Alibaba's Qwen team has released the full version of its next-generation AI model, Qwen4-Max. Following the preview version, which became available on the subscription-based Token Plan on November 11, the model became available through the pay-as-you-go API on November 16, 2026 (Japan time). According to the official Qwen blog, the new model uses a 2.4-trillion-parameter architecture and has been strengthened for coding, professional tasks, and long-running AI agent execution.
Qwen4-Max's defining feature is the combination of performance approaching top-tier models with relatively low API pricing. In Alibaba's comparison chart against other companies' models, Qwen4-Max took first place in only 4 of 16 benchmarks, and Claude Opus 5.5 outperformed it on many of the coding-related ones. Even so, its list price of $4.5 per million output tokens is less than a quarter of Opus 5.5's $20 and less than half of GPT-6.1 Sol's $10.
This price gap matters especially when operating AI agents that call a model dozens of times. However, even when per-token prices are low, the actual cost can flip if a model consumes more tokens to finish the same job or needs more retries after failures. Evaluating Qwen4-Max's competitiveness requires looking at the cost of completing a task, not just benchmark scores and API unit prices.
2.4-Trillion-Parameter Architecture Improves Processing Efficiency Over the Previous Generation
Qwen4-Max adopts the new Qwen4 architecture. The backbone has 2.4 trillion total parameters, but only 71 billion are active when processing a single token. Separately, it uses 328 billion parameters for "N-gram Embedding," which handles short sequences of tokens. The context length is 1 million tokens, and the model accepts images and video in addition to text.
The new architecture combines Gated DeltaNet and Qwen Sparse Attention. Gated DeltaNet compresses past context into a fixed-size state, while Qwen Sparse Attention selects the parts of a long input that need processing. According to Alibaba, combining three layers of Gated DeltaNet with Qwen Sparse Attention is designed to reduce computation when handling long contexts.
Compared with the previous-generation Qwen3.8-Max-Base, the company says active parameters per processing step were cut by 25% while performance improved in all four areas of its internal evaluation. BPB (Bits Per Byte), a measure of language prediction efficiency, fell by 5–13%. Lower BPB values indicate better predictive performance, and Alibaba explains that it improved computational efficiency while maintaining model scale.
API pricing has also been reduced from the previous generation. In the Singapore region, Qwen3.8-Max cost $2 per million input tokens and $6 per million output tokens, whereas Qwen4-Max costs $1.5 and $4.5 respectively, both 25% cheaper. The price for input hitting the implicit cache fell 70%, from $0.25 to $0.075.
Even in Alibaba's Own Comparison, Qwen4-Max Leads in Only 4 of 16 Benchmarks
The official Qwen blog includes a chart comparing Qwen4-Max, Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, and Moonshot AI's Kimi K3.1 across 16 benchmarks. Qwen4-Max achieved the highest score in four: DeepSWE for agentic coding (75.7), Agents' Last Exam for long-running real-world tasks (46.7), HLAB-Vals for legal agents (21.3), and BioMysteryBench for bioinformatics (85.0). Claude Opus 5.5, meanwhile, led in seven.
The gap is most noticeable in coding. On Terminal-Bench 4.0, which tests solving development tasks through terminal operations, Qwen4-Max scored 48.5, compared with 59.6 for Opus 5.5 and 67.2 for Kimi K3.1. On ProgramBench, which involves reimplementing existing software by observing its behavior from the outside, Qwen4-Max scored 31.5 while Opus 5.5 recorded 65.0, more than double. On CodeArena, a head-to-head web development evaluation, Qwen4-Max's 1750 also falls below Opus 5.5's 1816.
