Muse Wonder vs Gemini 4 Argon: One Is Callable Today, One Has No Endpoint

Muse Wonder and Gemini 4 Argon are, on the numbers, close enough to be a real decision — Muse Wonder (max) at 52 on Artificial Analysis' Intelligence Index, Gemini 4 Argon (high) at 53, priced at $4.00/$20.00 and $2.00/$10.00 per million tokens respectively. But only one of them has a model ID. Muse Wonder has been generally available since October 12, 2026 on the Meta Model API and inside Muse Code. Gemini 4 Argon was announced on September 30, 2026, has been benchmarked by independent evaluators, and cannot be called: there is no published model identifier, no public endpoint, and no general-availability date. That single fact reorders the comparison before any benchmark gets discussed.

So this is a piece about a matchup that is partly theoretical. Argon looks like the better buy on nearly every published axis. It is also, as of today, a set of numbers rather than an API. What follows separates the part of the comparison you can act on from the part you can only plan around.

The gap that decides it, for now

Gemini 4 Argon was announced by Koray Kavukcuoglu on September 30, 2026. It has been measured and placed on the same independent index that scores Muse Wonder, which means the numbers below are directly comparable in a way vendor tables never are. What has not shipped is access: no model identifier, no pricing page, no endpoint, and no committed date. What exists is a gated rollout under the Fairwind Program, which means a limited set of partners and no public path in.

Muse Wonder is the opposite shape. It has been callable since October 12 on the Meta Model API under the ID muse-wonder-1.0 and inside Muse Code, with a documented parameter surface: a reasoning-effort setting with values from minimal to max, a default of medium, an output budget cap of 943,718 tokens, and a published input and output price. You can build against it this afternoon, and you can forecast what a call will cost.

If you are choosing a model to ship on this month, the comparison ends here and the answer is Muse Wonder — not because it is better, but because Argon is not a thing you can call. Everything after this section applies to a decision you will face whenever Argon opens up, and is worth reading now so the evaluation is not a scramble later.

On paper, Argon looks like the better model

Every figure in this section is Artificial Analysis' independent measurement rather than a vendor claim, including the Argon ones:

• Intelligence Index (measured config) — Gemini 4 Argon 53, ranked #8 of 219; Muse Wonder 52, ranked #9 of 219

• Cost per Intelligence Index task — Gemini 4 Argon about $1.99; Muse Wonder about $4.52

• List price — Gemini 4 Argon $2.00 in / $10.00 out per million tokens; Muse Wonder $4.00 in / $20.00 out

• Blended price on a 7:2:1 cache-hit/input/output mix — Gemini 4 Argon about $1.47 per million tokens; Muse Wonder about $2.94

• Cache discount — Gemini 4 Argon roughly 95%; Muse Wonder roughly 95%

• Output tokens consumed in evaluation — Gemini 4 Argon about 110 million; Muse Wonder about 130 million

• Context window — both 1,048,576 tokens

• Modal inputs — Gemini 4 Argon text and image in, text out; Muse Wonder text, image, speech, video, and files in, text out

• Release status — Gemini 4 Argon announced 2026-09-30, no public endpoint; Muse Wonder generally available 2026-10-12

On those figures Argon is a point ahead on capability, less than half the cost per measured task, half the headline price, and leaner on output tokens. That is not a marginal win — cost per index task is the number that maps most directly onto a real invoice, and $1.99 against $4.52 is a different order of operating expense.

The one-point index lead deserves a caveat rather than a drumroll. A single index point between two models is inside the range that re-runs and configuration changes move, and the two rows are not measuring the same reasoning setting: Argon's published row is its high configuration, while Muse Wonder's is max. Argon has no published max row to compare against. The cost picture is robust; the ranking is not.

Where Muse Wonder is genuinely ahead

Three things, and only one of them is a benchmark.

The first is availability, already covered, and it is not a soft advantage. A model with an endpoint and a documented parameter surface can be benchmarked on your data, tuned, and put behind a budget. A model with neither can be planned around and nothing more.

