What Is Muse Wonder Full? Meta's #2 Flagship

Watermelon Ships at #2: A First Look at Muse Wonder Full

Muse Wonder Full is Meta's most capable frontier large language model, released on October 12, 2026 by Meta Superintelligence Labs and available only to Muse Code subscribers, not through the public API. It scores 56 on the independent Artificial Analysis Intelligence Index, placing second behind Claude Opus 5.5's 58 and ahead of GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 (each at 53). It runs at maximum reasoning effort, supports a roughly 1-million-token context window, and takes text, image, audio, and video as input while returning text.

Updated 2026-10-13: the first independent hands-on benchmarks are in, and they keep Anthropic ahead overall even as Meta widens its own claims (see the Update below).

Key Takeaways

• Muse Wonder Full is the larger of two models in Meta's new Muse Wonder family; the smaller Muse Wonder is the API version.
• It scores 56 on the Artificial Analysis Intelligence Index, ranking #2 behind Claude Opus 5.5 (58) and tying Claude Sonnet 5.5.
• It is not on the Meta Model API — access is restricted to paid Muse Code subscribers.
• Two coined capabilities define it: Active Memory (cross-conversation recall) and Active Focus (long-horizon persistence).
• It reportedly leads on long-horizon coding, scoring 57.5 on SWE-Marathon versus Claude Opus 5.5's 46.3, per Meta's own benchmark table.
• No open weights. This is a proprietary release, a departure from Meta's Llama lineage.

What Is Muse Wonder Full?

Muse Wonder Full is the flagship member of the Muse Wonder series, Meta's first genuine attempt at a frontier model after the lighter-weight Muse Spark line. Meta describes the family as "our most capable models to date," built for "personal and professional agentic tasks" — reading a lease, making sense of an insurance claim, or shipping a working app.

The release is confirmed and public. Meta Superintelligence Labs, the research group led by former Scale AI CEO Alexandr Wang, announced both Muse Wonder and Muse Wonder Full on October 12, 2026. According to TechCrunch's launch coverage, the models were trained at Meta's Prometheus data center in central Ohio and mark the company's first state-of-the-art release since Wang restructured its AI efforts roughly 14 months earlier. The internal codename circulating before launch was "Watermelon."

The split between the two models matters. Muse Wonder (model ID muse-wonder-1.0) is the standard version, available on the Meta Model API and in Muse Code. Muse Wonder Full is the bigger sibling, running at max reasoning for "the hardest, longest-running work," and Meta serves it only to Muse Code subscribers. Setting the API reasoning level to max does not get you Full.

Muse Wonder Full at a Glance

Attribute	Detail
Developer	Meta (Meta Superintelligence Labs)
Type	Proprietary frontier large language model
Modality	Input: text, image, speech/audio, video · Output: text
Context window	~1M tokens (1,048,576)
Reasoning	Yes; effort scales from "minimal" to "max" (Full runs at max)
Pricing	No published API price; bundled with Muse Code subscription (plan prices reported by community, not officially confirmed)
Availability	Muse Code subscribers only — not on the Meta Model API
License	Proprietary; no open weights
Release date	October 12, 2026
Artificial Analysis Index	56 (#2 overall)

How Muse Wonder Full Works / What Makes It Different

Meta's pitch is not raw chatbot intelligence. It is an agent that stays useful across a long job. Two named capabilities carry that argument.

Active Memory lets the model carry what matters about you across conversations and pull it back when a new task needs it. Meta measures this with its own MemoryBench, which salts long chats (64,000 to 100,000 tokens) with distractors plus facts that contradict earlier context and get updated partway through. Muse Wonder Full scores 85.5 there, with the standard Wonder at 85.6 and Claude Opus 5.5 at 81.5, per Meta's figures.

Active Focus keeps the model pursuing a goal when it hits obstacles — trying alternative approaches, working through blockers, and asking permission before consequential actions. Meta tests this on its own FocusBench, spanning housing, travel, IT, and personal finance tasks across multi-turn interactions.

