Muse Wonder vs Qwen 3.8 Max: The Cheaper Token That Costs More Per Task

Qwen 3.8 Max is cheaper than Muse Wonder on every published rate — $2.00 per million input tokens and $6.00 per million output against $4.00 and $20.00 — and it still costs more to finish a job. On Artificial Analysis's own harness the measured cost of completing one Intelligence Index task is about $5.41 on Qwen 3.8 Max and about $4.52 on Muse Wonder, because Qwen 3.8 Max emits roughly 190 million output tokens to get through that run while Muse Wonder emits about 130 million, against a board median near 85 million. Meta Superintelligence Labs shipped Muse Wonder on October 12, 2026 and it scores 52 on the same index; Qwen 3.8 Max scores 45. That is not a rounding gap on either axis, and it means the usual "cheaper per token wins" reflex gets this pairing exactly backwards.

The rate card is not the bill

Start with what the two vendors publish, because on paper there is no contest.

• Input — Qwen 3.8 Max $2.00 per million tokens; Muse Wonder $4.00, so it is half price

• Output — Qwen 3.8 Max $6.00 per million tokens; Muse Wonder $20.00, so it is roughly a third

• Cache discount — Qwen 3.8 Max 88% off cached input; Muse Wonder 95%. Because the percentage is taken off a list input price that is twice as high, the two land within a few cents of each other once rounded — about $0.24 against about $0.20 per million cached tokens — and our own catalogue prices Qwen 3.8 Max cache reads at $0.25. This is the one line where the cheaper sticker price buys the more expensive cache.

• Vendor's own framing — Artificial Analysis flags Muse Wonder's input as "somewhat expensive" against a board median of $2.00 and its output as "expensive" against a median of $10.00. Qwen 3.8 Max sits at or below both medians.

Every line of that favours Qwen 3.8 Max. Now look at what the same harness measured for a completed task, which is the only number a budget actually feels.

• Cost per Intelligence Index task — Muse Wonder about $4.52; Qwen 3.8 Max about $5.41

• Output tokens consumed on that index run — Muse Wonder about 130 million; Qwen 3.8 Max about 190 million, against a board median near 85 million

• What the harness says about the appetite — Muse Wonder is described as "somewhat verbose"; Qwen 3.8 Max is described as "very verbose"

The two rate cards and the two measured bills point in opposite directions, and the reason is entirely token volume. Qwen 3.8 Max spends about 1.5 times the output tokens Muse Wonder does on the same evaluation, which cancels its 3.3-to-1 advantage on the output line and then keeps going. If your workload looks like the index — long reasoning, long answers, high effort settings — the rate card is the wrong document to plan against. Your own token counts are the right one, and the gap between the two is only worth chasing if you have measured your own mix rather than borrowed somebody else's.

Seven index points is a tier, not a tie

Comparisons between models that finish within a point of each other are decided by price and shape. This one is not that.

• Intelligence Index — Muse Wonder (max) 52; Qwen 3.8 Max 45

• Shared measurement — both figures are Artificial Analysis's, produced on its own harness at maximum effort, not either vendor's preferred benchmark

Seven points on a composite that spans 219 models in this class is a real separation, and it comes from the side of the ledger where Qwen 3.8 Max is spending the extra tokens. It is generating more output and scoring lower. That combination — more tokens, lower score — is the expensive kind of verbosity, because you pay twice: once in dollars per task and once in whatever the score is standing in for.

The honest caveat is that a composite index is an average over somebody else's task mix. Seven points on the instrument tells you the two models are not peers; it does not tell you which one passes your evaluation. A single index number has never answered that and this pairing is no exception.

The spec sheet, line by line

• Context window — Muse Wonder (Max) 1,048,576 tokens, with the million-token window restated in Meta's own card summary; Qwen 3.8 Max 984k tokens as the independent card states it, against 1,000,000 tokens on our own catalogue page for the same named model. The two sources disagree by 16,000 tokens on that model's side; treat the ceiling as "about a million" on both and check the smaller number against the vendor before you design to it.

• Modal input — Muse Wonder text, image, speech and video; Qwen 3.8 Max text and image, with video also listed on our catalogue page. Speech is the input type only one of these two accepts.

• Output — text on both.