- Qwen4-Max
- Claude Opus 5.5
データを表で見る
| Qwen4-Max (Score) | Claude Opus 5.5 (Score) | |
|---|---|---|
| DeepSWE v1.1 | 75.7 | 74.2 |
| Agents' Last Exam v1.1 | 46.7 | 43.4 |
| APEX-Agents v1.1 | 72.1 | 73.5 |
| JobBench | 68.5 | 75.4 |
| Terminal-Bench 4.0 | 48.5 | 59.6 |
| ProgramBench | 31.5 | 65 |
In agent-related evaluations, strengths also vary. On APEX-Agents, which targets professional services work, Qwen4-Max scored 72.1 and Opus 5.5 scored 73.5, a gap of just 1.4 points. On JobBench, which covers job-related tasks, the scores were 68.5 vs. 75.4, and on AgentArena, which evaluates real-world agent tasks head-to-head, they were 7.5 vs. 18.2, so the gap with Opus 5.5 widens.
According to footnotes on the official comparison chart, APEX-Agents uses figures from Mercor's leaderboard and CodeArena uses figures from LM Arena's leaderboard, so not every evaluation is an internal Alibaba measurement. However, Alibaba chose which benchmarks to include. Even in this comparison, Opus 5.5 holds the advantage in difficult coding areas, so it cannot be said that Qwen4-Max surpasses it in overall performance.
Output Price Is Less Than Half of GPT-6.1 Sol and Under a Quarter of Opus 5.5
While Qwen4-Max falls short of competitors in some areas of performance, it has the advantage in API pricing. Standard API prices for major models, based on the official price lists from QwenCloud, OpenAI, and Anthropic, are as follows.
| Model | Input | Output | Cache read |
|---|---|---|---|
| Qwen4-Max (Singapore International) | $1.5 | $4.5 | $0.075 |
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| GPT-6 Astra | $10 | $50 | $1 |
List prices in US dollars per million tokens. Excludes batch processing, discounts, and additional tool fees. Prices for GPT-6.1 Sol and GPT-6 Astra apply to inputs of 272,000 tokens or fewer.
Looking at output pricing alone, Qwen4-Max's $4.5 is 45% of GPT-6.1 Sol's $10 and 22.5% of Claude Opus 5.5's $20. On input, Qwen4-Max's $1.5 is 25% cheaper than GPT-6.1 Sol and 62.5% cheaper than Opus 5.5. Under the same comparison conditions, Qwen4-Max also has the lowest cache-read price.
QwenCloud's prices vary by the region where inference runs. According to the full model's pay-as-you-go price list, the following unit prices apply to the Singapore International inference scope and to the Global scope covering Beijing, Tokyo, and others.
| Region / inference scope | Input | Output (thinking + answer) | Cache write (5 min) | Cache write (1 hour) | Cache read |
|---|---|---|---|---|---|
| Singapore International | $1.5 | $4.5 | $1.875 | $3 | $0.075 |
| Beijing China / Global | $1.413 | $4.24 | $1.767 | $2.827 | $0.071 |
In the Global scope, which includes Tokyo, output costs $4.24 per million tokens. At an exchange rate of 159 yen to the dollar, that is about 674 yen, a large gap from Opus 5.5's $20, which comes to about 3,180 yen at the same rate. QwenCloud also advertises a free allowance of 1 million tokens for new users of the Singapore pay-as-you-go API.
As an actual cost example, using 100,000 input tokens and 10,000 output tokens (thinking and answer combined) in Singapore International costs $0.15 for regular input and $0.045 for output, for a total of $0.195. For the same volume, if the cache-read price applied to the entire input, input would cost $0.0075, bringing the total with output down to $0.0525. Both are calculations based on published unit prices and exclude cache write costs, tool fees, taxes, free allowances, and discounts. The latter figure is not the total cost including writing to and reading from the cache, and it does not take into account the conditions under which a cache hit occurs.
Using built-in tools incurs separate charges. Web Search costs $10 per 1,000 calls in Singapore and $0.573411 in Global regions such as Tokyo. Function calling and MCP carry no additional tool fees, but tokens used for tool definitions are billed at the input rate.
Agent Operating Costs Aren't Determined by Per-Token Price Alone
Unlike one-off question answering, AI agents repeat cycles of planning, tool execution, result checking, and correction. A single job can involve dozens of model calls, each consuming thinking and input tokens. Because Qwen4-Max bills thinking tokens at the output rate, the longer the reasoning, the more the output price difference shows up in costs.
