Meta Muse Wonder Review (2026): Pricing, Benchmarks & What You Actually Get

Meta's first real frontier model, shipped under the codename Watermelon after a fourteen-month rebuild. The memory numbers genuinely beat Claude Opus 5.5. The intelligence score does not. And the model that wins Meta's own benchmark table is the one you cannot buy through the API.

By AI Tool Scout · Published October 13, 2026 · 17 min read

What Are Muse Wonder & Muse Wonder Full?

Muse Wonder and Muse Wonder Full are the first genuinely frontier-class models from Meta Superintelligence Labs (MSL), announced on October 12, 2026 by chief AI officer Alexandr Wang, with Mark Zuckerberg personally promoting the release. Internally the family was codenamed Watermelon — Zuckerberg teased it with a watermelon emoji back on September 2, when Muse Spark 1.3 shipped. They were trained at Meta's Prometheus data centre in central Ohio, the company's gigawatt-scale cluster, with a second site, Hyperion in Louisiana, scaling toward five gigawatts.

This is the payoff of a deliberate, expensive rebuild. After the Llama era left Meta behind on frontier capability, the company put $14 billion into Scale AI and installed Alexandr Wang as Chief AI Officer to lead MSL, then spent roughly fourteen months restructuring the AI organisation and committing well over $100 billion in capital expenditure on compute — much of it decided before anyone knew whether the models would work. Wang and Zuckerberg have since described the whole wager openly on the record. Muse Wonder is the first model Meta is willing to put head-to-head with OpenAI and Anthropic without an asterisk, and the first it is openly charging frontier prices for.

The two checkpoints are not two products so much as two access tiers. Muse Wonder (muse-wonder-1.0, in preview) is the model you can actually call: it is on the Meta Model API and it is included in every Muse Code plan. Muse Wonder Full is the larger, stronger sibling — and it is not offered on the API at all, running only inside Muse Code on two subscription tiers. Both share a 1,048,576-token context window, accept text, image, video, audio and PDF input and return text, and support up to 943,718 output tokens. Both are proprietary; no open-weights version of Muse Wonder has been announced. If you want the wider model landscape for context, our Claude review covers the family that currently tops the index.

Key Features: Active Memory, Active Focus, and Agent Reliability

Meta's whole pitch for these models is that raw benchmark chasing is the wrong goal for an agent. The company's framing is that real work unfolds over time — so the useful capability is remembering and persisting. Two named mechanisms carry that argument.

• Active Memory. Muse Wonder carries important context about you across separate conversations and then actively retrieves the detail relevant to the current task, rather than waiting for you to repeat yourself. Meta describes this as in-model memory rather than an external retrieval layer bolted onto the agent. It is the single most defensible claim in the launch: Muse Wonder's 85.6 on MemoryBench — a test that checks whether a model can pull the right detail out of a long, fact-dense conversation — beats Claude Opus 5.5 (81.5), GPT-6 Astra (81.4) and Claude Fable 5.1 (79.0). For anyone building an assistant that has to hold context over weeks, that is the number that matters.
• Active Focus. This is about persistence rather than recall. When a task gets hard, Muse Wonder keeps working — trying alternative approaches, working through blockers, recovering from its own errors and following through to a finished deliverable. Crucially, it also knows when to stop and ask: Meta says it asks for clarification when a prompt is ambiguous and checks in before taking consequential actions. Meta evaluates this through FocusBench (persistence in changing environments) and SWE-Marathon (long-running software projects), where Wonder Full scores 57.5 against 46.3 for Claude Opus 5.5.
• Long-horizon coding strength. The family is explicitly built for agentic coding and computer use. Muse Wonder scores 78.8 on DeepSWE 1.1 against 74.2 for Opus 5.5, and Muse Wonder Full reaches 68 on the independent Artificial Analysis Coding Agents Index at max effort inside Muse Code — level with Claude Code running Sonnet 5.5 (68) and ahead of Opus 5.5 (66).
• A million-token context with no long-context penalty. Both models take a 1,048,576-token window and Meta charges the same token rate no matter how much of that window you use — no surcharge for long prompts, which several rivals still apply. Cached input is heavily discounted at ~$0.20 per million.
• Multimodal input. Text, image, video, audio and PDF in, text out. Reasoning is always on and cannot be switched off; the effort ladder runs minimal, low, medium, high, xhigh and max, defaulting to medium, and reasoning tokens are billed at the output rate — a detail that quietly inflates real costs.
• Genuine agent safeguards. Meta says the models were trained not to share private information without permission and to refuse harmful requests, borrowing the approval-and-monitoring architecture built for the Muse agent. It also notes that a fuller safety evaluation report "will be available soon" — which, at launch, it was not.

