Meta Watermelon Is Muse Wonder: What the October 12 Launch Actually Confirmed

Meta Watermelon was Muse Wonder all along, and Meta released it on October 12, 2026 — generally available as muse-wonder-1.0 on the Meta Model API and inside Muse Code, at $4.00 per million input tokens and $20.00 per million output tokens. That price is the claim doing the rounds: 60% below Claude Fable 5.1, which the Claude maker lists at $10.00 in and $50.00 out, while Meta's launch material puts Muse Wonder ahead of it on benchmark after benchmark. It is also exactly the same $4/$20 as Claude Opus 5.5, the model currently sitting first on the independent intelligence index. The pricing half of that story survives arithmetic. The performance half survives only for one of Meta's two checkpoints — and the checkpoint a developer can actually call today is not that one.

This piece started life in September as a leak write-up of a watermelon emoji appended to a Muse Spark 1.3 launch post. The emoji has a product name now, so what follows is the same reporting re-read against a shipped release: what launched, what the price really is, what the benchmarks show once the vendor column is separated from the independent one, and the open-weights half of that September promise, still unfulfilled six weeks on. The four images on this page were built for the pre-launch version — the hero card still reads as a leak report, the confirmed-versus-unverified board was drawn when the codename was inference rather than fact, and the leaderboard capture is a September 6 snapshot on an earlier revision of that index. They are kept as the record of what was known then, and no claim below rests on their figures.

The codename is settled, and Meta's own launch material never mentions it

CNBC's launch coverage reports that Muse Wonder was the model Meta referred to internally as Watermelon, and that mapping lines up with a reporting trail running back months: The Information reported in August that Meta was targeting October for a model it called Watermelon, without establishing whether it would ship under the Muse name. It shipped as Muse Wonder, on the reported schedule, which is the resolution of the tease this article was originally written around. Note the provenance carefully. The mapping comes from CNBC's reporting of Meta's internal naming, not from Meta's launch material, which announces a Muse Wonder series and says nothing about fruit. The September 2 post that started all of this ended with "Next up 🍉 and Muse Spark open weights releases coming soon"; the first half of that line is now a shipped product, and the second half is not.

The consumer context landed too. The same August reporting tied Watermelon to a paid agent push, with a premium subscription reported at up to $199.99 a month, and the launch confirms the direction rather than the exact number: Muse Wonder reaches the Muse app only on paid tiers, at $20 or $100 a month, while Muse Wonder Full requires a Muse Code subscription at $50 or $200 a month.

What shipped on October 12

Meta released two checkpoints under one announcement, and the difference between them is the most consequential thing about the launch.

• muse-wonder-1.0 — the checkpoint served on the Meta Model API and inside Muse Code, Meta's terminal coding agent. This is the model anyone can call today.

• muse-wonder-full-1.0 — a separate, higher-scoring checkpoint that Meta describes as Muse Code subscription only. It is not a reasoning-effort flag on the API model; it is a different set of weights, reachable through a monthly product.

• Price — $4.00 per million input tokens, $0.20 for cached input, and $20.00 per million output tokens. Meta publishes the same rates for both checkpoints, so the subscription wall is not a pricing tier. It is the only way to reach the better model.

• Context — 1,048,576 tokens, with an output budget that caps at 943,718 tokens.

• Inputs and outputs — text, image, speech and video in; text out. Reasoning is mandatory rather than optional, with an effort ladder running from minimal to max and a default of medium.

• Named features — Active Memory, described as persistent state carried across sessions, and Active Focus, described as the model's own attention management on long tasks and a willingness to ask for clarification rather than stall. Both are Meta's descriptions of its own model, and neither is measurable from a model card.

• Consumer rollout — Meta says Muse Wonder will "soon power experiences in Muse" starting with subscribers. There is no date, and no statement about whether the consumer surface gets the same checkpoint.

What is confirmed, what is vendor-stated, what is still open

Separating those three is the whole job of this piece, and the launch let the first list grow. The board below was drawn for the pre-launch version — its left column is what was checkable in September, its right column what was inference then — and the confirmed list now reads differently.

• Confirmed — Meta released Muse Wonder on October 12, 2026, generally available on the Meta Model API and inside Muse Code, with Muse Wonder Full restricted to Muse Code subscribers.

• Confirmed — Muse Wonder's published rates are $4.00 per million input tokens and $20.00 per million output tokens, with cached input at $0.20, and the same rates are published for Muse Wonder Full.

• Confirmed, by arithmetic — those rates are 60% below Claude Fable 5.1's $10.00/$50.00 and identical to Claude Opus 5.5's $4.00/$20.00.

• Confirmed, independently — Artificial Analysis measures Muse Wonder (max) at 52 and Muse Wonder Full (max) at 56, against 53 for Claude Fable 5.1 (max with fallback) and 58 for Claude Opus 5.5 (max).

• Reported, not vendor-stated — that the model Meta shipped as Muse Wonder was internally codenamed Watermelon. The mapping comes from CNBC's launch reporting of Meta's internal naming, and Meta's own material does not use it.

