Independent AI intelligence Two editions daily · ET
Fervor AI

Analysis · November 16, 2026 · concept

Qwen4-Maxfrontier-modelsagent-infrastructure

Qwen4-Max Is Cheaper Per Token. Tokens Per Task Decide Your Bill.

Alibaba sold its new flagship on the cost of a finished job. Here is how to check that math on your own work before you switch.

Model prices have always been quoted per token, and almost nobody pays per token. You pay per job. A bug fixed, a report drafted, a ticket closed. Two models with the same rate card can send you very different bills, because one of them thinks for 3,000 tokens before answering and the other thinks for 30,000.

On November 16, a frontier lab finally put that on the front of the box.

Alibaba made Qwen4-Max, the first model of its Qwen4 generation, generally available after a preview that opened November 11. The headline was not the rate card. It was cost per task: on Alibaba's own benchmark, the average task got roughly 60% cheaper, because the price went down and the model spends fewer tokens getting to an answer.

That framing is right. It is also only half checked. The price cut is a fact you can read on a pricing page. The token savings are a claim about how the model behaves, measured by the company selling it, on tasks it picked. The useful move is to split the claim in two, believe the first half, and measure the second half yourself.

What Alibaba actually shipped

Qwen4-Max is a hosted model, so you call it over an API and never download it. Under the hood it is a mixture-of-experts model (a big network made of many smaller expert networks, where only a few wake up for each token), with a 2.4-trillion-parameter backbone and about 71 billion parameters active per token, according to MarkTechPost's read of the launch blog. It reads text, images and video.

The numbers that change your decision are in QwenCloud's model docs:

  • A context window of 1,000,000 tokens.
  • A combined cap of 262,144 tokens for thinking plus the final answer.
  • Thinking is on by default, at reasoning_effort=high, with a default thinking budget of 131,072 tokens.
  • You can pick low, high or xhigh effort, or set enable_thinking=false in Chat Completions to turn thinking off.
  • It speaks the OpenAI and Anthropic API formats, so most agent frameworks can point at it with a config change.

And the pricing page for Singapore: $1.50 per million input tokens and $4.50 per million output tokens, with no tiers for long inputs. The Beijing and Global regions list $1.413 and $4.24. The previous flagship, Qwen3.8-Max, was $2 and $6, so in Singapore that is a straight 25% cut on both sides. In the Global regions the cut is smaller, about 14%, per OrcaRouter's rate-card comparison, because the old model was already discounted there.

One line on that pricing page matters more than all the rest: thinking tokens count as output tokens and bill at the output rate.

The per-task claim, taken apart

Here is the claim as reported. On Work360, Alibaba's in-house suite covering 26 top-level work domains, average cost per task fell from RMB 26.58 to RMB 11.16, about 58%, per TechNode. MarkTechPost's figures are in dollars, $4.43 to $1.67, a 62% cut, and it reports Alibaba's token-use drops of about 24% at xhigh effort and 35% at high effort on the same suite.

Two things multiply to get that number:

  1. A lower price per token. Verified. It is on the rate card.
  2. Fewer tokens per task. Unverified outside Alibaba. Nobody has reproduced it yet.

That second factor is doing a lot of the work, and it is also the factor most sensitive to what you ask the model to do. A model that saves 35% of its thinking on Alibaba's office-style tasks might save 10% on your gnarly refactor, or spend more. The benchmark chart Alibaba published is a hint about where it might not: on ProgramBench, a coding test, it scores 31.5 against 65.0 for Claude Opus 5.5, by MarkTechPost's reading of Alibaba's own chart.

None of this makes the claim false. It makes it a starting hypothesis.

Why thinking tokens are the part of your bill you control

A reasoning model writes a private scratchpad before it answers. Those scratchpad tokens are invisible in the final reply, and on Qwen4-Max they cost the same as the answer: $4.50 per million.

So your effort setting is a budget dial. A simple way to see it:

def cost_per_task(input_tokens, output_tokens,
                  in_price=1.50, out_price=4.50):
    """Qwen4-Max Singapore list prices, USD per million tokens.
    output_tokens = thinking + final answer."""
    return (input_tokens * in_price + output_tokens * out_price) / 1_000_000

# Illustrative only: a task that reads 40k tokens and writes 12k
print(cost_per_task(40_000, 12_000))   # 0.114
# Same task, but the model thinks twice as long
print(cost_per_task(40_000, 24_000))   # 0.168

Same model, same task, same rate card. The second run costs about 47% more because the scratchpad doubled. That is why a vendor can cut its per-task cost without touching price: teach the model to stop thinking sooner.

It is also why the default matters. Qwen4-Max ships at high. If most of your traffic is short tool calls (look up an order, format a JSON blob, pick the next step), you may be paying for deep thought you never needed.