Benchmarks vs Reality: What Meta Claims and What Testers Found

This is where the launch gets interesting, because there are three different stories in circulation: Meta's own table, the independent evaluator, and people who actually ran the model.

What Meta's own table claims

Meta published a fourteen-row comparison against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, all at maximum reasoning effort. Treat every cell below as vendor-reported and not independently reproduced — Meta says so itself, and its model card notes that the rival SWE-Marathon numbers come from Codex and Claude Code harnesses, so that row is not a controlled same-harness comparison.

Read carefully, the table says two things. First, memory and long-horizon coding are genuine wins — Muse Wonder beats every rival on MemoryBench, and the family is well clear on SWE-Marathon. Second, on almost every other row the winner is either Claude Opus 5.5 or Meta's own subscription-only Full model. Opus 5.5 keeps a decisive lead on ProgramBench (70.5 vs 50.0 for Wonder), stays ahead on Terminal-Bench and OSWorld, and leads HealthBench Pro. Meanwhile Wonder Full out-scores standard Wonder across the board, including a striking gap on business workflows. That matters because Wonder Full is the model you cannot call via API.

What the independent evaluator says

• Artificial Analysis Intelligence Index v4.3: Muse Wonder (max) 52, Muse Wonder Full (max) 56, Claude Opus 5.5 (max) 58, Claude Sonnet 5.5 (max) 56, Claude Fable 5.1 53, GPT-6 Astra 53, Muse Spark 1.3 (max) 48. So Full is joint second, tied with Sonnet 5.5 and still two points behind Opus 5.5, while the API model — the one developers actually buy — sits one point below both Astra and Fable 5.1.
• Artificial Analysis calls it a material jump. Going from 48 to 52/56 in six weeks is a genuinely unusual release cadence, and AA reports the Full model's gains came across all ten of its evaluations, with the largest improvements in agentic tasks.
• Coding agents: Muse Wonder Full hits 68 on the AA Coding Agents Index at max effort in Muse Code, level with Claude Code plus Sonnet 5.5 and above Opus 5.5 at 66. Muse Wonder (max) scores 63.
• Cost per task favours Meta, but not by much. Because Claude models are famously verbose, AA's measured cost per index task is roughly $7.63 for Fable 5.1, $5.98 for Opus 5.5 and $5.46 for Sonnet 5.5, versus about $4.52 for Muse Wonder — cheaper than Anthropic despite identical rate cards. But it is more expensive per task than GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99) or Muse Spark 1.3 ($1.60). Since Wonder Full has no per-token price, no cost-per-task figure exists for the model with the better scores.
• Not everything Meta cites is live yet. Some figures Meta's materials reference had not appeared on Artificial Analysis's public pages on launch day, and no independent hands-on evaluation had been published when the numbers first circulated.

What the community found — and it was more mixed than the charts

• The marketing page mixes the two models. The most-upvoted observation on the launch-day Hacker News thread was that the benchmark scores displayed on Meta's main Muse Wonder page are Wonder Full's scores, sitting next to Wonder's price. That is a fair-and-square catch: the model behind the headline numbers is not the one the API sells.
• An image-to-HTML test was "much worse than Opus 5.5." One launch-day tester's verdict was that it performed closer to a mid-tier Claude model than to the frontier, and another described it as "still an entire generation behind in real-world performance." Treat individual prompts as anecdotes — but the direction matched the independent index, which is what makes them worth citing.
• A blind four-build comparison put Wonder Full last on all four. An independent reviewer ran Wonder Full against Opus 5.5, Fable 5.1 and GPT-6 Astra on the same four builds — a website clone, a launch campaign, an iPhone app and a 3D simulation — and ranked the results blind. Opus 5.5 came first on three, Fable 5.1 won the 3D test, GPT-6 Astra was consistently competitive, and Muse Wonder Full finished fourth every time. On the website clone it "ignored most of the site it was given and made up its own sections."
• Style complaints. A separate head-to-head on SVG generation called Wonder Full's output "good, and it even tries 3D generation" but "too gimmicky" and prone to over-indulging — a restraint problem rather than a capability one, and one worth noting because it is exactly the kind of thing a benchmark will never show you.
• Rough early demos. Creators generating 3D scenes and small games reported flickering, clipping and shadow artifacts, plus a "hyperrealistic" watermelon render that came out blurry. One tester reported an unprompted message from Wonder Full telling them not to investigate its development lineage — an output with no basis in the conversation.
• Tooling was the real complaint. Multiple developers said Muse Code's sandbox never worked across Mac or Linux on stock installs, that paid sessions threw errors, and that one run exhausted a five-hour limit plus 38% of a weekly allowance. Sign-in issues in containers were common enough that another vendor hit the same OAuth device-flow wall publicly on day one.
• The counter-evidence is real too. Several developers were enthusiastic — one said 90% of their work had already moved off OpenAI's Codex onto Muse Wonder Full, and Meta's own docs now support pointing Claude Code's CLI at the Meta Model API for a same-harness bake-off.