• Vendor-stated — every benchmark in Meta's launch table, including the 59.1 JobBench figure that belongs to the subscription-only checkpoint and the 59.0 index figure that Artificial Analysis publishes as 56. Meta's claim that it has reached the frontier is Meta's claim.

• Open — the weights question for Muse Wonder and Muse Wonder Full, the same question for Muse Spark 1.3, and the date for Muse Spark 1.2's promised open-weight release. "Soon" is not a date, and a promise is not a registry entry.

The 60% cheaper claim, checked against the rate cards

The comparison in circulation is against Claude Fable 5.1, and it survives contact with the numbers. The published rate for Claude Fable 5.1 is $10.00 per million input tokens and $50.00 per million output tokens — the same figures Artificial Analysis records on its card for the model, alongside a 98% cache discount. Against $4.00 and $20.00, that is exactly 60% off on both legs: not an approximation of "about 60%", but the same 0.6 factor applied to input and to output. For a workload that is mostly input, or mostly output, the saving comes out the same either way, which is unusually clean as price comparisons go.

Two corrections are worth making to how that number is being repeated. The first is that the 60% is against Claude Fable 5.1, not against the frontier generally — Claude Opus 5.5, the model Artificial Analysis currently ranks first on intelligence, is also $4.00 in and $20.00 out, so Meta has not undercut the top of the board. It has matched it. The second is that a lower token rate is not the same thing as a lower bill, and the section on real cost below works that through.

The outperformance claim, checked against the independent index

Meta's launch material is generous with benchmarks. Its own table puts Muse Wonder Full at 59.1 against Claude Fable 5.1's 52 on Meta's "Personal Superintelligence Index", at 75.3 against 27.6 on JobBench, and at 59.1 against 6.4 on Vals Finance Agent V2. Every one of those figures is vendor-reported, produced in Meta's harness, and unreproduced by anyone else — read them as Meta's claim, not as measurement. Meta's table also lists Muse Wonder Full at 59.0 on the Artificial Analysis Intelligence Index, where Artificial Analysis itself publishes 56 for the same checkpoint. When a vendor and the independent evaluator disagree by three points on the evaluator's own index, the evaluator's number is the one to plan against.

Artificial Analysis has re-based this index more than once since September, so the leaderboard card above is a September 6 capture from an earlier revision and its values are not comparable to the current ones. What the evaluator publishes today splits the claim along the checkpoint line:

• Muse Wonder (max) — 52, ranked ninth of 219 models, at $4.00 in and $20.00 out with a 95% cache discount, working out to about $4.52 of spend per index task.

• Muse Wonder Full (max) — 56, ranked fourth of 699 models on the evaluator's own card. No per-token price is published for it, because it is not sold per token.

• Claude Fable 5.1 (max with fallback) — 53, ranked fifth of 219, at $10.00 in and $50.00 out.

• Claude Opus 5.5 (max) — 58, ranked first of 219, at $4.00 in and $20.00 out.

So the headline resolves into something more precise than "better and cheaper". Muse Wonder Full at 56 does beat Claude Fable 5.1 at 53 — by three points, on the independent index, while costing 60% less per token. Muse Wonder at 52 does not: the checkpoint sold on the API sits a point below Claude Fable 5.1 and six points below Claude Opus 5.5. A reader who wants the claim to be true has to buy the subscription. A reader who wants an API key gets the model that loses the comparison.

The independent numbers also find the weak spot Meta's own table does not lead with. On Artificial Analysis's factual-knowledge evaluation, Muse Wonder Full scores 30 and Muse Wonder 25 — well behind Claude Opus 5.5 at 46, GPT-6 Astra at 43 and Gemini 4 Argon at 42, and consistent with the series being built on a smaller base model than the closed frontier. A low hallucination rate partly offsets it: the same evaluation puts Muse Wonder Full at a 23% hallucination rate, the second-lowest among models scoring at least 45 on the index, with Gemini 4 Argon lower still at 15%. The strength is agentic. Muse Wonder Full lands at 68 on the evaluator's coding agents index, tying Claude Sonnet 5.5, and its improvement over Muse Spark 1.3 on Terminal-Bench 4.0 is described as roughly doubled. Meta addresses the standing criticism that its models are tuned for benchmarks by arguing that the new base model generalises to tasks outside its training optimisation — a vendor claim about its own model, and one that only repeated public use will settle.

What the cheaper rate does to a real bill

Here is where the 60% headline gets smaller. Reasoning is mandatory on Muse Wonder and reasoning tokens bill as output, so the output leg is where the invoice is decided. Artificial Analysis records 130 million output tokens consumed during its evaluation run, against an 85 million median for comparable models — a verbose model, though less verbose than Claude Opus 5.5 at 260 million and about level with Muse Wonder Full at 140 million. That verbosity is already priced into the cost-per-task figure: $4.52 for Muse Wonder against $5.98 for Claude Opus 5.5 and $7.63 for Claude Fable 5.1. The rate card says 60% cheaper than Fable 5.1; measured per completed index task, the gap is closer to 40%, because Muse Wonder spends more tokens getting there.

