Muse Wonder vs MiniMax M3: 23 Index Points and 13× the Output Price, in Opposite Directions

M​iniMax M3 costs $0.30 per million input tokens and $1.20 per million output. Muse Wonder costs $4.00 and $20.00. That is a 13.3× gap on the output line and a 16.7× gap on measured cost per completed task — about $0.51 for MiniMax M3 against about $4.52 for Muse Wonder on the same independent harness. On capability, the same harness puts them 23 index points apart, 52 against 29. So the obvious reading is settled before the second paragraph: Meta Superintelligence Labs' October 12, 2026 release is the expensive, capable one, and M​iniMax's open-weight M3 is the cheap, weaker one. What makes the pairing worth a page is that the two models make opposite bets on the same engineering problem, and one of them publishes a number the other does not.

Muse Wonder states a 1,048,576-token window and prices every token in it identically. MiniMax M3 states the same window and then publishes a guaranteed usable minimum of 512K underneath it, because its sparse-attention design has a floor it has to promise. Which of those two statements is worth more to you depends entirely on where in the window your prompts actually sit.

The two models, and why they are not the same kind of object

MiniMax M3 is M​iniMax's flagship open-weight foundation model, announced on the vendor's blog on June 1, 2026 and listed in our own catalogue as available since May 31 — a one-day discrepancy of the kind these records always carry. It is a 428-billion-parameter mixture-of-experts model with roughly 23 billion parameters active per token, 60 layers, 128 routed experts with top-4 routing, and a 1,048,576-token context window with a 512,000-token maximum output. It accepts text, image and video input and returns text. Its three advertised claims are frontier-level coding and agentic performance, native multimodality including an ability to operate a desktop computer, and — the one that matters here — MSA, or M​iniMax Sparse Attention, a proprietary attention architecture the company proposed. It is open weights under the M​iniMax Community License, which is not Apache 2.0 and not MIT, and the distinction is worth reading rather than assuming.

Muse Wonder is Meta Superintelligence Labs' frontier release of October 12, 2026, shipped as a two-checkpoint series. Only muse-wonder-1.0 has an endpoint — on the Meta Model API and inside Muse Code — and it is the one every figure here applies to. Muse Wonder Full, the higher-scoring configuration, is sold only through a Muse Code subscription. 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, prices at $4.00 in, $0.20 cached and $20.00 out per million tokens, and runs reasoning mandatorily on an effort ladder from minimal to max that defaults to medium. There are no weights and no licence.

So the pairing is not simply cheap-versus-expensive. It is a downloadable 428B open-weights model against a rented proprietary one, which means the price comparison and the capability comparison are answering different questions.

What 23 index points are made of

Both figures come from Artificial Analysis on the same revision of the same instrument, at maximum effort, which is the only reason they can be put side by side at all.

• Intelligence Index — Muse Wonder 52, ninth of 216 models; MiniMax M3 29, eighteenth of 116 in its class

• Cost per completed index task — MiniMax M3 about $0.51; Muse Wonder about $4.52, roughly 8.9×

• Output tokens consumed on that run — MiniMax M3 about 120 million; Muse Wonder about 130 million, against a board median near 140 million for M​iniMax's cohort

• List rate — MiniMax M3 $0.30 in / $0.06 cached / $1.20 out per million; Muse Wonder $4.00 in / $0.20 cached / $20.00 out

• Context and output ceiling — both 1,048,576 in; MiniMax M3 512,000 out against Muse Wonder's 943,718

• Input surface — MiniMax M3 text, image and video; Muse Wonder text, image, speech and video

• Weights — MiniMax M3 downloadable under the M​iniMax Community License; Muse Wonder proprietary, none published

• Speed, independently measured — MiniMax M3 about 108 output tokens per second with a time to first token under a second; Muse Wonder about 115 tokens per second with a time to first token near 20 seconds

That last row is the one that changes how a workload feels rather than what it costs, and it is the largest practical difference on the list after price. A sub-second time to first token and a twenty-second one are different products for anything interactive, regardless of what each model scores when it finally starts writing.

Note also that the two models are close on output-token volume — 120 million against 130 million — which means the cost-per-task ratio is essentially the rate-card ratio and not an artifact of one model rambling. When a cheap model is both cheap per token and restrained in how many tokens it uses, the price difference is real and does not get eaten on the way to the invoice.

The cost arithmetic, in the two shapes that matter

Take two workloads. The first is a summarisation or extraction pipeline: 10 million tokens a month at a 70% input share, which is the default shape most estimators assume. The second is long-context retrieval: 5 million tokens a month at 800,000 tokens of prompt per call with a small response, cached.

• Short-call volume, no caching — MiniMax M3 about $5.70 a month; Muse Wonder about $11.30 a month. Roughly 2×, and for a workload at that size both numbers are rounding errors against an engineer's hourly rate.

