Meta shipped its most capable models to date on October 12, 2026, and this time it is charging frontier prices to match. Muse Wonder and Muse Wonder Full arrive from Meta Superintelligence Labs, built on the foundations of Muse Spark, and within hours of launch they had climbed into the top handful of the independent leaderboards.
The split is simple. Muse Wonder is the model anyone can use today, through an API or a coding tool, while Muse Wonder Full is the one you have to subscribe to reach. Almost everything that matters about this release, the benchmarks, the prices, and the access rules, lives in that gap.
Meta’s announcement framed the pair as its most capable models yet, built for “personal and professional agentic tasks” that range from making sense of a lease to building a working app. What changed is how Meta chooses to sell them.
From Muse Spark to Muse Wonder: What Changed
Meta has been climbing toward the frontier since Muse Spark, which was already trading wins with the top models on coding benchmarks when it shipped in September. So Muse Wonder is not Meta’s first competitive model. It is the next rung up the ladder, and it behaves like it: bigger, slower, and priced like a flagship rather than a value play.
The contrast with Spark is sharp. Muse Spark 1.3 launched with contributor pricing at $0.10 per million input tokens and $0.20 per million output tokens, a deliberate attempt to undercut rivals by an order of magnitude. Muse Wonder starts at $4 and $20 for the same units, roughly a forty-fold jump on input pricing and an even larger one on output.
Meta says the two models were trained at its Prometheus data center in central Ohio and developed under the AI division led by chief AI officer Alexandr Wang, who was hired in June 2025 and restructured Meta’s AI efforts that August. The models were internally codenamed “Watermelon,” a detail Mark Zuckerberg teased during an Instagram livestream over the weekend before launch.
Meta’s framing is grand: the pair marks “the next step on our scaling ladder and a major leap toward personal superintelligence,” aimed at expert domains including law, finance, and health, alongside frontier coding for people building everything from a game to a small business app.
Muse Wonder vs Muse Wonder Full: What’s the Difference?
This is where the launch gets unusual. Meta is not selling one model at two speeds. It is selling one model broadly and holding the better one back behind a coding subscription.
| Feature | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Meta Model API access | Yes | No |
| Muse Code access | Yes, all plans | Power Usage and Maximum Usage plans only |
| Context window | 1M tokens | 1M tokens |
| Max output | 131k tokens | 131k tokens |
| Input price (per 1M tokens) | $4.00 | Not sold per token |
| Cached input (per 1M tokens) | $0.20 | Not sold per token |
| Output price (per 1M tokens) | $20.00 | Not sold per token |
Muse Wonder Full is the more capable model, and it is available only through a Muse Code subscription. Because it has no per-token price, independent trackers have no cost-per-task figure for it, which is exactly what you would expect when a model is bundled rather than metered.
Muse Wonder is the accessible sibling. It runs on the Meta Model API and inside Muse Code, and Meta says it will soon power more of the Muse consumer app, rolling out first to subscribers, with access there tied to a monthly subscription of $20 or $100.
For developers, adoption is meant to be frictionless. Muse Wonder is OpenAI SDK compatible, so an existing project can switch with a single base_url change and set the model to muse-wonder-1.0. Meta’s developer blog frames it as keeping your code and pointing it at a new endpoint.
The Benchmarks: Joint 2nd on Artificial Analysis, Top 5 on Vals.ai
The independent numbers landed quickly, and they put the Muse Wonder series among the best models in the world, though not quite at the very top.
vals.ai Benchmark Results
On vals.ai, Muse Wonder Full lands around fifth of 47 models and Muse Wonder tenth, with the strongest showings in agentic and security work and the softest in legal and knowledge-heavy tasks.