Agents also often pass their earlier work logs and tool results back to the model as they proceed. If the same input can be read from cache, it can be processed more cheaply than at the regular input price. Qwen4-Max's low cache-read price of $0.075 per million tokens may work in its favor in long-running agent processing.
Alibaba explains that improvements to the model itself have also reduced the number of tokens required. In the company's evaluation, Qwen4-Max's token consumption fell 24–35% on Work360 and 48–49% on Terminal-Bench 4.0 compared with the previous-generation Qwen3.8-Max. In addition, average cost per task on Work360 dropped about 62%, from $4.43 to $1.67. Both token reductions from model improvements and the API price cut contributed, but no breakdown by factor is provided.
However, this cost comparison is against the previous-generation Qwen3.8-Max, not a direct comparison with Claude Opus 5.5 or GPT-6.1 Sol. The number of tokens needed to finish a task varies with a model's reasoning approach and how many times it uses tools. Even with a lower unit price, a model that uses a large number of tokens can end up costing more.
For example, the price gap versus Opus 5.5 is about 4.4x on output and about 2.7x on regular input and cache reads. In a simple comparison that excludes cache write costs and tool fees, Qwen4-Max would be cheaper as long as it completes a task using no more than about 2.6 times the tokens Opus 5.5 does, on both input and output. By contrast, the gap with GPT-6.1 Sol is only about 2.2x on output and about 1.3x on input and cache reads, so differences in token consumption can more easily reverse the cost advantage.
Success rate is another important factor. On difficult coding tasks such as ProgramBench, Qwen4-Max scores less than half of Opus 5.5. If that gap translates into more failures and retries in real development work, the savings from the lower API unit price could be lost.
Determining whether Qwen4-Max suits long-running AI agent operations requires third-party comparisons under identical conditions. If evaluators such as Artificial Analysis run the same tasks across multiple models and measure success rate, input and output token counts, tool usage, and cost to completion, it would be possible to judge its actual cost advantage over GPT-6.1 Sol and Opus 5.5.
API Terms Differ Between the Full Model and the Preview
Qwen4-Max is a full model available through the pay-as-you-go API, but "qwen4-max-preview," available since November 11, remains as a time-limited preview with a separate model ID. Comparing the full model's developer documentation, the Token Plan description, and the preview documentation as of November 16 reveals the following differences.
| Item | Full model | Preview |
|---|---|---|
| Model ID | qwen4-max | qwen4-max-preview |
| Access route | Pay-as-you-go API, Token Plan Personal / Team | Token Plan only; pay-as-you-go not supported |
| Availability period | Available via pay-as-you-go API | Nov. 11, 10:00 – Nov. 25, 24:00 (UTC+8); scheduled to end afterward |
| API key and endpoint | Separate for pay-as-you-go and Token Plan | Uses Token Plan connection conditions |
| Default reasoning effort | high | xhigh |
| Max output when thinking | 262,144 tokens combined for thinking and answer | 393,216 tokens combined for thinking and answer |
The full model has thinking mode enabled by default. Reasoning effort can be selected from three levels—low, high, and xhigh—and high applies if none is specified. Because the preview's default is xhigh, switching models may change reasoning behavior and token consumption even with the same prompt.
The default thinking_budget allocated to thinking in the full model is 131,072 tokens, a setting separate from the output cap covering thinking and the final answer combined. QwenCloud's developer documentation and model list state a maximum output of 131,072 tokens without thinking mode and 262,144 tokens combined for thinking and answer in thinking mode.
On the other hand, the full model's product page shows a maximum output of 393K, and 393K in thinking mode as well. The official materials contradict each other, and the output limit that is actually guaranteed could not be confirmed. As for fine-tuning, the product page lists it as a feature, but the developer documentation explicitly states the full model does not support it.
The Token Plan is a subscription service intended for conversational use within supported AI tools. It works by consuming Credits according to usage, and Personal and Team plans are available. Purchase and use are supported only in the Singapore region, and it uses a dedicated API key and endpoint separate from the pay-as-you-go API.
