Alibaba cut Qwen4-Max's cost per task by 67%, but GPT-6.1 Sol is still cheaper: $3.27 vs. $1.27
Alibaba has opened access to Qwen4-Max and promises "three times cheaper," but the price per million tokens is not the price of a solved task. According to Artificial Analysis, a task from its index costs $3.27, versus $10.02 for the previous version (−67%). However, 75% of that saving came from the price list, not from using fewer tokens. The model reasons too much, so GPT-6.1 Sol, despite its pricier tokens, finishes the same task for just $1.27. Choose Qwen4-Max for its 1M context, not for cheap coding.
AI-processed from TG: @ai_product (AI Product / Igor Akimov); edited by Hamidun News
Alibaba has opened general access to Qwen4-Max and calls it a "24/7 AI teammate" that is three times cheaper than the previous generation. The catch is that "cheaper" refers to the price per token, while you pay for a solved task. Imagine asking a model to fix a bug in a repository: if it "thinks" twice as long as a competitor, a low per-token rate won't save your bill.
That is exactly what happened. According to Artificial Analysis (an independent lab that compares models), running a task from its index costs $3.27, versus $10.
02 for Qwen3.8-Max, which is 67% less. But GPT-6.
1 Sol handles the same task for just $1.27. The practical takeaway for developers: compare models by the cost of a solved task on your own data, not by the price list.
What's inside and what it costs
Qwen4-Max is a giant model with 2.4 trillion parameters, of which 71 billion are active for each response. It is data-center-scale. There will be no open weights: Alibaba will release them only for the smaller models in the lineup. The context window is 1M tokens. The model accepts text, images and video as input, and outputs text only.
Pricing: $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 for cached input. That is 25% cheaper than Qwen3.8-Max and noticeably below the market leaders.
What Alibaba showed itself, and what independent tests show
The company built the vendor chart itself, so read it with that in mind:
- DeepSWE v1.1 (software development tasks): 75.7 vs. 74.2 for Opus 5.5 and 75.5 for Kimi K3.1. Five models landed within a point and a half of each other. That is noise, not leadership.
- HLAB-Vals (legal tasks): 21.3 vs. 9.2 for Opus 5.5. Here Qwen4-Max really did pull ahead.
- Terminal-Bench 4.0 (agent work in a terminal): 48.5 vs. 67.2 for Kimi K3.1 and 59.6 for Opus 5.5. Weak.
- ProgramBench: 31.5 vs. 65.0 for Opus 5.5. Also a failure.
- Chartography (chart reading) and Video-Index (video): 6–17 points behind the American models.
The independent picture is more modest. In the Artificial Analysis Intelligence Index, the model scored 43 points vs. 34 for Qwen3.8-Max. Ahead of it are Fable 5.2 (65), Gemini 4 Carbon (60), Opus 5.5 (59), GPT-6 Astra (57), GPT-6.1 Sol (55) and even the older Kimi K3.1 Pro (46). On the Vals Index, Qwen4-Max scores 63.20%, which puts it 11th out of 38 and makes it the best result among Chinese models. On Code Arena WebDev it ranks sixth with 1750 Elo, but the rating is still preliminary.
Why "three times cheaper" needs caveats
Of the 67% saving per task, 75% came from the price cut and only 25% from an actual reduction in token consumption. The model is still wordy: a single run of the index consumes 127 million output tokens, against a median of 78.6 million.
The most telling comparison is with GPT-6.1 Sol. At a price of $2 / $10, Sol costs $1.27 per task at Artificial Analysis and $3.61 per test at Vals. Qwen4-Max costs $3.27 and $9.33, respectively. Sol also scores slightly higher at Vals: 63.94% vs. 63.20%. So the model with the pricier token ends up both cheaper to run and slightly more accurate. A cheap token does not mean a cheap task: the extra reasoning eats the discount.
Who it's for
For agentic coding in the terminal, Qwen4-Max is not an option yet: Kimi K3.1, Sol, Opus 5.5 and Gemini 4 Carbon are still better here. Don't count on deep work with charts and video either, where the gap is noticeable.
The model does have a niche, though: tasks that need a 1M-token context without big costs, such as analyzing large documents, contracts or a codebase. Legal scenarios look like a separate strong point, but that result comes from the vendor chart, so verify it on your own data before adopting it.
Bottom line: this is an honest generational update, with 9 more index points and a price three times lower than the previous generation, but no magic. Before switching production over, run your own set of tasks and calculate the total cost of a solved task together with the share of correct answers.
- Nov 12, 2026Google opens preorders for Gemini glasses starting at $395 and blocks covert recording
- Nov 12, 2026Google brings Gemini Intelligence to Android Auto — for now, Pixel only
- Nov 12, 2026Lawsuit: Schoolgirl Spent Three Months Talking to Gemini About Suicide While Monitoring Stayed Silent for Weeks
- Nov 12, 2026Google and Adobe bring their products closer: Acrobat is coming to Gmail, Drive and Gemini
Key Questions Answered
What's inside and what it costs?
Qwen4-Max is a giant model with 2.4 trillion parameters, of which 71 billion are active for each response. It is data-center-scale. There will be no open weights: Alibaba will release them only for the smaller models in the lineup. The context window is 1M tokens. The model accepts text, images and video as input, and outputs text only. Pricing: $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 for cached input. That is 25% cheaper than Qwen3.8-Max and noticeably below the market leaders.
What Alibaba showed itself, and what independent tests show?
The company built the vendor chart itself, so read it with that in mind: DeepSWE v1.1 (software development tasks): 75.7 vs. 74.2 for Opus 5.5 and 75.5 for Kimi K3.1. Five models landed within a point and a half of each other. That is noise, not leadership. HLAB-Vals (legal tasks): 21.3 vs. 9.2 for Opus 5.5. Here Qwen4-Max really did pull ahead. Terminal-Bench 4.0 (agent work in a terminal): 48.5 vs. 67.2 for Kimi K3.1 and 59.6 for Opus 5.5. Weak. ProgramBench: 31.5 vs. 65.0 for Opus 5.5. Also a failure.
Why "three times cheaper" needs caveats?
Of the 67% saving per task, 75% came from the price cut and only 25% from an actual reduction in token consumption. The model is still wordy: a single run of the index consumes 127 million output tokens, against a median of 78.6 million. The most telling comparison is with GPT-6.1 Sol. At a price of $2 / $10, Sol costs $1.27 per task at Artificial Analysis and $3.61 per test at Vals. Qwen4-Max costs $3.27 and $9.33, respectively. Sol also scores slightly higher at Vals: 63.94% vs. 63.20%.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.