The PlaybookAI Models

Alibaba Launched Qwen4-Max at $1.67 a Task.

The price per token fell. The price per task is the number to route on.

John TanNovember 16, 20264 min read

Alibaba's Qwen team released Qwen4-Max on 16 November. It is the first model on the new Qwen4 architecture: a 2.4-trillion-parameter backbone with 71 billion parameters active per token, a one-million-token context window and native vision.

The number Qwen leads with is not a benchmark score. It is a bill. On Work360, the average cost per task drops from $4.43 on Qwen3.8-Max to $1.67 on Qwen4-Max, which Qwen puts at 62% below the old model.

Work360 is Qwen's own benchmark. The team built it in-house from a taxonomy of 26 domains, 182 subdomains and 721 specialised fields, and Qwen4-Max scores 52.69 on it against 47.74 for Qwen3.8-Max. That makes the $1.67 a real measurement on someone else's tasks. Yours are the ones that count.

Two scatter charts from Qwen. Left, Work360: Qwen4-Max's Low, High and Xhigh points sit between about 90K and 127K tokens per task and score higher than Qwen3.8-Max-0902's points, which sit between about 135K and 166K tokens. Right, Terminal-Bench 4.0: Qwen4-Max's points sit between about 120K and 152K tokens per task, Qwen3.8-Max-0902's between about 203K and 290K.
Qwen's chart · Nov 16Up is a higher score, left is fewer tokens per task. Every Qwen4-Max point sits up and to the left of the old model.

Token-efficiency chart from Qwen's Qwen4-Max launch post. Qwen ran both benchmarks and chose the comparison.

The token got cheaper. The task got cheaper faster.

Two things moved at once. The list price came down: on OpenRouter, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output, down from $2.00 and $6.00 for Qwen3.8-Max-0902. Cached input fell furthest, to $0.075 from $0.25.

And the model uses fewer tokens to finish. Qwen reports token use on Work360 down about 24% at Xhigh effort and 35% at High. On Terminal-Bench 4.0 the cuts are about 48% and 49%.

A per-token price sees only the first of those. The cost of a finished task sees both, which is why the price per token fell less than the price per task.

Figure 1. Two prices
The price per token fell less than the price per task.
PRICE PER MILLION INPUT TOKENS Qwen3.8-Max $2.00 Qwen4-Max $1.50 AVERAGE COST PER WORK360 TASK Qwen3.8-Max $4.43 Qwen4-Max $1.67 PRICE PER 1M INPUT TOKENS Qwen3.8-Max $2.00 Qwen4-Max $1.50 COST PER WORK360 TASK Qwen3.8-Max $4.43 Qwen4-Max $1.67
Each pair is drawn against the old model's value. Input prices from OpenRouter's model list, read 16 November 2026. Work360 costs as Qwen published them on 16 November 2026, on Qwen's in-house benchmark. Figure by nativefirst.

Whose tasks?

Artificial Analysis, which runs the same nine evaluations across every model in its index, puts Qwen4-Max at 43 on its Intelligence Index and $3.27 per task. That is a different set of tasks from Work360, so the two dollar figures do not contradict each other. They answer different questions.

On Artificial Analysis's tasks Qwen4-Max costs less than Kimi K3.1 Pro and Opus 5.5, at $3.68 and $10.04. GPT-6.1 Sol at Max costs less per task and scores higher: $1.27 and 55 points. Artificial Analysis also calls Qwen4-Max very verbose. Running the index it generated 127 million output tokens, more than the median of 65.4 million. A cheap token used freely is not a cheap task.

Figure 2. An outside bill
On Artificial Analysis's tasks, Qwen4-Max sits mid-table on cost.
COST PER TASK INDEX GPT-6.1 Sol $1.27 55 Qwen4-Max $3.27 43 Kimi K3.1 Pro $3.68 46 Opus 5.5 $10.04 59 COST PER TASK · INDEX GPT-6.1 Sol · 55 $1.27 Qwen4-Max · 43 $3.27 Kimi K3.1 Pro · 46 $3.68 Opus 5.5 · 59 $10.04
Artificial Analysis, weighted average cost per Intelligence Index task and Intelligence Index score, read 16 November 2026. GPT-6.1 Sol and Opus 5.5 at Max. Bars from zero. Figure by nativefirst.

Where the cheap task is

Qwen's own comparison chart says the same thing in another way. On Agents' Last Exam, the long-horizon work test, Qwen4-Max scores 46.7 against 43.4 for Opus 5.5. On Terminal-Bench 4.0, the coding agent's test, it scores 48.5, behind Opus 5.5 and Kimi K3.1 at 59.6 and 67.2.

Four bar-chart panels from Qwen's launch post: CodeArena, Terminal-Bench 4.0, Agents' Last Exam v1.1 and JobBench. Qwen4-Max's blue bar is tallest on Agents' Last Exam and shorter than the Anthropic and Kimi bars on Terminal-Bench 4.0 and shorter than the Anthropic bar on CodeArena and JobBench.
Qwen's chart, cropped · Nov 16Bars left to right, from Qwen's legend: Qwen4-Max, Qwen3.8-Max-0902, Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, Kimi K3.1. Agents' Last Exam shows five.

Four of sixteen panels from the comparison chart in Qwen's launch post. Qwen ran its own models; its footnotes say some rival scores are cited from public leaderboards.

Read together, the pitch is narrower than "your new teammate". Qwen4-Max is a candidate for long office and research jobs, the kind Qwen shows off with an 86-page due-diligence report and a machined car part whose manual programming time, Qwen says, fell from about 475 minutes to 24. It is not the obvious pick to drive a terminal.

Test it this week

If your company runs agent work in bulk, the cost of a finished task is the number to route on, and only your tasks can produce it. Qwen's $1.67 is on Qwen's tasks. Artificial Analysis's $3.27 is on theirs.

Pull 20 finished jobs. Take them from your logs, from the work you would hand an agent tomorrow: reports, research, document review. Keep the inputs and the accepted output.

Run them three ways. Qwen4-Max at High, Qwen4-Max at Xhigh, and the model you use now. It is live on OpenRouter, so this is a model-name change, not a contract.

Divide the bill by the passes. Total spend over jobs that passed your own review. Tokens per job and price per token are inputs to that number, not substitutes for it. Fable 5.2 showed the opposite case: cheaper per token, dearer per job.

Keep coding agents where they are until a terminal test says otherwise. Google's answer to the same problem was to split planning from typing across two models, and Carbon's price doubles on 1 December. A per-task sheet makes every one of those choices a lookup.

Price the task.

Portrait of John Tan
John Tan ↗

Founder of nativefirst.ai. Runs everything on agents and writes down what works.

Sources
1Qwen Team, Qwen4-Max: Your New Teammate Has Arrived, 16 November 2026: the architecture, Work360 and its cost per task, token-use reductions, the comparison chart and the token-efficiency chart.. qwen.ai
2OpenRouter, models API: Qwen4-Max and Qwen3.8-Max-0902 per-token prices, read 16 November 2026.. openrouter.ai
3Artificial Analysis, Qwen4 Max: Intelligence Index score, cost per Intelligence Index task and output tokens, read 16 November 2026.. artificialanalysis.ai
4Artificial Analysis, GPT-6.1 Sol, Claude Opus 5.5 and Kimi K3.1 Pro model pages: cost per Intelligence Index task and index scores, read 16 November 2026.. artificialanalysis.ai

Read next

AI is moving really fast. Keep up.

Read the Playbook