Qwen4-Max Released: Alibaba Flagship Now Closed, Context of 1 Million Tokens, and the Impact for Indonesian Developers
By Tim APIPedia, Tim Engineering APIPedia · Published · 9 min read
Alibaba released Qwen4-Max, a closed flagship model with a context of 1 million tokens and a price of $1.50/$4.50. These are the facts, claims and impacts for Indonesia.

Alibaba released Qwen4-Max on Monday, November 16, 2026, as a new flagship AI model of the Qwen family that can only be used via API and is not released as open-weight. The model has a context window of 1 million tokens, accepting text, image, and video input, priced at $1.50 per 1 million input tokens and $4.50 per 1 million output tokens in the Singapore region. This differs from many previous Qwen releases in that it can be downloaded and run on its own.
For developers and digital business owners in Indonesia, this release is important for two reasons: the price of the high-end models has dropped again, and the self-hosting option for the most powerful Qwen models is not available. This article summarizes the facts of the release, claims that still need to be tested, dollar cost calculations, and sensible steps to take before you move workloads.
Quick summary: what are the key facts about Qwen4-Max?
Qwen4-Max is a closed model with a mixture-of-experts architecture with a context of 1 million tokens available from November 16, 2026 via the Alibaba Cloud API.
| Items | Details |
|---|---|
| General release date | November 16, 2026 (limited preview since November 11) |
| Maker | Alibaba (Qwen team) |
| Model ID in API | qwen4-max |
| Architecture | Mixture-of-experts, 2.4 trillion total parameters, about 71 billion actives per token |
| Context window | 1,000,000 tokens (maximum input 991,808 tokens in non-thinking mode) |
| Maximum output | 131,072 tokens (non-thinking), 262,144 tokens for thinking plus answers |
| Modality | Input text, images, videos. Text output only |
| Singapore Price (International) | $1.50 input and $4.50 output per 1 million tokens |
| Prices in other regions | $1,413 input and $4.24 output per 1 million tokens |
| Model weight | Closed, can't download |
| API Protocol | OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, DashScope |
What is Alibaba actually releasing?
Alibaba released one big model called the Qwen4-Max which replaced the Qwen3.8-Max as the top model, and its weight was not shared with the public. The company calls it an “AI co-worker” for work, creativity and research. This model can be accessed via the Alibaba Cloud Model Studio and QwenCloud APIs, and installed in Alibaba's own products such as Qwen Studio, QwenWork, and Qoder.

Features noted in the official documentation
- Function calling, including calling multiple tools in parallel.
- Deep thinking mode with a summary of the thinking process.
- Structured output in the form of JSON Object and JSON Schema.
- Explicit (TTL 5 minutes and 1 hour) and implicit caching.
- Web search, streaming, and prefix completion.
- Not yet supported: batch inference, fine-tuning, and custom model deployment.
In which regions is the model available?
Qwen4-Max is available in six locations: Singapore (International), China (Beijing), Hong Kong, Japan (Tokyo), Germany (Frankfurt), and the US (Virginia). The documentation does not list the Indonesian region for this model, so the closest point for users here is Singapore. New accounts in the Singapore region get a free quota of 1 million tokens which is valid for 90 days.
Why is this release important?
This release is important because the world's most downloaded open-weight model family now puts its best models behind an API. So far, many teams have chosen Qwen precisely because it can be run on their own server. Alibaba says a smaller version of the Qwen4 will be released with open weights, but hasn't given a date.
The price is down compared to its predecessors and competitors
In the Singapore region, Qwen4-Max is around 25% cheaper than Qwen3.8-Max which is pegged at $2.00 input and $6.00 output per 1 million tokens. Compared to the list prices of other high-end models as of November 16, 2026, the gap is wider: Kimi K3.1 Pro $2 and $10, Claude Opus 5.5 $4 and $20.