| Benchmark | Muse Wonder Full | Muse Wonder |
|---|---|---|
| Vals Index | 64.57% | 61.06% |
| Code Migration | 61.03% | 62.55% |
| CyberBench v1.1 | 73.63% | 80.36% |
| EMB | 71.92% | 67.08% |
| Finance Agent (v2) | 59.06% | 56.64% |
| Legal Research Bench | 53.37% | 50.48% |
| MysteryMechanism | 35.13% | 27.48% |
| ProofBench v1.1 | 94.00% | 84.00% |
| Public Benefits Bench v1.1 | 73.75% | 70.23% |
| Tax Agent Bench | 43.00% | 36.27% |
| Vibe Code Bench v1.1 | 86.00% | 83.09% |
| BioMysteryBench | 73.33% | 67.04% |
| Harvey’s Legal Agent Benchmark | 21.67% | 17.08% |
| IOI | 79.83% | 66.94% |
| Terminal-Bench 4.0 | 58.08% | 49.49% |
| ProgramBench | 0.50% | Not reported |
| Terminal-Bench Science | Not reported | 22.86% |
Artificial Analysis Intelligence Index
On Artificial Analysis, Muse Wonder Full places joint second on the composite Intelligence Index, behind only Claude Opus 5.5, with Muse Wonder a step behind it and both models well clear of their predecessor.
| Model (max reasoning, unless noted) | Intelligence Index |
|---|---|
| Claude Opus 5.5 | 58 |
| Claude Sonnet 5.5 | 56 |
| Muse Wonder Full | 56 |
| GPT-6 Astra | 53 |
| Gemini 4 Argon (high) | 53 |
| Claude Fable 5.1 | 53 |
| Muse Wonder | 52 |
| Muse Spark 1.3 | 48 |
Meta’s claim of a material jump holds up on this tracker. Muse Wonder Full improves on Muse Spark 1.3 across all ten evaluations that make up the index, gaining eight points, while Muse Wonder gains four.
Artificial Analysis Per-Evaluation Results
The composite is built from ten evaluations covering agentic coding, knowledge work, scientific reasoning, and factual recall, with the separate Coding Agents Index and hallucination rate included for context. Muse Wonder Full places in the top three on six of them.
| Artificial Analysis evaluation | Muse Wonder Full | Muse Wonder | Leading rival |
|---|---|---|---|
| Intelligence Index (composite of ten) | 56 | 52 | Claude Opus 5.5 (58) |
| AA-Briefcase v1.1 (agentic knowledge work) | 61.45% (1729 Elo) | 57.95% (1659 Elo) | Claude Sonnet 5.5 (66.15%, 1823 Elo) |
| GDPval-AA v2.1 (real-world work tasks) | 67.85% (1857 Elo) | 63.50% (1770 Elo) | Claude Opus 5.5 (68.30%, 1866 Elo) |
| AutomationBench-AA (business workflows) | 71% | 68% | Gemini 4 Argon (78%) |
| Terminal-Bench 4.0 (agentic coding) | 57.1% | 52.0% | Claude Sonnet 5.5 (63.6%) |
| SciCode (coding) | 62.6% | 60.3% | Claude Opus 5.5 (66.9%) |
| Humanity’s Last Exam (reasoning and knowledge) | 53.6% | 50.1% | Claude Opus 5.5 (61.4%) |
| GDP.pdf (professional document reasoning) | 32.6% | 30.6% | GPT-6 Astra (32.2%) |
| CritPt (physics reasoning) | 32.9% | 28.3% | GPT-5.6 Sol (32.3%) |
| AA-LCR v1.1 (long-context reasoning) | 83.7% | 80.7% | Kimi K3 (88.7%) |
| AA-Omniscience (knowledge accuracy) | 43% | 41% | Claude Fable 5.1 (67%) |
| Coding Agents Index (max effort) | 68 | 63 | Claude Code + Sonnet 5.5 (68) |
| Hallucination rate (lower is better) | 23% | 27% | Gemini 4 Argon (15%) |
Meta’s strengths cluster in long-horizon and agentic work, scientific reasoning, and long-context tasks, where the Muse models sit near the top of the field. Factual accuracy and broad knowledge remain the weak spots, cushioned by a very low hallucination rate. Where the Claude models still pull clear is general reasoning and coding overall.
Speed and Latency
Speed cuts both ways. Muse Wonder outputs at 85.4 tokens per second, above average for its price tier, but its time to first token is 37.4 seconds against a median of 5.9, and vals.ai measures a latency of roughly 37 minutes per test run. These are models built to think at length before they answer.
Where Muse Wonder Actually Beats Wonder Full
The cheaper, more widely available model is not strictly worse than the flagship. On a handful of vals.ai benchmarks, including a cybersecurity evaluation where it tops the entire field, Muse Wonder actually beats Muse Wonder Full.