• Short-call volume, caching on — MiniMax M3 about $4.86; Muse Wonder about $9.63. The cache line costs M​iniMax five cents per million against M​eta's twenty, but the absolute gap stays small.

• Long-context retrieval at 800,000 tokens of prompt — MiniMax M3 about $0.35 a call; Muse Wonder about $4.00 a call, because neither model steps its rate and Muse Wonder's input line is 13× higher. At a thousand calls a month that is $350 against $4,000.

• The same retirement workload with caching — MiniMax M3 about $0.14 a call; Muse Wonder about $2.20 a call

None of those figures are estimates of your bill — they are list-price arithmetic on each vendor's published rates. The point is the shape: for small workloads the 13× is coffee money, and for genuinely long-context workloads it is the difference between a service you can afford to run at volume and one you route selectively.

MSA against a flat window, which is a bet about failure modes

This is the part of the comparison with no benchmark row, and it is where the two models made opposite choices.

M​iniMax attributes M3's long-context behaviour to MSA, which the company describes as supporting a million tokens with a guaranteed usable minimum of 512K. That word — guaranteed — is doing real work. A sparse-attention system degrades in ways a dense one does not, and publishing a floor is the vendor telling you where it believes the degradation becomes unacceptable. If your prompts sit between 512K and 1M, that is a statement about the risk you are taking, and it is a more useful piece of information than a bare window size.

M​eta's Muse Wonder publishes 1,048,576 tokens and says nothing comparable about a usable floor. It also charges the same rate at 900,000 tokens as at 900, and it caps a single response at 943,718 tokens, which is the second-largest output ceiling we track. What it does not do is tell you where its long-context recall starts to sag. For an agent that opens a large context and works inside it for a long session, the absence of that number is a gap you will discover on your own traffic rather than read on a page.

The honest summary is that M​iniMax has published a guarantee it may or may not keep, and M​eta has published a window with no promise attached to it. Neither statement substitutes for testing your own prompts at your own lengths.

Open weights, and what the licence actually says

MiniMax M3 is downloadable, which means three things a rented model cannot offer: you can fine-tune it on proprietary data, you can run it inside your own perimeter, and you can be confident that the artefact answering your calls in six months is the same one answering them today. The M​iniMax Community License is not an unrestricted licence, and anyone planning to build a product on top of it should read the terms rather than assume MIT-style freedoms — the practical question is what it permits for redistribution and for commercial use at scale, and the answer lives in the licence text, not in a blog post. Compare with Apache-2.0 open-weights checkpoints, which is a materially different grant.

Muse Wonder offers none of that. It is API-only through M​eta's own surfaces, which is the trade a frontier proprietary release always makes: no fine-tuning, no self-hosting, no version you control, in exchange for capability that open weights at this price tier do not currently reach. Two days into its life, the independently measured gap is 23 index points, and that is the number you are paying the rent for.

Running them against each other without a second contract

The comparison is unusually easy to run here, because one side of it is already on a key you may hold. OrcaRouter carries MiniMax M3 as minimax/minimax-m3 at M​iniMax's own list price with 0% markup — provider list price passed straight through, so a vendor price cut is live on our side the same day it is announced — sitting alongside more than 200 other models behind one credential, with automatic failover across providers when an endpoint is unhealthy. Both of the numbers this page argues about are therefore checkable against your own token counts without a second vendor relationship.

Muse Wonder is not in our catalogue, and there is no third-party serving coverage we can verify. A router cannot route a model nobody else is serving, and the only path to it today is M​eta's own API and Muse Code. Where the routing story does apply is on the boundary: a routing DSL that composes several models into one call is what lets a cheap model carry the volume and a frontier model take the hard requests, and automatic failover keeps a single unhealthy endpoint from taking the whole path down. That pattern needs both models to be reachable through the same interface, which is true for one of these two and not yet for the other.

Which one, and for whom

Take MiniMax M3 when the workload is high-volume and the budget is the constraint. At $0.30 and $1.20 per million it is roughly a sixteenth of Muse Wonder on input and a thirteenth of it on output, it is the faster model to first token by an order of magnitude, it holds a million-token window with a published floor, and it is the one of the two you can download, fine-tune and keep. For summarisation, extraction, classification at volume, and long-context retrieval where you would otherwise be paying four dollars per call for input alone, the 23-point index gap does not describe your workload.

Take Muse Wonder when the work is hard, long and agentic, when you need the capability rather than the price, and when a 943,718-token response ceiling or a mandatory reasoning ladder is a feature rather than a tax. It is the only one of the two with an independently measured score above 50, and it is priced at the cheapest headline rate any of the three major labs has published for a frontier model — which is a strange sentence to write about a model that costs 13× what M3 does, and a fair one.

The buyer who should not choose between them is the one whose traffic has two shapes. The cost model that puts an 800,000-token prompt on a $20-per-million output model is a different cost model from the one that puts a two-paragraph classification task on it, and the two models on this page are priced almost exactly as if the market already knew that.