Qwen4-Max costs $1.50 in and $4.50 out per 1M tokens with 1M context. API code, a price table vs Fable 5.2, GPT-6, Gemini 4, and who should switch.
Qwen4-Max is a closed-weights flagship from Alibaba that costs $1.50 per million input tokens and $4.50 per million output tokens on the international Model Studio list, with a 1,000,000-token context window.[1][2] That is 5.3x cheaper on input and 8.9x cheaper on output than Claude Fable 5.2 ($8 and $40).[8] The catch: Artificial Analysis scores it 43 on its Intelligence Index, #38 of 81 on its model page, and it writes a lot of tokens.[7] This guide covers the model ID, limits, a working curl and Python call from Alibaba's docs, a price comparison against Claude, GPT-6 and Gemini 4, what the open-weights promise actually covers, region caveats, and who should switch now versus wait.
What Qwen4-Max is
Qwen4-Max is the flagship of the new Qwen4 family. Alibaba's launch post describes a model with 2.4 trillion backbone parameters (71 billion activated per token, plus 328 billion N-gram embedding parameters), native vision and a one-million-token context window.[5] Model Studio describes it as built for coding, office productivity, frontend design and complex agentic tasks, and lists it as multimodal: text, image and video in, text out, with streaming, function calling, parallel tool calling, web search and JSON Schema structured output. Batch inference, fine-tuning and model deployment are not supported.[1]
A limited preview opened on November 11, and Alibaba announced the model on November 16.[6] The launch post says it is now available on QwenCloud.[5] The model ID for the API is qwen4-max.[1]
Alibaba's headline efficiency claim is on its own Work360 evaluation, where the average cost per task fell from $4.43 on Qwen3.8-Max to $1.67, a 62% drop.[5] That is a vendor number on a vendor benchmark. Treat it as a hypothesis to test on your workload, not a measurement.
Qwen4-Max pricing and limits
Model Studio publishes different list prices by deployment scope. The international (Singapore) list is the one most developers outside China will see first.
| Item | International (Singapore) | Global scope |
|---|---|---|
| Input, per 1M tokens | $1.50 | $1.413 |
| Output, per 1M tokens (includes chain of thought) | $4.50 | $4.24 |
| Cache read, per 1M tokens | $0.075 | $0.071 |
| Cache write, 5-minute TTL | $1.875 | $1.767 |
| Cache write, 1-hour TTL | $3.00 | $2.827 |
Source for every row: Alibaba's Model Studio pricing page.[2] There is a single price tier for requests up to 1M tokens, so a 900K-token prompt costs the same per token as a 9K one. For comparison, Qwen3.8-Max sits at $2 and $6 on the same international list, so Qwen4-Max is a 25% cut on both sides.[2]
Context and output limits
Context length is 1,000,000 tokens. Maximum input is 991,808 tokens in non-thinking mode and 983,616 in thinking mode. Maximum output is 131,072 tokens in non-thinking mode, and thinking plus the final answer can reach 262,144 combined.[1] Artificial Analysis lists the window as 984K, close to the 983,616-token thinking-mode input cap.[7]
Rate limits and free quota
The docs list a tokens-per-minute ceiling of 5,000,000 at the entry tier (monthly spend up to $8,000), 10,000,000 from $8,000 to $140,000, and 20,000,000 above that.[1] The international list shows a free quota of 1 million tokens, valid for 90 days from the latest of Model Studio activation, model release or application approval.[2] One million tokens is enough to try the API and not much more.
How to call Qwen4-Max
Qwen4-Max works through an OpenAI-compatible Chat Completions endpoint. It also supports the OpenAI Responses API, Anthropic Messages and Alibaba's native DashScope API.[1] The base URL is workspace-specific. Replace {WorkspaceId} with the ID from the Workspace Details page in the Model Studio console, and use an API key from the same region.[4]
curl
curl -X POST https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions \
-H "Authorization: Bearer $DASHSCOPE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen4-max",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Who are you?"}
]
}'This is the request shape from Alibaba's OpenAI-compatible reference.[4] The other regions use the same path with a different host prefix: us-east-1 for Virginia, eu-central-1 for Frankfurt, ap-northeast-1 for Tokyo, cn-hongkong and cn-beijing.[4]
Python with the OpenAI SDK
import os
from openai import OpenAI
client = OpenAI(
api_key=os.getenv("DASHSCOPE_API_KEY"),
# Replace {WorkspaceId} with your actual workspace ID. URLs vary by region.
base_url="https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1",
)
completion = client.chat.completions.create(
model="qwen4-max",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Who are you?"},
],
# extra_body={"enable_thinking": False},
)
print(completion.model_dump_json())Both snippets come straight from the Model Studio reference.[4] Existing OpenAI-SDK code usually needs only a changed key, base URL and model name.