So “Full” describes an average advantage, not a clean sweep. For specific security and code-migration work, the accessible model can be the stronger pick.
Pricing: API Costs and the Muse Code Plans
Muse Wonder’s per-token price is the point where the “value play” era of Meta models clearly ends. At $4 input and $20 output per million tokens, it sits at the top of the market rather than undermining it.
Muse Wonder API Pricing
Meta Model API pricing for Muse Wonder is $4.00 per million input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens. That exactly matches Claude Opus 5.5 and lands well below GPT-6 Astra and Claude Fable 5.1, as the table below shows.
Measured per completed task rather than per token, the picture is more nuanced. As the chart above shows, Muse Wonder lands mid-table: cheaper to run than the Claude models, but more expensive than GPT-6 Astra and GPT-6.1 Sol.
| Model | Input price (per 1M) | Output price (per 1M) | Cost per task |
|---|---|---|---|
| Muse Wonder | $4.00 | $20.00 | $4.52 |
| Muse Wonder Full | Subscription only | Subscription only | Not available |
| Claude Opus 5.5 | $4.00 | $20.00 | $5.98 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $5.46 |
| GPT-6 Astra | $10.00 | $50.00 | $3.26 |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.72 |
The Muse Code Plans That Unlock Wonder Full
Muse Wonder Full is not something you can buy by the token. It is bundled into a Muse Code subscription, which Meta lays out on its Muse Code pricing page across four tiers:
| Muse Code plan | Price per month | Flagship access | Usage vs Everyday |
|---|---|---|---|
| Everyday Usage | $5 | Muse model family (no Wonder Full) | Baseline, 10-50 prompts per 5 hours |
| High Usage | $15 | Muse model family (no Wonder Full) | 5x |
| Power Usage | $50 | Muse Wonder Full | 20x |
| Maximum Usage | $200 | Muse Wonder Full | 80x |
That puts Muse Wonder Full within reach at $50 a month, which is where the value question gets interesting. Against Meta’s nearest rivals, the top tier costs the same, but the price of entry into each vendor’s best coding model does not:
| Subscription | Price per month | What the headline tier unlocks |
|---|---|---|
| Muse Code Power Usage | $50 | Muse Wonder Full, 20x Everyday usage |
| Claude Max 5x | $100 | Higher Claude Code limits, plus $100 in monthly API credits |
| ChatGPT Codex Pro 100 | $100 | 10x Plus usage limits |
| Claude Max 20x | $200 | 20x Pro usage, plus $200 in monthly API credits |
| ChatGPT Codex Pro 200 | $200 | 20x Plus usage limits |
| Muse Code Maximum Usage | $200 | Muse Wonder Full, 80x Everyday usage |
Read on headline numbers alone, Meta looks like the value leader at the flagship tier. Its best model arrives at $50, a quarter of the $200 needed for the top Claude or Codex plan, and its $200 Maximum tier advertises 80x usage against the 20x that both Anthropic and OpenAI list at the same price.
Two things complicate that reading. Each vendor measures its “multiple” against a different baseline, so 20x means little without knowing the starting allowance. And Anthropic’s Max tiers bundle $100 to $200 in monthly API credits while Meta’s coding plans do not, which offsets part of the gap for anyone who also spends on the API. Meta has lowered the entry price into a frontier model, but the usage accounting is opaque enough that only running your own workload will settle which plan is cheapest.
Does Meta’s Automatic Caching Beat Anthropic’s?
One structural difference stands out. Meta’s published pricing separates fresh input ($4.00) from cached input ($0.20), a twenty-fold discount, and lists no separate cache-write fee. That is a simpler billing model than providers that charge a premium on the tokens you write into the cache in the first place.
In practice this can narrow the real-world gap on long agentic sessions, where most of what you pay for is re-sent context rather than new output. On a pure per-task basis the two flagship models are close enough that caching efficiency and reasoning-token use, not the headline rate, are likely to decide which is cheaper for a given workflow.
Active Memory and Active Focus: What the Models Can Do
Meta’s pitch rests on two named features. Active Memory carries important context across conversations, retrieves the details relevant to the current task, and keeps goals, constraints, and earlier decisions available as work progresses. Active Focus is the persistence layer: the models are trained to try alternative approaches when a task gets hard, work through blockers, and follow through, while asking for clarification when a prompt is ambiguous and checking in before consequential actions.
