Title card reading Qwen4-Max cuts cost per task 67%, with a headline figure of $3.27 against $10.02 per Intelligence Index task and a note that the split is roughly half rate card and half token discipline.
Engineering & Research

Qwen4-Max Cuts Cost Per Task 67%: Where the Saving Actually Comes From

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

A benchmark bill is a strange artefact: it is the only public document that prices a model's habits. Artificial Analysis publishes what it cost to run Qwen4-Max through its Intelligence Index — $3.27 per task — and it publishes the rate card it used to get there. Divide one by the other and you recover something the lab never states outright: how many tokens Qwen4-Max actually spends thinking. Do the same with Qwen3.8-Max, the model it replaces, and the 67% reduction in cost per task splits cleanly into two numbers that most coverage of the 16 November launch runs together. Roughly half of the saving is a price cut. The other half is Qwen4-Max talking less. And the line item that now dominates its bill is not the one the headline price suggests.

The vendor's own summary is accurate but compresses the interesting part. Qwen4-Max, Alibaba states on its launch page, uses approximately 24% fewer tokens than Qwen3.8-Max at Xhigh effort and 35% fewer at High on Work360, and 48% and 49% fewer on Terminal-Bench 4.0; the average cost per task on Work360 falls from $4.43 to $1.67, a 62% reduction. Those are company numbers on company benchmarks. The independent arithmetic below reaches 67% on a different benchmark through a different route, which is the useful kind of agreement: two methods, same direction, similar magnitude.

What the independent scoreboard actually says

On Artificial Analysis, Qwen4-Max scores 43 on the Intelligence Index against 34 for Qwen3.8-Max, a 27% gain in measured capability. The same page prices the two models at $1.50 per million input tokens and $4.50 per million output, against $2.00 and $6.00 — a flat 25% off both directions. Cached input is where the discount is dramatic: $0.075 per million against $0.25, a 70% cut, and the cache-hit discount rate moves from 87.5% to 95%.

Cost per Intelligence Index task, weighted across all nine evaluations in the index, lands at $3.27 against $10.02. Divide cost by score and the gap widens: Qwen4-Max delivers a point of measured capability for under eight cents, Qwen3.8-Max for about thirty. That is the number to hold onto if the only question is which model to move a workload to — capability per dollar improved by roughly three quarters, not by the 67% that cost per task alone suggests.

Inverting the bill to recover the token counts

Artificial Analysis breaks cost per task into components: input at full rate, cache write, cache read, and output, with output further split into reasoning and answer tokens. It also publishes the prices. Divide each component by its price and you have the tokens. The method validates itself on one line — the output component divided by $4.50 comes to 104,940 tokens per task, and the page separately reports 104,939.8 output tokens per task for Qwen4-Max. Exact agreement to the decimal, on a line that was computed two different ways.

Every other line inherits that confidence, with one assumption stated openly: that cache writes are billed at the full input rate, which is what the arithmetic across both models implies rather than something either vendor documents explicitly.

What falls out, per Intelligence Index task:

• Output tokens — Qwen4-Max 104,940 against Qwen3.8-Max 170,794, a 39% reduction.

• Reasoning tokens — 68,914 against 119,698, down 42%. The thinking budget shrank far more than the answer budget.

• Answer tokens — 36,025 against 51,096, down 30%. Qwen4-Max now spends 66% of its output on reasoning rather than the 70% its predecessor spent; it is not merely shorter, it is differently proportioned.

• Cache-read tokens — about 14.9 million against 22.7 million, down 34%. These are the replayed context tokens that dominate any long agentic run.

• Cache-write tokens — about 1.11 million against 1.66 million, down 33%.

Two things stand out. First, the token reduction is not concentrated in the visible answer, where a user would notice verbosity, but in reasoning and in replayed context, where nobody is watching. Second, the per-task token volumes are enormous in absolute terms — fourteen million cache reads per task — which tells you what the Intelligence Index actually measures: long-horizon work with a large persistent context, not single-turn question answering.

Two-column scoreboard comparing Qwen4-Max and Qwen3.8-Max on cost per task of 3.27 dollars against 10.02 dollars, output tokens per task of 104,940 against 170,794, cache-read tokens of 14.9 million against 22.7 million, and Intelligence Index scores of 43 against 34.

Split the 67% into price and behaviour

Cost is price multiplied by volume, so the reduction can be decomposed cleanly. Bill Qwen3.8-Max's token volumes at Qwen4-Max's rates and a task costs $4.97 instead of $10.02 — the price change alone accounts for a 50% cut. Apply Qwen4-Max's behaviour as well and it lands at $3.27, a further 34% off what remained. Multiplied together, 0.496 times 0.658 gives 0.326: the observed 67% reduction, fully accounted for, with nothing left over.

Which half deserves the credit depends on the order you charge it in, and it is worth saying so rather than picking the flattering number. In percentage terms, price is worth about 50% and token discipline about 34%. In dollars — $6.76 saved per task in total, $5.06 from rates and $1.70 from volume — price takes three quarters. Both framings are correct; they answer different questions. If you are forecasting next quarter's spend, the rate card is the larger and more certain lever. If you are asking whether the model is genuinely more economical, the 34% that comes from spending fewer tokens is the more durable one, because a price cut can be reversed and an architecture cannot.

The direction of the vendor's own numbers corroborates the split without being asked to. Alibaba attributes the reduction to both factors in the same breath, and its 62% on Work360 sits close to the 67% measured independently on the Intelligence Index — two harnesses, two sets of tasks, one conclusion.