Control thinking, because it controls your bill
Thinking is on by default. The defaults are enable_thinking=true, reasoning_effort=high and a thinking budget of 131,072 tokens.[1] The output price covers chain of thought plus the answer, so the default configuration spends the most.[2] On the OpenAI Chat Completions protocol, reasoning_effort accepts low, high and xhigh, and enable_thinking goes in extra_body from the Python SDK.[1] The reference also accepts none to disable thinking, and says reasoning_effort and thinking_budget cannot be set together for qwen4-max.[4]
# Cheap path for classification, extraction, routing
completion = client.chat.completions.create(
model="qwen4-max",
messages=[{"role": "user", "content": "Classify this ticket: ..."}],
extra_body={"enable_thinking": False},
)
# Agent path with lower reasoning cost
completion = client.chat.completions.create(
model="qwen4-max",
messages=[{"role": "user", "content": "Plan the migration steps: ..."}],
reasoning_effort="low",
)Turn on caching for long prompts
Cache reads cost $0.075 per million tokens against $1.50 for uncached input, a 95% discount.[2] Alibaba's Qwen4 caching docs show automatic caching through a prompt_cache_options field. The cached prefix must be at least 1,024 tokens, and a 5-minute cache write costs about 125% of the input price.[3]
completion = client.chat.completions.create(
model="qwen4-max",
messages=[{"role": "user", "content": "Summarize this document: " + long_text}],
extra_body={
"enable_thinking": False,
"prompt_cache_options": {"mode": "implicit"}, # automatic caching
},
)
print(completion.usage.prompt_tokens_details.cached_tokens)Put system prompts, tool definitions and reference documents first and the changing question last. A cache hit needs an identical prefix.[3] To size a prompt before you send it, paste it into our free token counter. Token counts differ by tokenizer, so treat any third-party estimate as approximate.
Qwen4-Max vs Claude, GPT-6 and Gemini 4 on price
All figures below are standard-tier list prices in USD per million tokens, taken from each vendor's own pricing page on November 16, 2026. OpenAI's page also offers a long-context view that this table does not use.
| Model | Input | Output | Cached input | Notes |
|---|---|---|---|---|
| Qwen4-Max (international) | $1.50 | $4.50 | $0.075 | 1M context, closed weights[2] |
| Claude Fable 5.2 | $8 | $40 | $0.20 | 1M context at standard pricing[8] |
| Claude Opus 5.5 | $4 | $20 | $0.20 | [8] |
GPT-6 Astra (gpt-6-astra) | $10 | $50 | $1.00 | Short-context list price[9] |
GPT-6.1 Sol (gpt-6.1-sol) | $2 | $10 | $0.10 | Short-context list price[9] |
| Gemini 4 Carbon / Argon | $2 (then $4) | $10 (then $20) | $0.10 (then $0.20) | Price doubles on December 1, 2026[10] |
| Gemini 4 Hydrogen | $1.50 | $7.50 | $0.075 | 1M context[10] |
On list price alone, Qwen4-Max ties Gemini 4 Hydrogen on input and undercuts every other row on output. Take a workload of 100 million input tokens and 20 million output tokens a month, with no caching and no thinking overhead. Qwen4-Max costs $240. Gemini 4 Hydrogen costs $300, GPT-6.1 Sol $400, Claude Opus 5.5 $800, Claude Fable 5.2 $1,600 and GPT-6 Astra $2,000. Gemini 4 Argon costs $400 until November 30 and $800 from December 1. For the Gemini change in detail, see our Gemini 4 Argon pricing breakdown, and for wider cross-vendor math, the LLM API pricing comparison.
List price is not task price
Here is where the table misleads. Artificial Analysis reports that Qwen4-Max generated 127 million output tokens across its Intelligence Index, against a median of 78.6 million for comparable models, and calls it very verbose.[7] It puts the cost per Intelligence Index task at $3.27 for Qwen4-Max.[7] Trending Topics, citing the same service, reports $10.04 for Claude Opus 5.5 and $1.27 for GPT-6.1 Sol, which makes Sol cheaper per task despite a higher list price.[11] A low per-token price buys less when the model spends more tokens. Run your own prompts and compare cost per finished task.
Quality: where it sits
Artificial Analysis lists Qwen4-Max as proprietary and scores it 43 on its Intelligence Index: #38 of 81 on the model page, above the median of 41 for comparable models.[7] Trending Topics, using the same index and each model's top setting, puts Claude Fable 5.2 at 65, Gemini 4 Carbon at 60, Claude Opus 5.5 at 59, GPT-6 Astra at 57 and GPT-6.1 Sol at 55.[11] The Next Web reports a mixed result: 75.7 on DeepSWE v1.1, ahead of Opus 5.5 at 74.2, but 48.5 on Terminal-Bench 4.0 against 59.6 for Opus 5.5.[6] It does not say who ran those evaluations, and the text of Alibaba's launch post does not repeat them, so treat them as unverified vendor-side figures. On the independent index, Qwen4-Max sits well below the leaders. The lower price does not buy matching quality.