On the coding side, Meta leans on the Personal Superintelligence Index (a self-built agentic eval) and long-horizon suites. The widest reported gap is on SWE-Marathon, which asks an agent to build an entire software project — a C compiler, a Slack clone — over 2 to 10 hours, counting a task only if every hidden check passes. Muse Wonder Full reportedly scores 57.5 there against Claude Opus 5.5's 46.3.

Source: @Ananth7e

One caution belongs here in confident plain terms: most of these comparison numbers come from Meta's own model-card table, run at max effort, and have not been independently replicated. One commentator reviewing the table noted that only two benchmarks included a Claude score and that Meta ran the tests itself.

What You Can Do With Muse Wonder Full

The demos Meta showed lean personal and professional. Reported and demonstrated uses include:

• Long coding runs — multi-hour agent sessions that plan work, call tools in parallel, hand parts to sub-agents, and recover from errors.
• Legal and document-heavy work — on the Harvey legal agent test, Meta reports 13.1 versus a rival best of 8.6, though pass rates are low across all models.
• Personal planning — building a Saturday running plan from your preferences, arranging wedding seating, or customizing a gift from Instagram photos.
• Business automation — running full workflows on Meta's AutomationBench, where it reports 45.4.

One hands-on tester reported building a Three.js castle in about 7 minutes 12 seconds, with a roof correction taking 58 seconds — a single anecdote, not a benchmark.

How Muse Wonder Full Compares

On the independent Artificial Analysis Index, the ranking is tight at the top. These figures are confirmed by Artificial Analysis's model page and VentureBeat's report.

Model	AA Intelligence Index	On public API?
Claude Opus 5.5	58	Yes
Muse Wonder Full	56	No (Muse Code only)
Claude Sonnet 5.5	56	Yes
GPT-6 Astra	53	Yes
Gemini 4 Argon	53	Yes
Claude Fable 5.1	53	Yes
Muse Wonder (standard)	52	Yes

The nuance: Muse Wonder Full ties Claude Sonnet 5.5 on the overall index, and Sonnet 5.5 is served publicly at roughly half the cost. Where Full reportedly wins is long-horizon agentic coding; where it trails Opus 5.5, per Meta's own card, is terminal coding (57.1 vs 59.6), from-scratch programming (59.0 vs 70.5), computer use (78.4 vs 81.8), and health Q&A (62.3 vs 65.6). For a closer read on that two-point gap, see our comparison of Muse Wonder and Claude Opus 5.5.

Source: @0x0SojalSec

If you want an API-accessible frontier model with similar pricing and a confirmed public endpoint, Claude Opus 5.5 is the model Muse Wonder's $4/$20 standard tier was benchmarked against — and the one Full trails by two index points.

Availability: How to Access Muse Wonder Full

Muse Wonder Full ships inside Muse Code, Meta's AI programming assistant, for paid subscribers. It is not callable from the Meta Model API. The standard Muse Wonder model is the API option, priced at $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20 per million — the same rate card as Claude Opus 5.5.

Community reports describe Muse Code plans at $50 (Power Usage) and $200 (Maximum Usage) per month, but Meta has not officially published Wonder Full's subscription pricing. Meta also says Muse Wonder will "soon" power its consumer agent product, Muse, starting with subscribers. Our earlier breakdown of the 56-point climb tracks how the standard model fits the API lineup.

Early adopters flagged heavy quota consumption: one user reported a single task burning 103% of a five-hour limit on the $50 plan. Alexandr Wang later posted that Meta had "identified and fixed" an issue causing unusually fast quota usage and reset Muse Code limits.

What We Don't Know Yet

• Muse Wonder Full's API pricing — unconfirmed; Meta has published no per-token rate because it is subscription-only.
• Exact Muse Code plan entitlements — the $50/$200 figures come from community posts, not Meta's docs.
• Independent benchmark replication — the headline coding wins rest largely on Meta-run evaluations; no broad third-party confirmation yet.
• Open weights — Meta's card mentions only that weights are protected; there is no statement on a future open release.
• The full safety report — Meta says a complete evaluation with additional safety data is still to come.