The second is that the parameter surface is knowable. Muse Wonder's cost behaviour is legible: reasoning is mandatory, the effort ladder runs from minimal to max with a medium default, and every reasoning token bills as output. That means a team can run extraction and classification at minimal and reserve high or max for the calls that need them, with a predictable effect on spend. Nothing comparable has been published for Gemini 4 Argon — no reasoning-effort enumeration, no output-token cap, no indication of whether thinking can be disabled or throttled. You cannot budget a dial you cannot see, and until the vendor documents that surface, every Argon cost projection rests on the single blended figure above.

The third is modality on the input side. Muse Wonder accepts speech and video directly; Gemini 4 Argon's published input modalities are text and image. For a pipeline that reasons over recorded calls or video, that is the difference between one call and a transcode-and-transcribe stage you build and pay for separately — and it narrows the effective price gap in a way the per-token table does not show. The Gemini 4 line has historically been strong on video, so this is a gap in what has been announced rather than a settled architectural difference, and it is worth re-checking when Argon ships.

There is also a sibling worth naming. Meta's muse-wonder-full-1.0 checkpoint measures 56 on the same independent index — above Argon's 53 — but it is Muse Code subscription only and not sold per token, so it competes with Argon on a subscription budget rather than a per-token one. It is not an API alternative to Argon and should not be treated as one.

What cannot be compared at all yet

Several of the axes that decide real deployments have no Argon data because there is no Argon service to measure. Throughput and latency do not exist for a model with no endpoint. Provider coverage does not exist either: Muse Wonder is on Meta's own API and in Muse Code, and its reach through gateways is whatever those gateways have implemented since October 12 — while Argon's eventual reach is unknown. Rate limits, regional availability, uptime history, and the cache-read rate that turns a 95% discount into an actual saving are all unanswered on the Argon side.

That asymmetry has a practical consequence. A cost-per-task number computed on a single evaluation run is a useful planning input and a poor commitment. Muse Wonder's $4.52 will move as its configuration surface gets exercised in production, and Argon's $1.99 is the only figure that will ever exist until someone can call the model and re-measure. Treat both as current best estimates, not as terms.

Neither is on OrcaRouter — here is what is

Being direct about this: Muse Wonder is not in OrcaRouter's catalogue, and neither is Gemini 4 Argon. There is no meta/muse-wonder entry and no google/gemini-4 entry. We will not present a route that does not exist, and Argon in particular has nothing to route to.

What we do carry from Meta Superintelligence Labs is the Muse Spark line — Muse Spark 1.2 and Muse Spark 1.1 — which is the closest thing to Meta's reasoning family a team can call through us today, and Claude Opus 5.5 and DeepSeek V4 Pro among the models Muse Wonder is priced against.

The reason a single endpoint matters more than usual for a comparison like this one is that both models are moving targets. Argon's status is a vendor product decision that could change on any given day; Muse Wonder's configuration surface and pricing are a day old and can shift as providers pick it up. OrcaRouter serves 200-plus models behind one OpenAI-compatible API with provider list prices passed through unchanged, so a model reaching the catalogue is a string change in a config file rather than a new integration — and when a vendor cuts a price, the number on our side moves the same day. When either of these two becomes routable, measuring them against each other costs one afternoon.

What to do with this in the meantime

• If you need a model now, choose Muse Wonder and treat the Argon column as a future migration candidate. Nothing in Argon's published numbers is reachable today.

• If you are planning rather than shipping, keep the Argon cost figures in your model as a target rate — $1.99 per index task is the number Muse Wonder will be judged against the moment Argon opens.

• Use Muse Wonder's effort ladder deliberately from day one. Defaulting everything to medium is how a model with a $20-per-million output rate produces a surprise invoice.

• Re-check three things when Argon becomes available: its reasoning-effort controls, whether video input is in the shipping model, and its first independent throughput measurement. Any of the three can move the conclusion.

• Do not plan a budget on either cost-per-task figure yet. Muse Wonder's comes from a two-day-old model and Argon's from a model nobody outside a partner program has called.

A matchup where the cheaper, slightly higher-scoring model cannot be reached is not much of a matchup — but it will be, and soon. The pricing spread between these two is wide enough that the answer is likely to change the day Argon's endpoint goes live, which makes this the comparison to re-run rather than the one to settle.