The line item that quietly became the biggest one

Here is where the headline price misleads. Before, cache reads were 57% of the bill and cache writes 33%. After, cache writes are 51% and cache reads 34%. The model got cheaper on every line, but the discount was lopsided — 70% off cache reads, 25% off cache writes — and it flipped the ranking of the two largest costs.

Cache writes are billed at the input rate, $1.50 per million, twenty times the cache-read rate of $0.075. A workload that rewrites its context more often than it replays it now pays for that habit in the most expensive currency on the card. This is the practical consequence of the pricing change that the launch coverage mostly missed, and it argues for the unglamorous work: stable prompt prefixes, long-lived sessions, fewer full-context rebuilds. On Qwen3.8-Max those habits mattered less because replaying was expensive enough to dominate the bill anyway.

The lesson generalises past this model. When a provider cuts one token class far harder than others, the optimal prompt architecture changes with it, and the bill you get is a lagging indicator of the architecture you built for the previous card.

Alibaba Cloud Model Studio documentation page for qwen4-max, showing the model name as the page title, a release date of 11-16, the model invocation ID qwen4-max, and the listed regions including Singapore and US (Virginia).

The efficiency is relative, and the absolute number is unflattering

A 39% reduction in output tokens sounds like a lean model. Set it against the field and it is a leaner version of a very talkative one. Across the whole index run Qwen4-Max generated 127 million output tokens; the median model generates 78.6 million. Qwen3.8-Max generated 150 million. Both Qwen models sit far to the verbose side, and Qwen4-Max merely moved from 1.9 times the median to 1.6 times it.

That is not a contradiction of the efficiency claim — it is the correct reading of it. The claim is that Qwen4-Max gets a better result than its predecessor while spending fewer tokens, and both halves hold. It is not a claim that the model is quiet, and any budget model built on the assumption that a high-scoring model must be a concise one will be wrong by a wide margin. If output volume drives your cost, moving to Qwen4-Max reduces the number and does not change the category.

The vendor's pre-training figures point the same way at a different layer. On in-house BPB evaluations, Qwen4-Max-Base is 5–13% better than Qwen3.8-Max-Base in every domain while activating 25% fewer parameters — the architecture is doing more work per unit of compute, which is consistent with a model that needs fewer tokens to reach a conclusion. Whether the 2.4-trillion-parameter backbone, 71 billion activated parameters and the extra 328 billion N-gram embedding parameters are the reason is a vendor explanation, not a measurement anyone outside Alibaba has reproduced.

What is missing from the cost picture

Latency, for one. The Intelligence Index card lists time per task as not applicable for Qwen4-Max; for Qwen3.8-Max it is 48.1 minutes against a 13.3-minute median, one of the slowest results measured. A model that is 67% cheaper per task and takes an unknown amount of wall-clock time per task has not been fully compared to anything. Cheap tokens you have to wait for are a different product.

Provider breadth, for another. The price behind every figure above is a single first-party rate from Alibaba Cloud, and the model currently appears with one host. There is no second source to arbitrage, so the rate card is the rate. And Qwen3.8-Max is now marked deprecated in favour of Qwen4-Max and benchmarked every eighteen hours rather than every six, which means the comparison baseline refreshes more slowly than the model being measured against it.

OrcaRouter model page for qwen/qwen3.8-max, showing the model id, input and output prices of $2.00 and $6.00 per million tokens, a 1,000,000-token context window, and the capability badges for vision, tools, JSON and reasoning.

Turning the arithmetic into a budget you control

Two levers follow from the decomposition, and they are independent. The first is price: a rate card you cannot negotiate is still a rate card you can route around, which is the case for paying list price and nothing above it. OrcaRouter passes provider list pricing through at 0% markup, so a vendor cut of the kind Alibaba made on 16 November reaches your bill the same day rather than at the next contract renewal — and the size of this particular cut, particularly the 70% on cached input, is exactly the sort of change that a pass-through arrangement makes visible instead of quietly absorbing.

The second lever is volume, and it is the one the model cannot make for you. The routing DSL composes several models into a single call, so the bulk of a pipeline can run on a cheap model while the hard tail — the tasks that genuinely need a 2.4-trillion-parameter MoE — is sent to one that has it. That split is how you capture the 34% behaviour dividend without paying frontier rates for work that never needed them, and it is the natural response to a workload where fourteen million context tokens per task is normal.

Being precise about what we do and do not offer: Qwen4-Max is not in our catalogue as of 16 November 2026. The Qwen model we route today is qwen/qwen3.8-max, the predecessor this article uses as its baseline. For a team wanting to test whether the token reduction holds on its own traffic before committing, automatic failover is the low-risk way in — route a slice, keep the fallback live, and compare real traces rather than benchmark tasks.

What would change this reading

Artificial Analysis publishes aggregate token counts per model, and this piece infers per-task volumes from its cost components. If the harness changed between the two runs — different task mixes, a different weighting — part of the 39% output reduction could be harness and not model. The exact reconciliation on the output line argues against that, but it does not rule it out, and anyone with traffic of their own should treat the derived numbers as a hypothesis to test rather than a fact to plan around.

The vendor's Work360 figures carry a different caveat, which Alibaba states itself: two of the checkpoints in the BPB comparison are post-trained rather than base, which weakens a like-for-like framing. The 62% cost reduction on Work360 is not affected by that, since it compares shipped models on a fixed task set, but the pre-training claims should be read with it in mind.

What survives all of that is the shape of the answer rather than its precision. A launch framed as a price cut is also, and roughly equally, a change in how much the model spends to reach an answer — and the second half is the part a competitor cannot copy by adjusting a rate card. That is the number to watch when the smaller Qwen4 models ship, because if the token discipline is architectural it will travel with them, and if it was a property of this one checkpoint it will not.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily