Muse Wonder vs Grok 4.6: A 1,048,576-Token Window Against One That Doubles Its Price at 200,000

Grok 4.6 is the cheaper model in this comparison by every published rate — $2.00 per million input tokens and $6.00 per million output against Muse Wonder's $4.00 and $20.00 — and there is a prompt length at which that advantage disappears completely. SpaceXAI's tiered rate card re-prices the entire request at double the base once a prompt crosses 200,000 tokens: input goes to $4.00, output to $12.00, cached input to $1.00. At 250,000 input tokens with 8,000 tokens of output, Grok 4.6 bills about $1.10 and Muse Wonder bills about $1.16. The model that looked 2.5× cheaper at short prompts is within 6% of its rival at long ones — and it is the rival, Meta Superintelligence Labs' October 12, 2026 release, that keeps a flat rate across the whole 1,048,576-token window.

That crossover is the most useful thing on this page, and it only exists because of a second difference. Grok 4.6 sells a 500,000-token context window and then charges the stepped rate for everything past 200,000 tokens; Muse Wonder sells twice that window and never steps. If your prompts live under 200,000 tokens, none of this applies and Grok 4.6 is simply the better buy. If they do not, the two models cost nearly the same and the comparison becomes about capability instead of price.

Two flagships, eight weeks apart

Grok 4.6 is SpaceXAI's flagship, released August 12, 2026, at $2.00 per million input tokens and $6.00 per million output with cached input at $0.50. It carries a 500,000-token context window, takes text, image and file input, returns text, supports configurable reasoning effort and native tool calling, and is priced in two tiers: everything up to 200,000 prompt tokens at the base rate, everything above it at double. Its publication record does not carry a maximum output figure for a single response at all, which is a gap rather than a zero — there is no number to plan a long-generation workload against.

Muse Wonder is Meta Superintelligence Labs' frontier release, shipped October 12, 2026 as part of a two-checkpoint series. The callable one is muse-wonder-1.0, served on the Meta Model API and inside Muse Code; the higher-scoring Muse Wonder Full is sold only through a Muse Code subscription and has no per-token price. The callable checkpoint accepts text, image, speech and video input, returns text, carries a 1,048,576-token context window, caps a single response at 943,718 tokens, bills $4.00 in, $0.20 cached and $20.00 out per million across the whole window, and runs reasoning mandatorily with an effort ladder from minimal to max that defaults to medium.

The independent board that has measured both puts them eight points apart — Muse Wonder at 52, Grok 4.6 at 44 — and also flags Grok 4.6 as superseded by Grok 4.7. That last detail matters for a reason that is not obvious: Grok 4.7 carries the same $2.00/$6.00 list price as the model it replaces, so the successor is not a cheaper option. It is a differently-shaped one.

The rate cards, and the one row that is not a number

Everything in this list is a vendor rate-card fact per million tokens, except the last two rows, which are independent measurements.

• First-tier input — Grok 4.6 $2.00 per million; Muse Wonder $4.00

• First-tier output — Grok 4.6 $6.00 per million; Muse Wonder $20.00

• Cached input — Grok 4.6 $0.50 per million; Muse Wonder $0.20

• Above 200,000 prompt tokens — Grok 4.6 goes to $4.00 in, $12.00 out and $1.00 cached on the whole request; Muse Wonder is unchanged

• Context window — Grok 4.6 500,000 tokens; Muse Wonder 1,048,576

• Maximum output in one response — Grok 4.6 not published; Muse Wonder 943,718 tokens

• Input surface — Grok 4.6 text, image and file; Muse Wonder text, image, speech and video

• Reasoning control — Grok 4.6 low through xhigh with a high default; Muse Wonder minimal through max with a medium default, and no off switch

• Intelligence Index, measured independently — Grok 4.6 44, thirty-seventh of 216; Muse Wonder 52, ninth of 216

• Cost of completing one index task, measured the same way — Grok 4.6 about $1.48; Muse Wonder about $4.52

The last row is where the price story gets honest. Three times the cost per completed task is a bigger gap than the blended rate card implies, and it is not caused by Grok 4.6 being extravagant — the opposite. On that evaluation Grok 4.6 consumed about 94 million output tokens against Muse Wonder's 130 million, with a board median near 85 million. The cheaper model is also the less talkative one, which is unusual in this price bracket and means the sticker gap is real rather than an artifact of one model writing more.

Where the crossover actually lands

Take a plausible workload: 70% of the prompt bill on input, a stable system prefix that caches, and a response of a few thousand tokens. The arithmetic below is on each vendor's published list prices and is our arithmetic, not a vendor's benchmark.

• 30,000 input and 8,000 output, uncached — Grok 4.6 about $0.11; Muse Wonder about $0.28. Grok is roughly 2.6× cheaper.