Frequently Asked Questions

What is Muse Wonder Full?

Muse Wonder Full is Meta's most capable frontier large language model, released October 12, 2026 by Meta Superintelligence Labs. It runs at maximum reasoning effort for long, hard agentic work and is available only to Muse Code subscribers, not through the Meta Model API.

Is Muse Wonder Full available on the API?

No. Muse Wonder Full is not served through the Meta Model API. Meta reserves it for paid Muse Code subscribers; the standard Muse Wonder model is the API-accessible version at $4/$20 per million tokens.

How much does Muse Wonder Full cost?

Meta has not published a per-token API price for Muse Wonder Full because it is not sold through the API. Access comes bundled with Muse Code's paid subscription tiers; community reports put those plans at $50 and $200 per month, which Meta has not officially confirmed.

Is Muse Wonder Full open source?

No. Muse Wonder Full is a proprietary model with no publicly downloadable weights, a break from Meta's earlier open-weight Llama models. The model card says the weights are protected under Meta's weight-security program.

What is the difference between Muse Wonder and Muse Wonder Full?

Muse Wonder is the standard model, available on the Meta Model API and in Muse Code, scoring 52 on the Artificial Analysis Intelligence Index. Muse Wonder Full is the larger, more capable sibling running at max reasoning, scoring 56, and is limited to Muse Code subscribers.

Muse Wonder Full vs Claude Opus 5.5 — which is better?

Claude Opus 5.5 leads the overall Artificial Analysis Intelligence Index at 58 versus Muse Wonder Full's 56. Muse Wonder Full reportedly leads on long-horizon coding benchmarks like SWE-Marathon and DeepSWE, while Opus 5.5 stays ahead on terminal coding, from-scratch programming, and health tasks.

Who made Muse Wonder Full?

Muse Wonder Full was built by Meta Superintelligence Labs, the AI organization led by former Scale AI CEO Alexandr Wang. The models were trained at Meta's Prometheus data center in central Ohio.

What to Watch Next

Three signals will decide whether the launch hype holds. First, independent benchmark replication — the SWE-Marathon and DeepSWE leads are Meta-run until a third party confirms them. Second, whether Muse Wonder Full ever reaches the API or stays locked to Muse Code, which shapes who can actually build on it. Third, how the quota mechanics settle after Meta's usage-accounting fix, since early testers hit five-hour limits fast.

Update — 2026-10-13

The third-party replication the article flagged as pending has started to arrive, and the early read is more nuanced than Meta's card. In a hands-on 105-bug test, @PawelHuryn scored Wonder Full (max) at 44 — narrowly ahead of Claude Fable 5.1 (43) and Claude Opus 5.5 (41.7), but behind GPT-6 Astra (45). A leaderboard from @AskClash has Opus 5.5 leading overall 80.1 to 71.7 and winning five of six shared tests, with Full's clear win confined to DeepSWE (79.0 vs 74.2). The pattern is consistent: Full's independent edge is real but narrow and concentrated in long-horizon coding, while Anthropic stays ahead across the broader board.

Meta has since widened its framing, claiming Full beats Opus 5.5 on 8 of 14 published benchmarks (@Mikadzyki_NFT) — still a vendor-run tally that the independent tests above do not corroborate. On access and quality, reception splits: @QuixiAI called Muse Code "Fable-level" and roughly twice as fast with more generous limits, enough to replace Fable 5.1 for hard tasks, while other testers report 3D and from-scratch detail still trailing Opus 5.5 and GPT-6.1 Sol.

On economics, a single cost analysis from @StragglerLiu argues Full is locked behind subscriptions because of heavy token burn — pegging standard Wonder at roughly $4.52 per task against GPT-6 Astra's $3.26, consuming about 7.69M input tokens per task. Those figures come from one account and remain unconfirmed elsewhere.

Try related models on Kie.ai: Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1 (Coming Soon).