Muse Wonder vs Grok 4.7: The Model That Costs Twice as Much per Token and Only 1.2× as Much per Task

On rate cards, Muse Wonder looks like it belongs in a different price bracket from Grok 4.7. Meta charges $4.00 per million input tokens and $20.00 per million output for Muse Wonder, which it released on October 12, 2026. xAI charges $2.00 and $6.00 for Grok 4.7, released September 21, 2026. That is double on input and more than triple on output. Then Artificial Analysis runs both through the same harness and reports the price of finishing one intelligence task: about $4.52 for Muse Wonder (max) and about $3.74 for Grok 4.7 (xhigh). A 3.3× rate-card gap collapses to 1.2× at the till, because Grok 4.7 is the most verbose model on the board and Muse Wonder is only the second most.

Where the rate-card gap goes

The two labs price tokens very differently and the models spend them very differently, in opposite directions. One number, output tokens consumed during the index evaluation, explains almost all of it:

• Output tokens on the index run — Grok 4.7 about 240 million; Muse Wonder about 130 million, against a board median near 82 million

• List rate — Grok 4.7 $2.00 in / $0.50 cached / $6.00 out per million; Muse Wonder $4.00 in / $0.20 cached / $20.00 out

• Measured cost per index task — Grok 4.7 about $3.74; Muse Wonder about $4.52

• Intelligence Index — Muse Wonder (max) 52, ninth of 219; Grok 4.7 (xhigh) 46

Grok 4.7's verbosity is the reason its cheap-looking rate card does not produce a cheap task. At 240 million output tokens per evaluation, a Grok 4.7 call that says the same thing as a Muse Wonder call takes roughly twice as many tokens to say it, and the saving on each token gets spent buying more of them. Muse Wonder's $20.00 output rate is nearly three times Grok's, but it emits about half the tokens — the two effects very nearly cancel, and what survives is a 1.2× premium for six index points.

That reframing matters more than it looks, because it inverts the usual advice. If you price models off the rate card you will conclude Grok 4.7 is the budget option and Muse Wonder is a luxury; if you price them off completed work, they are within 22% of each other.

The cache line is the one that does not cancel

One row breaks the pattern, and it breaks it in Meta's favour. Grok 4.7's cached input rate is $0.50 per million, unchanged from Grok 4.6 and five times GPT-6.1 Sol's $0.10. Muse Wonder's is $0.20. On a prompt-caching-heavy pipeline — a long system prompt and a fixed corpus re-sent on every call — that is a 2.5× advantage on the largest token category in the request, and it does not get eaten by verbosity because cached input tokens are not output tokens.

The other non-cancelling difference is the shape of the price curve, and here Meta wins outright:

• Context window — Muse Wonder 1,048,576 tokens; Grok 4.7 500,000 tokens

• Long-prompt repricing — Muse Wonder none; Grok 4.7 re-prices the whole request at $4.00 in and $12.00 out once a prompt reaches 200,000 tokens

• Maximum output in one response — Muse Wonder 943,718 tokens; Grok 4.7 450,000 according to the vendor's own listing

• Effort ladder — Muse Wonder minimal through max, defaulting to medium; Grok 4.7 low, medium, high, xhigh, defaulting to high, with xhigh new in this generation

Grok 4.7's 200,000-token step is the steepest cliff either lab sells. A prompt at 199,000 tokens and one at 201,000 tokens are billed at double the input rate and double the output rate on the whole request — not just on the excess. Muse Wonder charges a 400,000-token prompt at exactly the same per-token rate as a 400-token one. For retrieval-style workloads that routinely carry a large fixed context, that single row can outweigh everything above it.

Where the six index points come from

A six-point composite gap is a real gap, and it does not distribute itself evenly. Artificial Analysis scores both models on the same revision of its index; these are its measurements rather than any vendor's claims.

• AA-Briefcase — Grok 4.7 1,657 Elo at xhigh, fourth on the board; Muse Wonder not published on the same harness

• GDPval-AA — Grok 4.7 1,715 Elo, a high-water mark among non-Anthropic models; Muse Wonder not published

• Code generation on its own harness — Muse Wonder not directly comparable; Grok 4.7 57.4% on SciCode and 43.1% on Humanity's Last Exam

• Cost per index task — Grok 4.7 $3.74; Muse Wonder $4.52

Grok 4.7's profile is broad rather than spiky: it is competent across reasoning, long-context recall and professional document work without leading on any of them, and it is genuinely fast at 82.6 tokens per second, which Muse Wonder has not had published on the same harness. Muse Wonder's shape is the opposite — a higher composite, an unpublished speed figure, and a vendor table that concentrates its wins in terminal and long-horizon software work.