• Reasoning — both reason, and neither publishes a no-reasoning mode that a caller can rely on. Reasoning tokens bill as output on both, which is exactly the line where Qwen 3.8 Max's token appetite gets expensive.

• Output speed — Qwen 3.8 Max about 36 tokens per second, which the harness calls notably slow; Muse Wonder not yet published on the same harness. That asymmetry matters more than it looks: it means one side of this comparison has a publicly measurable latency profile and the other does not.

• Weights — proprietary on both, neither published.

• Release — Muse Wonder on October 12, 2026. Qwen 3.8 Max is the older model here: Artificial Analysis measures the 0902 checkpoint, and our catalogue lists its own Qwen 3.8 Max entry under a different date, 2026-08-03. Those two ids are not interchangeable, and if you are comparing prices or scores it is worth being sure which checkpoint you are actually calling.

Latency is the axis only one of them documents

The speed line above is the least comfortable part of this page, because it is the one place where the comparison stops being symmetric. Muse Wonder has no published output speed on the independent harness — not a slow number, no number — so any claim that one of these two answers faster is unsupported in at least one direction.

What can be said is what our own traffic shows. On OrcaRouter, Qwen 3.8 Max carries a measured p50 time-to-first-token of about 4.92 seconds. That is the wait before the first token arrives, on a model the independent harness already rates as notably slow at sustained output. For an interactive surface that is a long silence; for a batch job it is nearly irrelevant. Which of those two you are building decides how much weight this section deserves, and if you are building the interactive kind, a model whose latency nobody has published is a risk you should price before you adopt it rather than after.

Running both, or running one

Qwen 3.8 Max is on OrcaRouter as qwen/qwen3.8-max, at list $2.00 input and $6.00 output per million tokens with a $0.25 cache-read rate and the million-token window intact. Nothing is marked up on our side — provider list price passes through, so when the provider changes a rate the number on our page changes the same day. That makes this side of the comparison a one-string change away from any harness you already have, which is the cheapest possible way to find out what your own task mix does to that $2.00.

Muse Wonder is not in our catalogue, and this page will not pretend it is — Meta's own Model API and Muse Code are where it runs today. What we do carry from the same lab is Muse Spark 1.2, the checkpoint Meta built the rest of this family on, at $1.25 input and $4.25 output per million tokens across a 1,048,576-token window with $0.15 cache reads. If what you actually want from Muse Wonder is the Meta house style at a fifth of the price rather than the frontier score, that route exists on the same key you already have.

The practical case for one OpenAI-compatible endpoint across 200-plus models is precisely this measurement problem. The $4.52-versus-$5.41 spread is an average over somebody else's tasks, and the only way to know whether it holds for yours is to run it. Doing that through a single key means pointing your existing evaluation at a second model name and re-running it, not signing a second contract and rebuilding your billing. Automatic failover covers the case where you ship the newer, unproven one first: if the upstream provider degrades, the request moves rather than fails.

Which one, and on what evidence

• Take Muse Wonder when the score is the point. Seven index points at a lower measured cost per task is a straightforward win, and nothing in this pairing reverses it.

• Take Qwen 3.8 Max when the input is long and the output is short. A rate card of $2.00 in against $4.00 is a real 50% saving on workloads that are all prompt and almost no answer, and the verbosity penalty never gets a chance to apply because there is nothing to be verbose into.

• Take Qwen 3.8 Max when you need a published latency profile. Its speed is documented, and if procurement wants a number before sign-off, "not published" is a harder answer to defend than "36 tokens per second".

• Do not take Qwen 3.8 Max for its headline price. The $2.00 input rate is a real discount on input and a misleading summary of the model: on the independent harness it costs more per completed task than a model charging double for input and more than triple for output.

• Do not assume either of them is the cheap option until you have counted your own tokens. The per-task figures on this page come from one harness running one evaluation. The inversion they show is the reason to measure, not the measurement itself.

What makes this pairing interesting is that it breaks the shortcut. Rate cards are easy to read and easy to trust, and here the easy read gives the wrong answer by about 20%. The models are also a generation apart on release date but close enough in shape that the choice is not obvious from a spec sheet. Both of those are answerable from your own logs this week — and neither will be answered by the next benchmark table either lab publishes.