Muse Wonder API Access and Pricing: What $4/$20 Buys on Meta's First Frontier Model

Meta Superintelligence Labs began selling Muse Wonder on October 12, 2026, at $4.00 per million input tokens and $20.00 per million output tokens — and the more capable half of the release, Muse Wonder Full, does not have a price per token at all. That is the whole access story in one sentence: two checkpoints shipped, two different doors, and only one of them opens with an API key.

So this page answers the questions that come before any benchmark comparison. Which identifier you can call, what the rate card contains beyond the two headline numbers, what M​eta says the model does, what an outside evaluator has measured so far, and what the current answer is if you wanted to reach it from a router today. Nothing on this page implies OrcaRouter serves Muse Wonder — it does not, and we say so in the section where it matters.

Which Muse Wonder you can actually call

M​eta's material describes a series rather than a model. The two members behave differently as products, and the difference is administrative rather than technical.

• muse-wonder-1.0 — the checkpoint served on the Meta Model API and inside Muse Code, M​eta's terminal coding agent. This is the identifier a developer provisions against, and every per-token figure below belongs to it.

• muse-wonder-full-1.0 — a separate checkpoint that M​eta describes as the higher-scoring configuration. It is not an effort flag on the API model and it is not a second API endpoint; M​eta sells it through a Muse Code subscription. M​eta's published plans for that product are $50 a month for Power Usage and $200 a month for Maximum Usage, which means Full's real unit of purchase is a seat-month, not a million tokens.

That is an unusual shape for a frontier release, and it has a practical consequence worth stating plainly: there is no API path to the better-scoring checkpoint. A team that benchmarks the API model and likes it cannot simply turn a dial and get Full. They either buy the subscription, or they accept that the model they measured is the model they get.

The rate card, line by line

M​eta's per-token pricing applies to the API checkpoint. The rows below are vendor rate-card facts unless marked otherwise, all per million tokens.

• Input — $4.00 per million tokens

• Cached input — $0.20 per million tokens, which the independent evaluator Artificial Analysis reads as a 95% cache discount

• Output — $20.00 per million tokens

• Blended cost at a 7:2:1 cache-hit/input/output mix — about $2.94 per million tokens, per Artificial Analysis' own model page

• Measured cost of completing one Intelligence Index task — about $4.52, again Artificial Analysis' figure rather than M​eta's

• Context window — 1,048,576 tokens stated and served

• Maximum output in a single response — 943,718 tokens, which is the largest single-response ceiling on the board we track

• Input modalities — text, image, speech and video, against text output

• Reasoning control — mandatory. M​eta exposes an effort ladder from minimal to max, defaulting to medium, but not an off switch

One structural detail is easy to miss and matters to anyone pricing retrieval or long-agent work: the input rate does not step at a token threshold. A 400,000-token prompt bills at the same per-token rate as a 400-token one. Several competing flagships re-price the whole request once a prompt crosses a documented boundary, so this is a genuine difference in the shape of the curve rather than only in its height.

What M​eta says it scores, and what that is worth

M​eta shipped a benchmark table covering both checkpoints at their maximum reasoning settings, and it is the only scoreboard that covers Muse Wonder Full at all. It is vendor-reported, it was produced in M​eta's own harness, and none of it has been independently reproduced — read the following as M​eta's claim rather than as a measurement.

The rows M​eta promotes are the sustained-agent ones. On SWE-Marathon, a long-horizon software task set, M​eta places Muse Wonder Full ahead of every model in its comparison set. On DeepSWE 1.1 it makes a similar claim, and on the composite index it publishes the series carries its own name. Independent commentary has repeated those two rows selectively, and at least one widely shared post paired a Muse Wonder Full figure with a lower competitor number without noting that the underlying table is the vendor's, unreproduced, and unreleased as a harness.

The awkward part is internal to the table. Muse Wonder Full is not uniformly better than the cheaper API checkpoint: the two profiles trade rows, which suggests M​eta trained toward two different objectives rather than turning one dial up. If that is right, then "Full is the better model" is not a safe assumption for your own workload, and the only way to know which profile fits is to measure both — which currently means buying the subscription.

What an outside evaluator has measured

Artificial Analysis is the only evaluator with a published row for Muse Wonder as of October 13, 2026, and its numbers are a different species from M​eta's: same harness for every model on the board, published configuration, and no vendor input.

• Intelligence Index — 52 at maximum effort, ninth on a board covering more than 200 models

• Cost per index task — about $4.52, which the evaluator flags as above the median for the cohort once price is folded in

• Output tokens consumed on that run — about 130 million against a board median near 85 million, which the evaluator describes as somewhat verbose

• Serving speed — about 115 output tokens per second, with a time to first token near 20 seconds

• Muse Wonder Full — an index of 56 at maximum effort, ranked near the top of the board, with no per-token price because it is not sold per token

Note the last row against M​eta's table: on the composite index M​eta publishes 59.0 for Muse Wonder Full while the independent board reads 56. Three points on a scale like that is not a rounding difference, and the independent figure is the one to plan against, because it was produced the same way as the numbers for every other model you are weighing.

Where you cannot call it yet

Muse Wonder is available through M​eta's own API and inside Muse Code, and that is currently the whole list. There is no third-party serving coverage in the catalogue we can verify, and it is not in OrcaRouter's catalogue — a router cannot route a checkpoint nobody else is serving, and we would rather say that than imply otherwise.

What that means in practice is a two-vendor setup if you want to try it: M​eta's key for the new model, and whatever you already use for everything else. That is the situation a router is built for on the other side of the fence. OrcaRouter passes provider list price through with 0% markup across more than 200 models behind one key, which includes M​eta's own Muse Spark 1.2 and Muse Spark 1.1 — so the M​eta models you can call today are already reachable in one place, and a rate change at M​eta lands here the same day it is announced rather than waiting for a middleman to re-price. When a second provider does start serving Muse Wonder, automatic failover across endpoints is what keeps a single unhealthy path from taking your traffic down.

What would change this answer

Four things are checkable, and any of them would make this page worth re-reading rather than re-guessing.

• A model identifier appearing for Muse Wonder Full on the Meta Model API, which would end the subscription-only boundary

• A revision to the independent index that changes the 52, or the arrival of a second evaluator with its own row

• A price change on either rate-card line, which at 0% markup is live on our side the same day

• Muse Wonder appearing in a third-party catalogue with a published rate, which would make the availability question a routing question instead of a vendor question

Until then the honest summary is that M​eta has priced its API checkpoint 20% under the current A​nthropic and O​penAI flagships on input and output alike, made its cheaper model the one you can reach programmatically, and left the interesting configuration behind a monthly seat. The $4/$20 is real and it is the cheapest frontier rate M​eta has ever published. It is also the price of the model you can actually call.