The Token Plan pricing page shows a promotional price for new customers of "from $3." However, the final price is whatever appears on the checkout screen after login. This is a subscription plan fee, not the unit price of the full model's pay-as-you-go API.
A 1-Million-Token Context and an Output Budget Shared by Thinking and Answers
Qwen4-Max supports the OpenAI-compatible API and the Anthropic-compatible API, and can be used from compatible tools such as Claude Code and Cursor by changing the endpoint. The parameter for specifying reasoning effort differs by API type: reasoning_effort for OpenAI-compatible Chat Completions and DashScope, reasoning.effort for the Responses API, and output_config.effort for the Anthropic-compatible Messages API.
The handling of output token limits has also changed from earlier Qwen models. According to the API compatibility change notice, for new Qwen models from qwen4-max-preview onward, max_tokens is treated as equivalent to max_completion_tokens in OpenAI-compatible Chat Completions and DashScope. In thinking mode, the specified output budget is shared between thinking and the final answer.
For example, specifying max_completion_tokens=10000 does not allocate 10,000 tokens each to thinking and the answer; the total of both is capped at 10,000 tokens. For tasks requiring long reasoning, if thinking consumes many tokens, fewer remain for the final answer.
The full model also supports parallel tool calling and JSON Schema, and can connect to MCP through the Responses API. The context length is 1 million tokens, and according to the model list, it accepts images up to 16 megapixels each and video up to two hours and 2 GB. While this accommodates uses involving long documents and video, input capacity and output budget each need to be considered in actual processing.
For long-running agents, explicit context caching is also available. Caching repeatedly used inputs such as system prompts, tool definitions, and reference documents can reduce processing time and input costs. Cache retention can be chosen from 5 minutes or 1 hour, and separate prices are set for writing and reading.
However, which APIs can use caching differs. According to the full model's developer documentation, explicit caching for the new Qwen4 is supported on Chat Completions, Responses, and Anthropic Messages, but not on the DashScope API. Implicit caching also does not guarantee a hit every time.
Even in the Tokyo Region, Inference Does Not Necessarily Run in Japan
When using Qwen4-Max from Japan, it is necessary to distinguish between the region of the API endpoint and where inference actually runs. According to QwenCloud's official regional documentation, the region determines where requests are received and data is stored, while the inference scope specifies where processing is executed.
For the full model "qwen4-max," the inference scope that the developer documentation specifies for the Tokyo region is Global. With Global, inference may run on any available node, including those in mainland China, so simply using the Tokyo API endpoint does not confine processing to Japan.
The Tokyo region also offers an inference scope called Japan, which restricts inference processing to within Japan. However, supported models and prices differ by inference scope. In the official materials checked as of November 16, it could not be confirmed whether qwen4-max supports the Japan scope.
When connecting to the API, the region in which the API key was created must match the endpoint. Even in Tokyo, where the same Base URL is used, the workspace's API key determines the inference scope, and that scope is fixed when the workspace is created.
The Singapore pay-as-you-go API and the Token Plan also use different endpoint domains. The pay-as-you-go API uses ap-southeast-1.maas.qwencloudapi.com, while the Token Plan uses token-plan.maas.qwencloudapi.com. Because the API keys for the two cannot be used interchangeably, migrating from the existing preview to the full model requires changing both the contract and the connection settings.
When deploying Qwen4-Max in production, it is important to compare reasoning effort and output budget, cache usage conditions, and actual task completion rates on an equal footing. Because the price gap with GPT-6.1 Sol is not as large as with Opus 5.5, the number of tokens consumed and retries needed to complete the same work will determine which is cheaper.
Alibaba has also said it plans to release smaller Qwen4 models, built with the same generation of technology as Qwen4-Max, as open source. No release date has been given, but while Qwen4-Max itself is offered via API and apps, small models that can be run in a company's own environment would add an operating option, separate from API use, for firms wanting to keep AI agents running around the clock.