Performance claims: where are the facts, where are the vendor claims?
Specifications, price and availability are facts recorded in official documentation, while all performance figures below are claims by Alibaba that have not been independently verified.
- Coding: Alibaba says Qwen4-Max is at the top of its coding benchmarks. The release charts show mixed results: 75.7 on DeepSWE v1.1 (Claude Opus 5.5: 74.2), but 48.5 on Terminal-Bench 4.0 (Claude Opus 5.5: 59.6, Kimi K3.1: 67.2).
- Efficiency: on Work360, Alibaba's internal benchmark, the average cost per task dropped from $4.43 to $1.67 (around 62%) versus Qwen3.8-Max, with around 35% less token usage in high reasoning mode. This comparison is against Alibaba's own legacy model, not against competitors.
- Secure code: the pass rate for code that is both functional and secure in the Last Secure Code Bench rose from 22% to 31.8%.
- Autonomous working: Alibaba claims this model works autonomously for 48 hours to design chip cooling materials which are then manufactured and tested in the lab.
Changes on the security side
Alibaba said training data for defensive security tasks (finding and patching gaps) was retained, while data aimed at improving offensive capabilities was discarded. There is also additional security training for chemical, biological, radiological and nuclear risks. The evaluation was carried out by Alibaba's internal team, no external red team test results have been published yet.
What is the impact for developers and businesses in Indonesia?
The direct impact is threefold: the API costs for high-end models are getting cheaper, Qwen's flagship models cannot be self-hosted, and the data you send will be processed outside of Indonesia.
How much does it cost in rupiah?
With the November 16 2026 reference exchange rate of IDR 18,259 per US dollar, the price of Qwen4-Max in the Singapore region is equivalent to around IDR 27,400 per 1 million input tokens and IDR 82,200 per 1 million output tokens. Example calculation for a chatbot with 10,000 requests per day, 2,000 input tokens and 500 output tokens each:
| Model | Input 20 million tokens | Output 5 million tokens | Total per day |
|---|---|---|---|
| Qwen4-Max (Singapore) | $30.00 | $22.50 | $52.50 (around Rp. 959 thousand) |
| Qwen3.8-Max (Singapore) | $40.00 | $30.00 | $70.00 (approximately IDR 1.28 million) |
| Claude Opus 5.5 | $80.00 | $100.00 | $180.00 (approximately IDR 3.29 million) |
This calculation uses list prices without cache and without token thinking. Thinking mode increases output tokens, so real bills can be higher. If you're not used to counting tokens, read the explanation of AI tokens like credit.
Who is most affected?
- Teams that use expensive API models for coding or agents. There are new candidates that are much cheaper to test, especially on tasks with long contexts.
- The team running Qwen on its own servers. The flagship model is not available on-premise. The choice is to stick with the previous generation open Qwen model or wait for the small version of the Qwen4.
- Businesses that process personal data. Requests are processed in overseas regions. Check cross-border data transfer obligations in the Personal Data Protection Law (Law 27/2022) and internal policies before sending customer data.
- Preview users. According to Alibaba documentation, model
qwen4-max-previewstops on November 25, 2026 at 24:00 (UTC+8), so change the model ID toqwen4-maxbefore that date.
What do you need to do now?
The safest move is to test Qwen4-Max on your own data first, rather than immediately moving production based on vendor benchmark numbers.
- Try with free quota. Activate Model Studio in the Singapore region, create an API key in the same region as the endpoint being called.
- Test with 20 to 50 real cases. Take the prompt from the production log, compare the response quality, latency from Indonesia, and number of tokens with the model you are currently using.
- Recalculate costs with token thinking. Note the actual output token, because active thinking mode can change the calculation.
- Make use of cache for long prompts. The context of 1 million tokens is tempting to include all the documents, but the costs go up accordingly. See how prompt caching works.
- Set up a fallback. Don't rely on one provider for your application. Set a retry limit and 429 error handling.
- Review the data flow. Remove or obfuscate personal data before sending, and note the processing region for compliance needs.
Example of a call via an OpenAI compatible endpoint
The following example follows the Model Studio documentation format for the Singapore region. Replace {WorkspaceId} with your workspace ID.
curl --location 'https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
"model": "qwen4-max",
"messages": [
{"role": "user", "content": "Ringkas perbedaan REST dan gRPC dalam 3 poin."}
]
}'Because the protocol is compatible, applications that already use the OpenAI SDK simply change the base URL, API key, and model name for testing.
Frequently asked questions
Is Qwen4-Max open source?
No. The Qwen4-Max is a closed model and its weights cannot be downloaded. Alibaba says the smaller Qwen4 model will be released with open weights, but there's no official date yet.
How much does the Qwen4-Max cost?
In the Singapore region the price is $1.50 per 1 million input tokens and $4.50 per 1 million output tokens. In Beijing, Hong Kong, Tokyo, Frankfurt, and Virginia the prices are $1,413 and $4.24. Cache is calculated separately.
Is Qwen4-Max better than Claude or GPT?
It's not certain yet. Alibaba claims to be ahead in some coding benchmarks, but its own release charts show Qwen4-Max lagging behind in Terminal-Bench 4.0. All numbers are from the vendor, so test them in your own case.
Can Qwen4-Max be run on your own server or laptop?
Can't. This model is only available via API and Alibaba products, and Model Studio does not yet support custom deployment or fine-tuning for this model.
Is there an Indonesian region for Qwen4-Max?
This model documentation does not include the Indonesian region. The closest region is Singapore (International), so request data is processed outside Indonesia.
Related reading on APIPedia: Qwen-based Jev and Qev decision models for low-cost classification tasks, as well as migration guides from OpenAI's Agent Builder and Evals for those of you reorganizing your AI stack.
Source
- Alibaba Cloud Model Studio: Qwen4-Max documentation (specifications, price, region)
- Alibaba Cloud Model Studio: OpenAI compatibility and base URL per region
- Qwen's official announcement on X, November 16, 2026
- Bloomberg via The Star: Alibaba unveils new flagship AI model in response to US rivals
- The Next Web: Alibaba launches Qwen4-Max flagship AI model without open weights
Topics: #ai #qwen #alibaba #llm #harga api #berita
About the author
Tim APIPedia — Tim Engineering APIPedia
Tim engineering di balik APIPedia yang membangun dan memelihara API data Indonesia - ongkir, OCR KTP, data wilayah, BMKG, hingga server MCP - dan menulis panduan integrasinya dari pengalaman langsung.