• 250,000 input and 8,000 output, uncached — Grok 4.6 about $1.10 at the stepped rate; Muse Wonder about $1.16 at its flat rate. The gap is about 5%.

• 400,000 input and 8,000 output, uncached — Grok 4.6 about $1.70; Muse Wonder about $1.76. Still level, and Grok 4.6 is now only 100,000 tokens from the end of its window while Muse Wonder is using 38% of its own.

• Cached reads at 250,000 tokens of context — Grok 4.6 $1.00 per million against Muse Wonder's $0.20, a 5× edge to M​eta once the cache is warm and the prompt is long enough to have tripped the step

The pattern is the point. Grok 4.6 is dramatically cheaper for the work most teams do most of the time, and it is level or slightly behind for the work that made people look for a million-token window in the first place. A cost model built from the headline rate badly overstates Grok's advantage on long-context traffic.

What eight index points buy, and what they do not

Measured independence cuts both ways here, and it is worth being precise about which numbers came from where.

Grok 4.6's strongest published figures are third-party and one of them is unusually good. On GPQA Diamond, a graduate-level science question set, it scores 94.9 — the highest first-tier score in its price class, and a place where Muse Wonder has no published independent row at all. It also posts 61.0 on Humanity's Last Exam and 76.8 on the independent coding assessment, both from the same evaluator. On question-shaped work at volume, those are the numbers that predict whether the model will embarrass you, and at $2.00/$6.00 they are a strong bargain.

Muse Wonder's strength is a different shape, and it is the reason the eight-point index gap is not the story either. M​eta's own launch table — vendor-reported, run in M​eta's harness, and unreproduced — is built almost entirely from long-horizon agentic and terminal tasks, and independent commentary has already selected from it. M​eta's DeepSWE 1.1 and SWE-Marathon rows put the series ahead of the models it compares against, and the 943,718-token output ceiling is a structural advantage no rate card captures: a workload that emits a very large single artifact, whether that is a generated application or a long structured document, has no Grok 4.6 number to compare against because Grok 4.6 does not publish one.

So the split is not better against worse. It is cheap-and-quick-versus-long-and-flat, and the axis that decides it is prompt length.

Three models now, not two, once the successor is in the frame

The independent board flags Grok 4.6 as superseded and points to Grok 4.7, which scores 46 against Grok 4.6's 44, runs at the same $2.00/$6.00 list with the same $0.50 cache read and the same 500,000-token window, publishes a 450,000-token output ceiling that its predecessor does not, and offers one more effort setting at the top. It is the obvious upgrade on paper.

The bill says something else. Grok 4.7's cost per completed index task is about $3.74 against Grok 4.6's $1.48, because it consumed roughly 240 million output tokens on the run where its predecessor used 94 million. Newer, higher-scoring and same-list-price still came out 2.5× more expensive per finished job. That is the rare case where the deprecation flag on a model page is not by itself a reason to move, and it is worth knowing before you migrate traffic you did not need to migrate.

For anyone building this as a two-model pipeline rather than a two-vendor project, the routing side is straightforward for one half and not the other. OrcaRouter carries Grok 4.6 as grok/grok-4.6 at SpaceXAI's own list price with 0% markup — provider list price passed through, so the tier prices and any change to them arrive here on the same day — behind the same key as more than 200 other models, with automatic failover across providers when an endpoint wobbles. Muse Wonder is not in our catalogue, and there is no third-party serving coverage we can verify, so the only way to call it today is M​eta's own API and Muse Code. A model fusion call that runs the cheap model and the strong one side by side and picks between the answers is one of the things a single endpoint makes easier — but it can only include models somebody is serving.

The pick, by prompt length rather than by verdict

Under 200,000 tokens of prompt, at volume, on question-shaped work: Grok 4.6, without much agonising. It is 2.6× cheaper on the request shapes most teams actually send, it scores 94.9 on the one independent benchmark where it beats everything in its class, and it is roughly 2× faster to first token than its rival. The one caveat is that Grok 4.7 exists at the same list price with an extra effort tier — check your own token counts before assuming it costs more, because the published figure says it will.

Over 200,000 tokens of prompt, or several hundred thousand tokens of context held open across a long agentic run, or a job whose output is enormous: Muse Wonder. The crossover is real and it is sharp, and the 943,718-token response ceiling is not a number Grok 4.6 can answer with one of its own.

Between them sits the honest observation that this is the one pairing in M​eta's launch window where the cheaper model is also the leaner one, so the price difference is buying capability rather than buying tokens. Two days after Muse Wonder shipped, that capability is described by one independent row and one vendor table. Grok 4.6's has been accumulating since August.