A generated title card reading Qwen 4 vs Muse Wonder, with the subtitle The cheaper model is the one nobody has timed, and two cards: Cost per index task showing $3.27 above $6.95, and Measured speed showing not published above 118 tokens per second.
Guides & Insights

Qwen 4 vs Muse Wonder: The Cheaper Model Is the One Nobody Has Timed

Author

jinhao.song

Date Published

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

Muse Wonder will cost you roughly $6.95 to finish one Intelligence Index task. Qwen4-Max will cost you $3.27. That is the entire economic case for Alibaba's new flagship against Meta's, and it comes with one large caveat: the $3.27 belongs to a model that has no published output speed, no measured time per task, and no timing of any kind on the board that produced the figure. Muse Wonder has all three — 118 tokens per second, 13.5 minutes per index task, 19.64 seconds to first answer token. So the honest statement is that Qwen 4's flagship is about half the price and an unknown amount of the speed, and only one of those two differences you can plan around today. Muse Wonder was released on 12 October 2026; Qwen4-Max on 16 November 2026. Both dates are context, not news.

The two profiles, side by side and unsmoothed

Muse Wonder is Meta's frontier reasoning model, built for agentic work with an unusually wide input surface: text, images, speech and video in, text out, inside a one-million-token context. It is a proprietary API model at $4.00 per million input tokens and $20.00 per million output, with a $2.94 blended rate and a 95 per cent cache discount. Meta also publishes a higher-effort variant that scores 41 rather than 44 and runs faster in wall-clock terms — 9.5 minutes per index task against 13.5 — which is a useful reminder that "max effort" is not automatically the right setting.

Qwen4-Max is Alibaba's fourth-generation flagship: a 2.4-trillion-parameter sparse mixture-of-experts with 71 billion active per token, native vision, video input, a one-million-token context, and a 262,144-token budget shared between thinking and answer. It prices at $1.50 in and $4.50 out per million, with cache reads at $0.075.

• Score — Qwen4-Max 43 (rank 38 of 81) vs Muse Wonder 44 (rank 35 of 81)

• Input price — Qwen4-Max $1.50 vs Muse Wonder $4.00 per 1M

• Output price — Qwen4-Max $4.50 vs Muse Wonder $20.00 per 1M

• Cost per index task — Qwen4-Max $3.27 vs Muse Wonder $6.95

• Output speed — Qwen4-Max not published vs Muse Wonder 118 tokens per second

• Time per index task — Qwen4-Max not published vs Muse Wonder 13.5 minutes

• Input surface — Qwen4-Max text, image, video vs Muse Wonder text, image, speech, video

• Weights — neither publishes any

Price per token and price per answer disagree by a factor of two

The headline ratio is 4.4× on output and 2.7× on input. The per-task ratio is 2.1×. The gap between those numbers is the amount of work each model does to produce an answer, and it runs the opposite way from what the sticker suggests: Muse Wonder is spending more tokens per task, not fewer, which is why a $20 output price turns into a $6.95 bill rather than a $40 one.

Work it through on a concrete job. A 50,000-token document review producing a 5,000-token memo, no cache: Qwen4-Max bills $0.075 of input and $0.0225 of output, under ten cents. Muse Wonder bills $0.20 and $0.10, thirty cents. Three to one, and both are trivial sums — which is the point. At small volume neither price matters and the score difference of one index point is irrelevant next to the speed question.

Scale it. Ten thousand of those reviews a month: Qwen4-Max $975 against Muse Wonder $3,000. Now the price matters, and now the missing stopwatch matters more, because a job that costs a third as much but takes twice as long may not fit a nightly window at all.

That is the shape of the decision. This is not a comparison where one model is better; it is one where the cheaper option has an unmeasured dimension that happens to be the one production systems care about most.

What "no published speed" actually costs you

An unmeasured model is not a slow model, and it would be wrong to read it that way. It is a model whose latency you have to establish yourself before you can put it behind anything with a timeout. That has three practical consequences, in ascending order of annoyance.

First, you cannot size autoscaling from documentation. You have to run a pilot, measure p50 and p95 time-to-first-token at your own prompt lengths, and then size. Kimi K3.1 Pro's 30 tokens per second and 39.7 minutes per task were measured and published; Qwen4-Max's figure will be too, eventually, but it is not there now.

Second, reasoning models degrade unpredictably under load in ways that a token-per-second number only partly captures. Muse Wonder's 19.64-second time to first answer token at max effort is a measured, reproducible figure. A model with no such figure has no baseline against which to notice that a Tuesday is twice as slow as a Monday.

Third — and this is the one that bites — a routing layer can fail over on latency only if it knows what normal looks like. If you put an untimed model into a pool with a timed one, the pool's behaviour is defined by the slower, unknown member. Failover makes that survivable: a request that stalls can be retried against a second model without the client knowing. It does not make it measurable, and it does not substitute for knowing your own numbers.

