Summarize this article with:
Alibaba released Qwen4-Max on November 16, 2026. On Alibaba's international list it is 25% cheaper per token than Qwen3.8-Max. Reasoning is on by default and billed as output. Unlike the previous generation, there are no downloadable flagship weights. This article covers what changed, what the benchmarks show, what a request costs, and who should switch.
Alibaba released Qwen4-Max on November 16, 2026, under the model ID qwen4-max. On Alibaba's international list it costs $1.50 / $4.50 per 1M input / output tokens, 25% less than Qwen3.8-Max. Reasoning is on by default, and the flagship ships without downloadable weights.
That is a real tradeoff. API users get lower token prices and cheap cache reads. Teams that self-host lose the path the previous generation offered through the open, text-only Qwen3.8-2.4T-A95B. Thinking is also billed as output, so your savings depend on how you set reasoning effort and token limits.
This article covers the key facts, the changes from Qwen3.8-Max, what the benchmarks show, what a request costs, and who should switch.
Qwen4-Max key facts
| Released | November 16, 2026, by Alibaba |
|---|---|
| Architecture | 2.4T-parameter MoE, 71B active per token |
| Context | 1,000,000 tokens (unchanged) |
| Inputs and output | Text, image, video in; text out (unchanged) |
| Reasoning | On by default at high; low, high, xhigh |
| Output limit | 262,144 tokens, thinking and answer combined |
| International price | $1.50 / $4.50 per 1M input / output tokens |
| Weights | Not released; API-only |
Alibaba keeps the 2.4-trillion-parameter mixture-of-experts backbone of Qwen3.8-Max. According to the South China Morning Post, Qwen4-Max activates 71 billion parameters per token, against 95 billion for Qwen3.8-Max. Yotta Labs' launch summary adds 328B n-gram embedding parameters and Qwen Sparse Attention with Gated DeltaNet layers. A smaller active count does not by itself prove faster responses.
What changed from Qwen3.8-Max
| Attribute | Qwen4-Max | Qwen3.8-Max | What it means |
|---|---|---|---|
| Input / output per 1M | $1.50 / $4.50 | $2.00 / $6.00 | 25% lower list price |
| Active parameters | 71B | 95B | Speed gain not shown |
| Reasoning | On by default (high) | Configurable effort | Output budgets must include thinking |
| Open weights | None for the flagship | Text-only Qwen3.8-2.4T-A95B | Qwen4-Max is API-only |
| Context and inputs | 1M; text, image, video | 1M; text, image, video | No change |
Sources: Alibaba Cloud qwen4-max documentation, Yotta Labs, SCMP and the Qwen3.8-2.4T-A95B model card. International (Singapore) prices, checked November 17, 2026.
Distribution is the bigger change. The Qwen3.8-2.4T-A95B model card offers downloadable weights for a text-only model that always thinks. Vision input and a non-thinking mode came only with the hosted Qwen3.8-Max. According to Yotta Labs, Alibaba says smaller Qwen4 models will be open-sourced, but it has given no names or dates.
Alibaba also tightened safety. According to SCMP, it stripped offensive cyberattack capabilities from the training data while keeping vulnerability detection. It also added post-training mitigations for chemical, biological, radiological and nuclear (CBRN) risks.
What the benchmark evidence shows
Almost all published numbers come from Alibaba's launch materials. They show a clear step up from Qwen3.8-Max, not an overall winner.
On Work360, a new Alibaba benchmark covering 721 specialized fields of professional work, Qwen4-Max scored 52.69 against 47.74. Average cost per task fell 62%, from $4.43 to $1.67. Alibaba designed this test and ran it on its own workload. According to Yotta Labs, Alibaba credits the lower rate plus fewer tokens: about 24% fewer at xhigh effort and 35% fewer at high.
| Benchmark | Qwen4-Max | Qwen3.8-Max | Top score in the chart |
|---|---|---|---|
| DeepSWE v1.1 (agentic coding) | 75.7 | 69.3 | Qwen4-Max, 75.7 |
| HLAB-Vals-1.1 (legal agents) | 21.3 | 15.4 | Qwen4-Max, 21.3 |
| Terminal-Bench 4.0 (terminal coding) | 48.5 | 38.9 | Kimi K3.1, 67.2 |
| ProgramBench (software replication) | 31.5 | 28.0 | Claude Opus 5.5, 65.0 |
| Humanity's Last Exam | 51.7 | 43.9 | Claude Opus 5.5, 59.8 |
Source: Alibaba's November 16, 2026 launch chart, as reported by Yotta Labs. Qwen3.8-Max is the 0902 snapshot. Settings vary by row.
Yotta Labs counts Qwen4-Max ahead of Qwen3.8-Max on all 15 benchmarks where both have a score. It has the top score on 4 of the 16. It trails the leaders clearly on terminal coding and ProgramBench. Alibaba's footnotes say several rows come from third-party leaderboards such as Artificial Analysis and Vals, while Alibaba ran others itself.
Independent results are still few. Artificial Analysis lists Qwen4-Max at 43 on its Intelligence Index as of November 16, with an average cost of $3.27 per task on that run. That run uses a different workload from Work360, so the two costs are not comparable.
Our Kimi K3 vs Qwen 3.8-Max comparison covers the previous generation. Compare like with like: run representative tasks on each model with the same prompts, tools and output limits.
Pricing and availability
Qwen4-Max is available through several channels:
- Alibaba Cloud Model Studio and QwenCloud's pay-as-you-go API
- The credit-based Token Plan
- Alibaba's QwenWork, Qoder and QwenStudio apps
The API accepts the OpenAI Chat Completions and Responses formats, as well as the Anthropic Messages and DashScope formats. The default limit is 5,000,000 tokens per minute.
Prices depend on the region. The Singapore (international) list charges $1.50 / $4.50 per 1M input / output tokens. Beijing and the global regions (Frankfurt, Hong Kong, Tokyo, Virginia) list $1.413 / $4.24. Singapore pay-as-you-go also includes a free quota of 1 million tokens, valid for 90 days.
| Item (Singapore list) | Price per 1M tokens | Example |
|---|---|---|
| Input | $1.50 | 500,000-token prompt: $0.75 |
| Cached input (cache read) | $0.075 | Same prompt from cache: $0.0375 |
| Cache write | $1.875 (5 min) / $3.00 (1 h) | For explicit cache breakpoints |
| Output, including thinking | $4.50 | Full 262,144-token budget: about $1.18 |
Sources: Alibaba Cloud qwen4-max documentation and Yotta Labs. Examples computed from list prices. Checked November 17, 2026.
Thinking is billed as output and is on by default. The default thinking_budget is 131,072 tokens, so a request that uses all of it pays about $0.59 for reasoning before the answer. In Chat Completions, max_tokens now equals max_completion_tokens. It caps thinking and the answer together: up to 262,144 tokens with thinking on, and 131,072 with it off. A limit sized for the answer alone can cut the response short.
The 25% rate cut accounts for only part of the 62% Work360 saving. The rest depends on how many tokens the model uses, so you will see that saving only if your tasks behave like Alibaba's.
If you tested the preview, note that qwen4-max-preview is a separate model ID. It is available only through Token Plan, from November 11 until 24:00 on November 25, 2026 (UTC+8). It defaulted to xhigh effort and allowed 393,216 output tokens. Alibaba recommends the production model.
Should you switch to Qwen4-Max?
Switch if hosted access is acceptable and your workload uses reasoning or repeated context. Don't switch for the 1-million-token context or video input alone, because Qwen3.8-Max already had both.
Decision by workload
Who gains from Qwen4-Max
Self-hosting
Stay on Qwen3.8
There are no Qwen4 flagship weights. The text-only Qwen3.8-2.4T-A95B remains the downloadable option, and smaller open Qwen4 models have no announced dates.
