with Qwen3.8-Flash-Next until the smaller Qwen4 open models arrive.
For a wider table of providers, see my AI API pricing comparison.
Moving From the Preview
If you set up qwen4-max-preview last week, here is what changes.
- Model name: change
qwen4-max-previewtoqwen4-maxbefore 25 November, 16:00 UTC, when the preview is retired. - Token Plan users: write-ups say Qwen4-Max is on Token Plan Personal and Team. Alibaba has not published its credit rate yet, so watch the Analytics page.
- Output cap: drops from 393,216 to 262,144 tokens (thinking plus answer). Few jobs will notice.
- Key type: Token Plan keys (
sk-sp-) only work on the token-plan URL. For pay-as-you-go, make a new key and use the regional URL. - Backends: the Individual Token Plan bans app backends and scripts. If you want Qwen4-Max inside your own product, use pay-as-you-go.
Limits, Privacy and What’s Still Unknown
- Data location: pick a region close to you, but read QwenCloud’s data terms before sending work code. Only paste what your employer allows.
- Region-locked keys: a key created in Singapore fails against the Frankfurt URL, and vice versa.
- No batch discount on the international page and no fine-tuning or dedicated deployment.
- Rate limits are per account and not published for this model; bursts can return 429.
- Unknown: independent benchmark results, the Token Plan credit rate, release dates for Qwen4-Plus, Qwen4-Flash and Qwen4-27B, and licences for the open models.
Do not rely on unofficial numbers: an “Artificial Analysis score of 53” circulated at launch, but the index had not listed the model. Wait for the official listing.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
model_not_found or “model does not exist” |
Region does not offer qwen4-max yet, or a typo such as qwen-4-max |
Use exactly qwen4-max; try the Singapore endpoint |
| 401 Invalid API-key provided | Key from another region, a Token Plan key on the pay-as-you-go URL, or a stray space | Create the key in the same region as the base URL |
| 403 invalid api-key | Pay-as-you-go key sent to the token-plan URL | Match key type and URL |
400 on max_tokens |
Value above 32,768 with thinking on | Use max_completion_tokens |
| Request rejected when setting thinking | reasoning_effort and thinking_budget sent together |
Send only one |
| Answer stops mid-sentence | Thinking used up the output limit | Raise max_completion_tokens, lower effort, or set CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000 |
| 429 Throttling | Account rate limit or no balance after free quota | Back off and retry; top up the account |
| Free quota not used | Calls going to a non-Singapore region | Use a Singapore key and endpoint |
| Claude Code: “Unable to connect to Anthropic services” | Still pointing at api.anthropic.com | Check /status, set hasCompletedOnboarding, open a new terminal |
| Timeout on long tasks | Non-streaming request held open too long | Turn on stream: true |
My Take
This is a far better launch than the preview. A clear per-token price, a real benchmark chart and multi-region access make Qwen4-Max something you can actually plan around.
The price is the story. At $4.50 per million output tokens, it is about 78% cheaper than Claude Opus 5.5 and 91% cheaper than GPT-6 Astra, while matching them on Alibaba’s DeepSWE test. For coding agents that burn millions of tokens, that gap is hard to ignore.
The weak spot is terminal work, where it sits more than 10 points behind. Alibaba’s charts also include its own Work360 test, so I would wait for independent results before moving anything critical. For side projects and cost-sensitive agents, the free 1M tokens make it an easy thing to try this week.
FAQ
Is Qwen4-Max free?
New QwenCloud accounts in the Singapore region get 1M free tokens for Qwen4-Max, usually valid for 90 days. After that it costs $1.50 per million input tokens and $4.50 per million output tokens.
Is Qwen4-Max open source?
No. Qwen4-Max is closed and only available through Alibaba’s API and apps. Alibaba says smaller Qwen4 models will get open weights later, with no date yet.
What is the Qwen4-Max model ID?
Use qwen4-max. The old qwen4-max-preview ID stops working after 25 November 2026.
How big is Qwen4-Max?
Alibaba says 2.4 trillion total parameters with 71 billion active per token, plus 328 billion N-gram embedding parameters. It is a sparse mixture-of-experts model.
Does thinking mode cost more?
The per-token price is the same with or without thinking. Thinking still costs more in practice because the reasoning tokens are billed as output.
Can I use Qwen4-Max with Claude Code?
Yes. Set ANTHROPIC_BASE_URL to https://ap-southeast-1.maas.qwencloudapi.com/apps/anthropic, use your QwenCloud key, and set ANTHROPIC_MODEL to qwen4-max.
Is Qwen4-Max better than Claude Opus 5.5?
On Alibaba’s own chart it scores slightly higher on DeepSWE v1.1 (75.7 vs 74.2) but much lower on Terminal-Bench 4.0 (48.5 vs 59.6). It is far cheaper. Independent tests are not out yet.
Can I run Qwen4-Max locally?
No. There are no downloadable weights. For local use, run an open Qwen3.8 model now and watch for the smaller Qwen4 releases.
Sources
- Qwen team blog: “Qwen4-Max” (qwen.ai/blog?id=qwen4-max), 15 November 2026
- QwenCloud docs: Text generation models, Model prices, Regions and endpoints, Free quota, Thinking guide (docs.qwencloud.com), read 16 November 2026
- Alibaba Cloud Model Studio: Qwen4-Max-Preview page (retirement dates)
- FourWeekMBA: “Qwen4-Max Launches at $4.50/M Output, 91% Below GPT-6 Astra”, 16 November 2026
- AlphaSignal: “Alibaba’s Qwen4-Max Slashes Agent Task Costs 62% With 1M-Token Context”
- DataLearner: Qwen4-Max model page (architecture and benchmark list)
- ai-tldr.dev: Qwen4-Max release page
- KuCoin News (AIBase repost): Alibaba launches Qwen4-Max, 16 November 2026
- Hacker News discussion of the Qwen4-Max blog post, 16 November 2026