Run the arithmetic on a concrete month — say 10 million input tokens and 2 million output tokens. Muse Wonder bills $40.00 of input and $40.00 of output, $80.00. Claude Fable 5.1 bills $100.00 and $100.00, $200.00. Claude Opus 5.5 bills $40.00 and $40.00, $80.00. Against Fable 5.1 the saving is real and large. Against the model that actually leads the index, Meta has priced to match rather than to undercut. And because reasoning cannot be switched off, the only cost levers are the effort level and the output budget — at medium effort on short calls this is a normal frontier bill, and at max effort inside a long agentic loop it is a large one.

The open-weights half is now nine weeks old

The second clause of Zuckerberg's September 2 sign-off — "Muse Spark open weights releases coming soon" — has not been honoured. Meta promised the Muse Spark 1.2 weights on August 10, 2026, which makes that promise sixty-four days old today, and a direct check of Meta's model organisation on Hugging Face this morning returns exactly four repositories, all of them Muse Glimmer variants: Muse Glimmer 30B, its GGUF and assistant builds, and an ExecuTorch export. There is no Muse Spark checkpoint of any version, no license, no hardware spec and no date. Muse Glimmer — the 30-billion-parameter Apache 2.0 model Meta released alongside the promise — remains the only downloadable open-weight model in the Muse family.

The launch did not clarify Muse Wonder's own distribution either. Meta has not said whether it will publish the weights for either new checkpoint. That omission carries consequences beyond developer preference: under Article 53 of the EU AI Act, a genuinely free and open model escapes some of the documentation it would otherwise owe, and the exemption depends on the licence being real, with weights, architecture and usage information public. Meta had not decided the same question for Muse Spark 1.3 either, and Zuckerberg has signalled caution about opening the most powerful systems. The practical read for anyone planning a self-hosted path is unchanged from September: the window keeps being described as "soon", the flagship keeps shipping closed, and the only weights on the registry are the 30B distilled one.

Where Muse Wonder is, and where OrcaRouter fits

Muse Wonder is not on OrcaRouter. There is no muse-wonder entry in our catalogue of 207 models, and this piece will not describe one as if there were. What we do carry from the same lab is the Muse Spark line, and the same rate-card logic applies to the comparison shopping this launch invites: both of Muse Wonder's pricing peers are already on our catalogue at their vendors' own list prices. Claude Fable 5.1 is there at $10.00 in and $50.00 out, Claude Opus 5.5 at $4.00 in and $20.00 out. Muse Spark 1.2 is on OrcaRouter at the provider's list price — $1.25 per million input tokens and $4.25 per million output tokens, passed through with 0% markup, so the rate on our side is Meta's own rate and moves the same day Meta moves it.

That pass-through matters more than usual this week, because this is a pricing launch as much as a capability one: a vendor's rate change shows up on our side without a contract renegotiation, and a model that undercuts one competitor by 60% while matching another makes the comparison worth running rather than reading. It is also the cheapest way to test the claim yourself. Muse Wonder's marketing rests on comparisons against models already callable through a single OpenAI-compatible endpoint, so the honest evaluation — same prompts, same harness, the alternatives side by side — costs one API key instead of three vendor accounts. And for a checkpoint this new, whose distinguishing feature is a subscription wall rather than a track record, automatic failover is the pattern that lets you try it without betting a production path on it: point a test route at the new model, keep a proven fallback on the same endpoint, and let the router move traffic back if the reasoning loop stalls or the bill runs hot.

What to watch next

Four things are unresolved, and each one changes how a reader would plan around this launch.

• Whether muse-wonder-full-1.0 gets an API path. The checkpoint that beats Claude Fable 5.1 by three points on the independent index is reachable only through a $50-to-$200 monthly subscription, which turns the headline comparison into a product decision rather than an engineering one. Meta has not said it will open it.

• Whether Meta publishes Muse Wonder's weights. It did not decide this for Muse Spark 1.3 either, and the answer determines both what a team can self-host and what Meta owes under the EU AI Act.

• A Hugging Face repository for Muse Spark 1.2's promised open weights, now sixty-four days overdue. Check Meta's model organisations on the Hub rather than waiting for an announcement — the registry moved before the blog post last time.

• Whether independent harnesses reproduce the JobBench and Vals Finance spreads, and whether Active Memory holds up outside a demo. A 44-point gap between two checkpoints from the same lab on one benchmark is either a genuine capability split or a harness artefact, and only third-party runs can settle which.

What the launch did settle is the question this article was written around in September. The watermelon emoji was a real flagship, it shipped on the reported schedule, and it shipped into the price bracket the frontier already occupies. The interesting part is no longer whether Meta could reach the frontier. It is that the model Meta priced to match the leader is not the model Meta benchmarked against it.