The input surface is Muse Wonder's real differentiator

Speech is not a checkbox. Muse Wonder accepts audio as a first-class input, alongside text, image and video, and returns text. Qwen4-Max takes text, image and video. If your pipeline ingests call recordings, meeting audio, or broadcast media, that difference decides the model before price is discussed — transcribing to text first is a separate model, a separate bill, and a separate failure mode that loses prosody and speaker overlap.

Video is the other axis where both compete and the details matter. Both accept it; neither publishes a frame-rate, duration, or resolution limit in the material we read, which means a video workload is a pilot on both sides rather than a specification comparison.

Where Qwen4-Max wins on input is cost at volume. A million tokens of video-derived context at $1.50 costs $1.50 on Qwen4-Max and $4.00 on Muse Wonder. Long media analysis is exactly the workload where the 2.7× input gap compounds, and exactly where Qwen4-Max's price advantage is most defensible.

A generated two-column scoreboard for Qwen4-Max and Muse Wonder: Artificial Analysis Index 43 against 44, cost per task $3.27 against $6.95, input $1.50 against $4.00, output $4.50 against $20.00, speed not published against 118 tokens per second, and audio input no against yes.

One point of index score, and what it is worth

44 against 43 on the same Artificial Analysis Intelligence Index, read the same day. That is not a meaningful difference in capability; it is within the noise of how a single aggregate index responds to a handful of hard questions. Anyone building a case for Muse Wonder on that point is building it on nothing.

What is meaningful is that Muse Wonder has a high-effort variant scoring 41, which tells you the effort dial moves the score by about three points in either direction. If Qwen4-Max's effort dial behaves similarly — and Alibaba's own reporting suggests effort changes token consumption substantially, with Work360 usage down 24 per cent at xhigh and 35 per cent at high — then the 43 is a single sample from a distribution, not a fixed property. Vendor-reported, relative to Qwen3.8-Max, unreproduced: treat the direction as informative and the magnitude as a claim.

Both vendors' launch benchmarks should be read with the same scepticism. Meta's JobBench and Vals figures are first-party. Alibaba's Work360 52.69, its cost-per-task drop from $4.43 to $1.67, and its bits-per-byte improvements across five base checkpoints are first-party. Neither set has been independently reproduced, and neither vendor publishes the eval harness that would let you try.

Where each one is the obvious answer

Choose Muse Wonder when audio is in the pipeline, when you need a published latency number to write an SLA against, or when the work is agentic and long-horizon and you would rather pay $6.95 for a task you can time than $3.27 for one you cannot. The 118-tokens-per-second figure is genuinely fast for a model at this score, and speed is a feature customers notice.

Choose Qwen4-Max when volume is the constraint, when the workload is long-context reading rather than long-horizon reasoning, or when the input is video at scale and the 2.7× input gap is the whole bill. The flat $1.50/$4.50 card with no long-context tier is easy to forecast against, which matters more than it sounds when you are writing a quarterly budget.

Both are proprietary with no published weights, so neither can be self-hosted or pinned. If that is a requirement, this comparison has no winner and the answer is a model that publishes checkpoints — Kimi K3.1 Pro, among the flagships, does.

If you want to avoid choosing, the useful compromise is a pool rather than a decision: route by input modality (audio to Muse Wonder, video-at-scale to Qwen4-Max), keep the price advantage where it is real, and let failover cover the latency you cannot yet predict. Worth being precise about what is routable today — neither Muse Wonder nor Qwen4-Max is among the 215 identifiers we serve, so that pool is a design, not a product page. What we do carry on the Meta side is the Muse Spark line at $1.25 in and $4.25 out, and on the Alibaba side qwen/qwen3.8-max at $2.00/$6.00, priced at the provider's list with nothing added.

Artificial Analysis page for Muse Wonder, captured 17 November 2026 in English, showing Intelligence #35 of 81, 19.64 seconds to first answer token, 118 output tokens per second, $4.00 input and $20.00 output per 1M tokens, 13.5 minutes and $6.95 per Intelligence Index task.

The question to ask before the next release

Every comparison of a weeks-old model against a month-old one has the same weakness: the evidence is asymmetric in ways that flatter whichever model shipped first, because independent testing has had longer to look at it. Muse Wonder's numbers are richer not because Meta is more transparent but because it has had five more weeks of measurement.

The thing to watch is not another benchmark. It is whether Qwen4-Max's speed and per-task timing appear on the same board before the price advantage gets spent. A model that is half the cost per task and twice the latency is a different product from one that is half the cost and comparable, and right now nobody outside Alibaba knows which of those two it is.

OrcaRouter model page for Muse Spark 1.2, captured 17 November 2026 in English, showing the identifier meta/muse-spark-1.2, text, image, video, file and audio input, a 1M-token context, a pricing panel and the Meta vendor label.

Compared in this article1

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