Around those sit the standard frontier-agent skills. The series supports multi-step reasoning over long horizons, end-to-end agentic task completion with tool invocation and failure recovery, coding agents that work across large repositories and multi-file refactors, and multimodal reasoning over screenshots, charts, documents, and video.
Meta says Muse Wonder Full tops the independent long-horizon coding leaderboards, finishing ahead of both GPT-6 Astra and Claude Opus 5.5, and that it outperforms Fable, Opus, and Astra on measures of memory. Those are Meta’s own comparisons rather than independent ones, but they point the same direction as the neutral trackers.
The Adoption Challenge
Benchmarks are not the same as habit, and this is where Meta’s climb gets hard. Claude Opus 5.5 still holds the top spot on the Artificial Analysis Intelligence Index at the same $4 and $20 price as Muse Wonder, and Anthropic’s reputation for coding work is well earned. GPT-6 Astra averages slightly better than Muse Wonder across both benchmark sites while costing less per task, and GPT-6.1 Sol undercuts the entire field.
That leaves Meta with a marketing problem rather than a capability problem. Muse Wonder is good enough to sit in the top handful, but being roughly at parity on price with the model that already leads, and near-parity on quality with several others, is not by itself a reason to switch. The clearest wedge Meta has is the subscription: if Wonder Full is genuinely best-in-class at long-horizon coding and it comes bundled into a $50 or $200 plan, that is a concrete reason to try Muse Code where a pure API rate comparison would not be.
A deeper head-to-head, including Wonder Full against Opus 5.5 on the tasks developers actually care about, is worth its own piece. For now, the gap to the frontier is far narrower than the earlier GPT-6 Astra comparison would have predicted, without yet pulling clear of it.
Why This Release Matters for Meta
Roughly fourteen months after beginning an overhaul of its AI efforts, Meta now has models in the global top handful, and, more importantly for the business, a way to monetize them. The Muse app subscriptions and the Muse Code plans together represent Meta’s clearest attempt yet to build AI revenue outside advertising.
It also resets expectations for the next release. Having shown it can ship a model this close to Opus 5.5, Meta will be measured against that standard from now on, in a way it was not while Muse Spark was framed as the cheap option.
Frequently Asked Questions
Is Muse Wonder free?
No. Muse Wonder is billed per token through the Meta Model API at $4 input and $20 output per million tokens, or used inside Muse Code, which is subscription based. The model that powers the Muse consumer app sits behind a $20 or $100 monthly subscription.
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder Full is the more capable model, and it is available only through the Power Usage and Maximum Usage Muse Code plans. Muse Wonder is the broadly available version, offered through the Meta Model API and every Muse Code plan, and it is the one most developers will use.
How much does Muse Wonder cost?
Muse Wonder costs $4.00 per million input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens. Measured per task on Artificial Analysis’s Intelligence Index, it averages about $4.52.
Can I use Muse Wonder Full through the API?
No. Muse Wonder Full is not served through the Meta Model API. It is included with a Muse Code subscription on the Power Usage or Maximum Usage plan, which is also why there is no published cost-per-task figure for it.
Is Muse Wonder open weight?
No. Both Muse Wonder and Muse Wonder Full are proprietary models available through Meta’s own surfaces. Meta has signaled open-weight releases on its roadmap, but these models are not part of that.
How fast is Muse Wonder?
Artificial Analysis measures Muse Wonder output at 85.4 tokens per second, which is above average, with a time to first token of about 37 seconds. The models are built to reason at length before answering, so short prompts can feel slow.
Bottom Line
Muse Wonder and Muse Wonder Full are the strongest models Meta has shipped, and the first ones it has priced like a true frontier lab. Muse Wonder Full lands joint second on Artificial Analysis and fifth on vals.ai (for now), Muse Wonder is a clear notch below but still beats its bigger sibling on a handful of benchmarks, and the $4 and $20 API rate puts Meta at parity with Claude Opus 5.5 rather than under it.
The open question is adoption. Winning the benchmark race by a nose is not the same as winning developers, and Meta’s best argument, for now, is the one it does not have to make on token price: a bundled flagship inside a coding subscription. Whether that is enough to pull people away from Anthropic and OpenAI is the test that starts now.