The flip side is real too. Cutting effort on hard tasks can make the model fail more, and a failed task that you rerun costs you twice. That is the trap in comparing models by price alone.

The metric that settles it: cost per successful task

Price per token tells you what one token costs. Cost per successful task tells you what one finished job costs, after you count the retries.

The formula is plain:

total spend on a batch of tasks ÷ number of tasks that actually succeeded

If a cheap model finishes 6 of 10 tasks for $1.20, you paid $0.20 per success. If a pricier model finishes 9 of 10 for $1.62, you paid $0.18. The pricier model was cheaper. Per-token comparisons hide this completely.

Tencent's Hy4 flash, which launched the same day, makes the point from the other direction. It lists at about $0.070 per million input tokens, roughly 21 times cheaper than Qwen4-Max on uncached input. Tencent's evidence is a blind test of its own staff on office tasks, and Tencent itself says to keep its bigger model for long software work. A cheap token that fails your task is not cheap.

Put this into practice

You can run this in an afternoon. You need an API key, ten tasks and a spreadsheet.

1. Pick ten real tasks. Not benchmark questions. Pull them from last week's agent logs: the bug your coding agent fixed, the summary it wrote, the three-step lookup it ran. Write down what "done" means for each one, so you can mark pass or fail without arguing with yourself later.

2. Run each task at two or three effort settings. Start with low and high. Add xhigh only for the tasks that fail at high. The parameter names are in the thinking docs; the model ID is qwen4-max. Run your current model on the same ten tasks as the control.

3. Log the usage block from every response. Each API response reports input and output token counts. Output includes the thinking tokens. Put them in the spreadsheet next to pass or fail, and note which region you called, since prices differ by region.

4. Compute cost per successful task for each model and setting. Use the formula above. Expect at least one surprise: usually a lower effort setting that passes everything, or a "cheap" setting that fails often enough to lose.

5. Cache what repeats. Qwen4-Max supports an explicit cache (you mark the parts of a prompt that repeat, and the provider stores them so later calls reuse them) for shared prefixes like system prompts and tool definitions, with a 5-minute or 1-hour lifetime. Cache reads cost $0.075 per million tokens against $1.50 for fresh, uncached input, per the pricing page. Agents resend the same system prompt and tool list on every turn, so this adds up fast.

6. Check your model IDs. Secondary coverage reports that the preview IDs (qwen4-max-preview and qwen4.0-max-preview) retire at 24:00 on November 25, China time (UTC+8). If you joined the preview, change the string now.

What to expect: on short, structured tasks, low effort often matches high at a fraction of the output tokens. On long reasoning or code, the gap opens and the cheaper setting starts failing. Your spreadsheet will show you where that line sits for your work, which no vendor chart can.

Honest limitations

Every per-task number in this launch is Alibaba's. Work360 is an in-house suite. The token-efficiency claim has not been reproduced outside the company, and the two secondary reports even disagree on the size of the cut (58% in yuan, 62% in dollars).

No open weights. You cannot self-host it, quantize it or run it on your own hardware. Alibaba says it plans to open-source smaller Qwen4 models, with no date.

The docs contradict themselves in places. The model docs say fine-tuning is not supported and list batch inference (queueing many requests to run later at a discount) as unsupported, while the model page shows fine-tuning and batch tiles under features. Trust the developer docs, and confirm before you plan around either.

Explicit caching is not free or universal. Cache writes cost more than plain input ($1.875 per million for the 5-minute cache, $3 for 1 hour), so a prefix you only use once loses money. DashScope, Alibaba's own native API format, does not support the new explicit cache, and batch and cache discounts cannot be combined on the same request.

Coding may not be its strength. On Alibaba's own chart, Qwen4-Max trails Claude Opus 5.5 by a wide margin on ProgramBench. If your agents mostly write code, test that first.

The bill was always per task

Per-token pricing was a convenience for vendors. It made comparisons look easy and kept the expensive part, how long a model thinks, out of view. Qwen4-Max's launch drags that part into the light, and that is a good thing even if the 60% figure shrinks under real workloads.

So take the half that is proven: the rate card is lower, by 25% in Singapore and about 14% in the Global regions. Treat the other half as a question. Ten tasks, two effort settings, one spreadsheet. Then switch, or don't, on a number you measured yourself.

Sources: QwenCloud Qwen4-Max docs, QwenCloud pricing, QwenCloud model page, TechNode, November 16, 2026, MarkTechPost, November 16, 2026, OrcaRouter daily brief, November 16, 2026. Prices and limits as of November 16, 2026. The qwen.ai launch blog could not be read directly; benchmark figures come from MarkTechPost's reading of it.