Repeated context
Test the cheaper cache
At $0.075 per 1M cached input tokens, repeated documents, system prompts and tool definitions get much cheaper. Measure your real cache-hit rate.
Price-sensitive reasoning
Set effort and limits
Pick low, high or xhigh per task and cap max_tokens. With high effort on by default, thinking can become the largest item on the bill.
Coding agents
Run your own tasks
Alibaba's chart puts Qwen4-Max first on DeepSWE but well behind the leader on Terminal-Bench 4.0. Test with fixed repositories, tools and timeouts.
Security tools
Check the guardrails
According to SCMP, vulnerability detection stays but offensive capabilities were removed from training data. Test authorized security workflows before migrating.
Regulated workloads
Check hosting terms
Qwen4-Max runs only on Alibaba's hosted service, in Singapore, Beijing or global regions. Review data controls, region choice and the new CBRN restrictions.
A lower list price is only one part of the bill. Our guide to reducing LLM API costs covers output limits, caching and routing. To weigh what giving up self-hosting costs, see open versus proprietary models.
How to prepare with Eden AI
A gateway sits between your code and the model providers, so changing models becomes a configuration change. Eden AI gives you one API key and one OpenAI-compatible endpoint for more than 1,000 AI models, from LLMs to OCR, speech and image. You can list models and their capabilities through the API, so check which Qwen4 IDs are live before you hard-code one.
Portability has limits. The same client code does not make reasoning controls, safety behavior or tool calling identical across models. Provider limits also change. For example, Qwen4-Max's output ceiling went from 393,216 tokens in preview to 262,144 at launch. Our post on provider-specific limits covers this kind of change.
Python
from openai import OpenAI
client = OpenAI(
api_key="YOUR_EDENAI_API_KEY",
base_url="https://api.edenai.run/v3",
)
# Find Qwen4 model IDs in the live catalog
qwen4_ids = [m.id for m in client.models.list().data if "qwen4" in m.id.lower()]
print(qwen4_ids)
if qwen4_ids:
response = client.chat.completions.create(
model=qwen4_ids[0],
messages=[{"role": "user", "content": "Summarize this specification in five bullets."}],
max_tokens=4000, # on Alibaba's API, this caps thinking and answer together
)
print(response.choices[0].message.content)
For production, set up automatic fallback to backup models, so a rate limit or an outage at one provider does not stop your application. The price of every model is listed in the catalog. Check it before you move traffic.
What to keep in mind
- No self-hosting: Qwen4-Max is API-only. Smaller open Qwen4 models have no announced names, sizes, dates or license terms.
- Savings are Alibaba's numbers: the 62% task-cost cut comes from Work360, Alibaba's own benchmark.
- Thinking is billed as output: high effort and a 131,072-token thinking budget are the defaults, so set effort and
max_tokensyourself. - Independent evidence is thin: the launch chart mixes Alibaba's runs with third-party leaderboards. Artificial Analysis's index score of 43 is the main independent figure so far.
- Prices vary by region: Singapore, Beijing and the global regions use different price lists.
FAQ
When was Qwen4-Max released?
Is Qwen4-Max open weight?
How much does the Qwen4-Max API cost?
Is reasoning enabled by default in Qwen4-Max?
How large is the Qwen4-Max context window?
Does Qwen4-Max accept video input?
Can Qwen4-Max be used for cybersecurity work?
How can I check whether Qwen4-Max is on Eden AI?
Further reading
- Alibaba's Qwen4-Max documentation: capabilities by region, reasoning parameters, prices and rate limits.
- Qwen4-Max preview documentation: preview dates, Token Plan access and the preview's output limits.
- Yotta Labs' Qwen4-Max summary: Alibaba's full benchmark chart, cache prices and Token Plan tiers.
- Qwen3.8-2.4T-A95B model card: the open weights, architecture and serving options of the previous generation.
- Kimi K3 vs Qwen 3.8-Max: benchmarks, pricing and API access for the previous Qwen flagship.
- Qwen models on Eden AI: our earlier announcement on Qwen model pricing.



