Muse Wonder vs GPT-6.1 Sol: A 943,718-Token Output Ceiling Against a 128,000-Token One

Muse Wonder and GPT-6.1 Sol finish within a fifth of a point of each other on the Artificial Analysis Intelligence Index — 52 for Muse Wonder (max), 51.8 for GPT-6.1 Sol (max) — and a single index task costs $4.52 on the first and $0.72 on the second. Meta Superintelligence Labs shipped Muse Wonder on October 12, 2026 at $4.00 per million input tokens and $20.00 per million output. GPT-6.1 Sol shipped on September 29, 2026 at $2.00 and $10.00. The rate card says roughly double; the measured bill says 6.3 times, and the reason is verbosity — 130 million output tokens against 67.2 million to land on almost the same score. But the number that actually decides which of these two you can put on a job is neither the index nor the price. It is the output ceiling: 943,718 tokens for Muse Wonder, 128,000 for GPT-6.1 Sol.

The ceiling is the spec line nobody reads

Both models advertise a million-token context. That is where the similarity in the spec sheet stops being useful.

• Context window — Muse Wonder 1,048,576 tokens vs GPT-6.1 Sol 1,050,000 tokens

• Maximum input — Muse Wonder not published separately; GPT-6.1 Sol 922,000 tokens

• Maximum output in one response — Muse Wonder 943,718 tokens; GPT-6.1 Sol 128,000 tokens

• Reasoning effort — Muse Wonder minimal, low, medium (default), high, xhigh, max; GPT-6.1 Sol low, medium (default), high, xhigh, max, with no off switch on either

• Modal input — Muse Wonder text, image, speech and video; GPT-6.1 Sol text, image and file

• Weights — both proprietary, neither published

Read the third line again. GPT-6.1 Sol buys a 1,050,000-token window but lets you spend only about an eighth of it on output; the remaining 922,000 is input-only. Muse Wonder lets a single response occupy nine-tenths of its window. That is not a rounding difference. A synthesis pass over a large corpus, a long-form generator, a transcript-to-transcript transformation, a harness that has to emit a whole file rather than a patch — those live above 128,000 output tokens, and on GPT-6.1 Sol they do not fit in one call at any price. You paginate, and pagination means the model loses the plan it made in the first response.

Muse Wonder's ceiling comes with its own trap, which Meta does not hide but does not foreground either: reasoning tokens bill as output on both models, and Muse Wonder's ladder runs to a max setting that will happily consume the whole budget. The 943,718 figure is a limit, not an allowance.

Two rate cards, and the cliff hiding in one of them

GPT-6.1 Sol is cheaper on every line, and cheaper still in the middle of its range. What it does not have is a flat curve.

• Standard rate — Muse Wonder $4.00 in / $20.00 out per million; GPT-6.1 Sol $2.00 in / $10.00 out

• Cached input — Muse Wonder $0.20 per million; GPT-6.1 Sol $0.10 per million

• Long-prompt behaviour — Muse Wonder flat across the whole window; GPT-6.1 Sol reprices the entire request above 272,000 input tokens at $4.00 in and $15.00 out

• Cache write — Muse Wonder not published; GPT-6.1 Sol $2.50 per million

• Discount paths — Muse Wonder none published; GPT-6.1 Sol batch and flex at 50%, fast mode at 2×

That 272,000-token threshold is the most expensive line on GPT-6.1 Sol's card, because crossing it re-prices work you have already done. A request at 271,000 tokens and a request at 273,000 tokens differ by under 1% of input and by 2.01× on the invoice. Muse Wonder has no equivalent discontinuity — a 900,000-token prompt is charged at the same per-token rate as a 900-token one. For a workload that habitually sits just above 272,000 tokens, that is the single strongest argument on Meta's side of this page, and it is worth more than the cache discount GPT-6.1 Sol wins on every other row.

Below that threshold the ordering is unambiguous: GPT-6.1 Sol is half the price to run, half the price to cache-read, and it is the only one of the two with a batch lane at 50% off.

Where the 0.2 index points actually live

A composite index that lands 52 against 51.8 tells you the two models are peers. It tells you nothing about which one will pass your workload, and on the sub-scores the pair separates into recognisable shapes rather than a winner and a loser.