What the open-weights promise covers
Alibaba's launch post says only that smaller models in the Qwen4 family are coming to open source, with no date and no license named.[5] The Next Web reports that Qwen4-Max itself cannot be downloaded and modified, and Artificial Analysis states the weights are not publicly available.[6][7]
So plan around what is documented:
- Qwen4-Max is API-only today. There are no weights to self-host, and Model Studio lists fine-tuning and model deployment as unsupported.[1]
- Open weights are promised for smaller Qwen4 models, with no date and no license named. Do not build a roadmap on an unannounced release.
Region and data-residency caveats
Model Studio tells developers to pick the region and service deployment scope that meet their data storage and inference location requirements before calling a model.[12] Qwen4-Max is offered in China (Beijing), Singapore, Germany (Frankfurt), US (Virginia), Japan (Tokyo) and Hong Kong, each with its own endpoint, and the pricing page lists International, Global and Chinese mainland scopes.[1][4]
Four practical consequences follow:
- API keys vary by region, so use a key from the same region as the host you call.[4]
- The cheaper $1.413 and $4.24 prices belong to the Global scope. Check what that scope means for where inference runs before sending regulated or customer data through it.[2]
- Prompt caches are reused only across requests from the same user.[3]
- Legal and compliance teams still need to review the provider's data retention and privacy terms. Model Studio's docs list a data retention policy and a privacy page. This article does not assess them.
Who should switch and who should wait
Switch a slice of traffic now if you run high-volume, long-prompt workloads where a mid-tier model is good enough: document summarization, extraction, internal search, support triage, first-pass code review. The 95% cache discount makes repeated 100K-token prefixes cheap. Route 5 to 10% of production prompts to qwen4-max, log cost per completed task, and decide from your own numbers.
Wait if your workload is hard agentic coding, terminal-heavy automation or anything where the top of the leaderboard matters. The 22-point Intelligence Index gap to Claude Fable 5.2 is real, and the Terminal-Bench 4.0 score The Next Web reports trails Opus 5.5.[11][6] Also wait if you need self-hosting, a fixed data region you have not verified, or the open-weights Qwen4 models that have no date yet.
A reasonable middle path is a router. Send routine calls to Qwen4-Max with thinking off, and escalate failures to a stronger model. Because the endpoint is OpenAI-compatible, the change is a base URL and a model string. Before you commit, use the AI token and cost estimator to turn your monthly token volumes into dollars, then add the thinking-token overhead you measure in testing.
Sources
- qwen4-max (Latest Model), Alibaba Cloud Model Studio documentation (accessed November 16, 2026)
- Model inference pricing, Alibaba Cloud Model Studio documentation (accessed November 16, 2026)
- Explicit cache (Qwen4), Alibaba Cloud Model Studio documentation (accessed November 16, 2026)
- OpenAI compatible Chat API reference, Alibaba Cloud Model Studio documentation (accessed November 16, 2026)
- Qwen4-Max: Your New Teammate Has Arrived, Qwen Team (accessed November 16, 2026)
- Alibaba launches Qwen4-Max flagship AI model without open weights, The Next Web, November 16, 2026 (accessed November 16, 2026)
- Qwen4 Max: Intelligence, Performance and Price Analysis, Artificial Analysis (accessed November 16, 2026)
- Pricing, Claude Platform Docs, Anthropic (accessed November 16, 2026)
- API pricing, OpenAI developer documentation (accessed November 16, 2026)
- Gemini API pricing, Google AI Studio documentation (accessed November 16, 2026)
- Alibaba's Giant Qwen4-Max Model Trails Claude, GPT and Moonshot AI, Trending Topics, citing Artificial Analysis (accessed November 16, 2026)
- Regions and service deployment scopes, Alibaba Cloud Model Studio documentation (accessed November 16, 2026)
Written by
WOWHOW
The WOWHOW team brings 14+ years of production engineering experience. Every tool and product in the catalog is built, tested, and curated in-house.
Monday memo · free
One insight, every Monday. 7am IST.
1 field report, 3 links, 1 tool we actually use. No fluff, no spam.
Beyond the writing
Need production-ready templates?
Free browser tools with no signup, plus premium dev templates and starter kits.
Comments · 0
Beta: comments are stored locally on your device and not visible to other readers.
No comments yet. Be the first to share your thoughts.