That vendor table is Meta's own and has not been reproduced by anyone. Its Terminal-Bench 4.0-AA figure of 85.6 for Muse Wonder sits against 79.0 for Claude Fable 5.1 and 81.4 for GPT-6 Astra in the same column; Grok 4.7's independently measured Terminal-Bench 4.0 result is 0.258 on the same benchmark name, which is a different harness and a different number. Do not put those two in the same sentence as a comparison — Meta's figure is a vendor claim about its own model and Artificial Analysis' figure is a measurement of xAI's, and the harnesses are not the same.

One of these has a subscription twin

Meta shipped Muse Wonder as two checkpoints. muse-wonder-1.0 is the one on the Meta Model API at $4/$20. muse-wonder-full-1.0 scores 56 on the independent index — ten points above Grok 4.7 — and cannot be bought per token at all; it comes with a Muse Code subscription at $50 a month for Power Usage or $200 a month for Maximum Usage. Grok 4.7 has no equivalent split. What you see is what you buy, on one rate card, at one endpoint.

For a team weighing a $200 monthly subscription against per-token billing, that is the genuinely hard part of this comparison. The Full checkpoint's ten-point lead over Grok 4.7 is larger than the API model's six, and a flat subscription is a very different purchase from a metered rate — it removes the verbosity penalty entirely, which is the penalty that made this pairing close in the first place.

The routing side, which is asymmetric here

Both of these are Meta and xAI's own models, sold through their own APIs, and one of them is also a route we serve. Grok 4.7 is in the OrcaRouter catalogue as grok/grok-4.7, dated 2026-09-21, at xAI's own $2.00 input, $0.50 cache-read and $6.00 output per million tokens, with the 500,000-token window, the 450,000-token output ceiling and the 200,000-token repricing rule all intact. Pass-through pricing means a change on xAI's rate card shows up on ours the same day, and the same key reaches the rest of the catalogue.

Muse Wonder is not one of our routes. There is no meta/muse-wonder entry in our catalogue and this page is not going to imply one — Meta's Model API and Muse Code are where it runs. What we do carry from Meta Superintelligence Labs is meta/muse-spark-1.2, the larger checkpoint this family descends from, at $1.25 input and $4.25 output per million across a 1,048,576-token window with $0.15 cache reads.

That asymmetry is actually the practical argument for the comparison. The 1.2× cost ratio and the six-point score gap are both averages over a benchmark suite; on your traffic the ratio will be different and you cannot know by how much without running it. Grok 4.7 is a string change away. When Muse Wonder reaches a provider we route, it will be the same string change — and until then, the honest position is that you can test one half of this pairing cheaply and the other half not at all.

Choosing between them

• Grok 4.7 for anything where you already like the shape: a 500,000-token window covers most retrieval work, the 450,000-token output ceiling is generous, the effort ladder goes to xhigh, and the composite is close enough that 1.2× rarely decides anything.

• Grok 4.7 for reasoning-heavy professional document work specifically. Its GDPval-AA Elo is the strongest non-Anthropic reading on the board, and that is the sub-score that predicts contract review, financial analysis and long-form drafting.

• Muse Wonder when prompts are large and fixed. A 1M window with no repricing step and a $0.20 cache read is a materially different deal from a 500K window that doubles the whole request at 200,000 tokens, and that difference compounds across every call in a loop.

• Muse Wonder when the output has to be enormous. 943,718 tokens against 450,000 is the difference between emitting a whole artefact and paginating one.

• Muse Wonder when speech or video is the input. Grok 4.7 takes text and images; Meta's model takes text, image, speech and video, which collapses a pipeline stage rather than improving one.

The tempting summary — cheaper model versus better model — is wrong on both halves. Grok 4.7 is not cheaper per completed task by anything like its rate card suggests, and Muse Wonder's advantage is concentrated in places a composite score cannot show. What separates them in practice is prompt size and output size, and both of those are printed in your own logs.