Artificial Analysis is the source for both, measured on its own harness at max effort, and these are its numbers rather than either vendor's:

• Intelligence Index — Muse Wonder (max) 52, ninth of 219 models; GPT-6.1 Sol (max) 51.8, tenth

• Cost per index task — Muse Wonder about $4.52; GPT-6.1 Sol about $0.72

• Output tokens consumed on the index run — Muse Wonder about 130 million; GPT-6.1 Sol about 67.2 million, against a board median near 82 million

• Output speed — GPT-6.1 Sol about 66.8 tokens per second; Muse Wonder not yet published on the same harness

The verbosity gap is the whole story of the price gap. Muse Wonder spends nearly twice the tokens to produce a score half a point higher, and at $20 per million output against $10 it pays twice as much for each of those tokens. Cost per task multiplies the two effects. Nothing about that ratio is fixed — reasoning effort is the lever, and both models expose it — but the default settings are where most traffic runs, and the defaults are what produced these figures.

Meta's own launch table, which is vendor-reported and has not been independently reproduced, puts Muse Wonder ahead of GPT-6 Astra and level with Claude Opus 5.5 on Terminal-Bench 4.0-AA at 85.6 against 81.4 and 81.5, and ahead on SWE-Marathon at 78.8 against 74.1 and 74.2. Read those as Meta's claims. Note also that the same table puts Muse Wonder behind on AutomationBench-AA (60.5 against 65.3 and 65.6), GDP.pdf (39.1 against 41.6 and 41.2) and DeepSWE 1.1 (52.0 against 59.1 and 59.6) — the vendor's own scoreboard is not uniformly flattering, and the two places it is flattering are terminal and long-horizon code, not business-process automation.

What the subscription twin does to this comparison

Muse Wonder is not one model. Meta shipped a second checkpoint, muse-wonder-full-1.0, which scores 56 on the same independent index and ranks fourth of 699 — four points above the API model and above GPT-6.1 Sol's 51.8. It has no per-token price because Meta does not sell it per token: it is reachable only through a Muse Code subscription at $50 a month for Power Usage or $200 a month for Maximum Usage.

For a team already paying for Muse Code, the effective price of a four-point index gain is a flat monthly fee, and the comparison against GPT-6.1 Sol changes shape entirely. For everyone else, the better half of this launch is a product decision, not an API decision, and the honest reading is that Meta's frontier number and Meta's reachable number are two different figures.

Running both, or running one

The uncomfortable fact at the centre of this page is that neither of these is a decision you can make from a spec sheet, because the 6.3× cost ratio is an average over somebody else's task mix. Your own mix is the only one that matters, and measuring it should not require two vendor accounts.

GPT-6.1 Sol is on OrcaRouter as openai/gpt-6.1-sol, dated 2026-09-29, at list $2.00 input, $0.10 cache-read and $10.00 output per million tokens with the 1,050,000-token window and the 272,000-token repricing rule intact. Nothing is marked up on our side; when the provider changes a rate, the number on our page changes the same day. That makes the Sol half of this comparison a one-string change away from any harness you already have.

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 your reason for looking at Muse Wonder is the Meta house style rather than the frontier score, that route costs a fifth of it.

The practical use of a single OpenAI-compatible endpoint across 200-plus models is exactly this comparison. The $4.52-versus-$0.72 spread is large enough to be worth testing rather than believing, and testing it means pointing an existing evaluation at a second model name and re-running — not signing a second contract. Automatic failover then covers the case where you ship the unproven one first: if the upstream provider degrades, the request moves rather than fails.

Which one, and on what evidence

• Take GPT-6.1 Sol unless something specific rules it out. It is cheaper on every published rate, it has a batch lane, its cache read is half Muse Wonder's, and its independent index is a fifth of a point behind. That is the default answer and it should stay the default until a workload disagrees.

• Take Muse Wonder when the output has to be long. A 128,000-token response ceiling is a hard wall, and if your job needs a single 200,000-token generation, no amount of saving on input tokens makes GPT-6.1 Sol able to do it.

• Take Muse Wonder when the input is speech or video. Meta accepts both as first-class inputs; GPT-6.1 Sol takes text, image and file. Collapsing a transcription step and a reasoning step into one call is worth more than a rate difference on some pipelines and nothing at all on others.

• Take Muse Wonder when your prompts sit just past 272,000 tokens. That is the one region of the price curve where Meta's flat rate beats the Sol cliff, and it beats it decisively.

• Do not take Muse Wonder for its headline index score. Fifty-two against 51.8 is a tie on the instrument that produced both numbers, and paying 6.3 times as much per task for a tie is not a strategy.

What makes this pairing genuinely hard is that the two models are close enough on quality that the decision is entirely about shape — how long the output has to be, what the input is made of, and where your prompts sit on the price curve. Those are answerable from your own logs this week, and they will not be answered by the next benchmark table either lab publishes.