The honest synthesis: Muse Wonder is excellent at exactly what it claims to be excellent at — memory and long-running agent work — and mid-frontier at everything else, at frontier prices. If your job is a personal assistant that remembers across weeks, or a coding agent that runs for hours, it is a legitimate first choice. If your job is one-shot hard programming, the same money buys Opus 5.5, and the evidence is that you should take it.

Pricing in 2026: Three Price Books, One Model Family

Meta published three separate price books on launch day, and the same two model names appear in all of them. Every figure below is approximate, in US dollars, checked against Meta's own API and Muse Code pages and launch-day coverage on October 13, 2026 — always confirm current pricing on Meta's official pages before you subscribe, because these are brand-new products with regional variation and active promotions.

What is missing from all three tables: a per-token price for Muse Wonder Full. There is not one. Meta does not sell its strongest checkpoint by the token, which means a ~$50/month subscription cannot be converted into a dollars-per-token comparison against any rival — and the plan's allowance is expressed as a multiple of an unpublished baseline (20x or 80x the Everyday plan's 10–50 prompts), not as a token count. Reviewer arithmetic on Meta's own advertised numbers does suggest Power is not a premium rate — it works out cheaper per unit of Everyday usage than High — so you are buying volume, with Full riding along.

A footnote worth reading twice: choosing maximum reasoning on the API does not turn Muse Wonder into Muse Wonder Full. It changes how hard the model thinks, not which checkpoint you are talking to. Any comparison that does not state its effort setting is comparing a throttled run against an unthrottled one. For the wider market, see our LLM API cost guide and the GPT-6 Astra review, whose $10/$50 rate is more than double Wonder's.

Which One Do You Get? Muse App vs Muse Code vs API

This split is the most confusing part of the launch, so here it is plainly.

Two consequences fall straight out of that table. First, the best model is behind a paywall that cannot be bought per token — if your workload genuinely needs Wonder Full, the cheapest door is the ~$50 Power plan, and you will be running it inside Meta's coding agent rather than your own harness. Second, the API and the subscription have opposite data terms: the API runs Standard mode with no training on your content, while the Muse Code plans that unlock Full default to Contributor mode. If your code is confidential, the settings toggle is not optional.

Pros & Cons

5 Best Muse Wonder Alternatives

Comparison Table

Which One Should You Choose?

• Choose Muse Wonder if memory and long-running agents are the actual job. If you are building an assistant that has to recall context across weeks, or a coding agent that runs for hours, the MemoryBench and DeepSWE leads are real, the $5 Everyday plan is a cheap way in, and the API's Standard mode keeps your data out of training.
• Choose Claude Opus 5.5 if you want the best model at this price. At identical token rates it scores 58 against Muse Wonder's 52, and it is available per token rather than locked to a subscription. This is the single most important decision in this review.
• Choose the ~$50 Muse Code Power plan if you specifically need Muse Wonder Full. It is the only affordable way in, and it comes with 20x the Everyday allowance — but run it with Contributor mode off.
• Choose Gemini or DeepSeek if the budget is the constraint. A free tier and sub-$1-per-million rates will cover most routine work; Muse Wonder's price only makes sense when its specific strengths apply.
• Wait if you are in Hong Kong, the EU or Asia. The Muse app route is closed outside the US and Canada, and Meta has named no date. The API is the only realistic door today, and even that varies by region.
• Watch two dates. The Muse Spark open-weights release Meta promised in September is still outstanding, and a wider Muse app rollout has not been scheduled. Both would materially change this recommendation.

Frequently Asked Questions