Meta released its first frontier AI models on October 12, 2026: Muse Wonder and the stronger Muse Wonder Full. Here’s the short version for anyone running a freelance or creator business: the standard Muse Wonder costs exactly the same as Claude Opus 5.5 — $4 per million input tokens and $20 per million output tokens — while the best version, Wonder Full, isn’t sold per token at all. You can only get it inside a $50 or $200 per month Muse Code subscription. Meta’s own benchmarks look strong, but the independent scores tell a more careful story, and two of Meta’s headline wins come from its own internal tests. This breakdown covers the real pricing, what the benchmarks actually prove, and whether either model belongs in your workflow.
Two Models, Two Doors: What Meta Actually Released
The launch came from Meta Superintelligence Labs, the AI research organization Meta set up last year and put under former Scale AI CEO Alexandr Wang. Mark Zuckerberg announced it on X on October 12, 2026, calling it “our new family of frontier models with state of the art performance for personal and professional agentic work.”
This is Meta’s first model family it places at the frontier — the same tier as OpenAI’s GPT-6 line, Anthropic’s Claude 5.5 line, and Google’s Gemini models. But the release is split in a way that matters a lot for how you’d actually use it:
- Muse Wonder — the standard model, available through the Meta Model API and inside Muse Code, Meta’s terminal-based coding agent. This is the one most developers will touch.
- Muse Wonder Full — the bigger, more capable model, available only to Muse Code subscribers on the $50/month Power Usage or $200/month Maximum Usage plans. There is no API access and no per-token price for it.
That split is the first thing you should notice. Meta didn’t just release two models — it released two doors. One door charges you per token. The other charges you per month and won’t let you pay per token at all. Which door you walk through changes the math of every project you run.
The models are built for agents — AI systems that do multi-step work on your behalf instead of answering once and stopping. Think planning a schedule, filing an expense report, or reviewing a codebase while you sleep. Meta trained the Wonder series at its Prometheus gigawatt-scale data center (the models were developed under the codename “Watermelon,” per launch coverage), and they support roughly a million tokens of context, text plus image, audio, and video inputs, and tool calling, search grounding, and structured output.
The Pricing, Line by Line
Let’s get the numbers out of the way, because this is where the story gets interesting.
Muse Wonder API pricing
Through the Meta Model API, Muse Wonder costs:
- $4.00 per million input tokens
- $0.20 per million cached input tokens
- $20.00 per million output tokens
A token is roughly three-quarters of an English word, so a million input tokens is around 750,000 words — far more text than any single freelance deliverable, but agents chew through tokens fast because they “think” in long reasoning chains before they act.
Now here’s the detail that should make every freelancer sit up: Claude Opus 5.5 — Anthropic’s flagship, the model that beats Wonder on the independent benchmarks — costs exactly the same. $4 in, $20 out. Identical list price. (Our earlier look at Claude Opus 5.5’s cheaper launch covers why Anthropic priced it the way it did.)
That match is awkward for Meta because Alexandr Wang had said publicly that Wonder would be priced “meaningfully below” OpenAI and Anthropic. It wasn’t. As one newsletter put it, this is “second place at first-place prices” — a harsh line, but it lands, because the standard Wonder trails Opus 5.5 on the independent Intelligence Index (52 vs. 58) while charging Opus money. Even Claude Sonnet 5.5, which costs roughly half, outscores Wonder on that index.
The broader flagship pricing picture helps here — our comparison of Gemini 4 Argon, GPT-6 Astra, and Claude flagship pricing in 2026 lays out what the top tier costs across vendors. GPT-6 Astra, for instance, runs $10/$50 per million — 2.5x Wonder’s rate, and it scores 53 on the independent index, one point above Wonder’s 52. So Wonder is cheap relative to Astra but expensive relative to its own benchmark position.
The subscription wall: Muse Wonder Full
The better model, Wonder Full, doesn’t have a per-token price. It’s included with Muse Code’s Power Usage ($50/month) and Maximum Usage ($200/month) plans. Meta’s consumer Muse app, meanwhile, unlocks Wonder on its $20 and $100 per month subscription tiers, according to launch-day breakdowns.
This is a deliberate pricing architecture, not an accident. Meta is doing two things at once: competing with Anthropic on per-token API pricing, and copying the subscription playbook for its best technology. For a solo freelancer, $50/month is a real line item — it’s more than a ChatGPT Plus subscription and it only pays for one tool’s ecosystem. For an agency running long agentic jobs, $200/month might be nothing if the model finishes client work unattended that you’d otherwise bill hours for. The math depends entirely on your volume, which is why I’ll run the scenarios in the “should you switch” section below.
What the Benchmarks Really Say
Benchmarks are where this launch needs the most careful reading. There are two scorecards in play — Meta’s own and the independent one — and they tell different stories.
Meta’s own scorecard (vendor-reported)
Meta published a 14-row benchmark chart with the launch. On that chart:
- The Wonder family leads 7 of 14 rows, including long-horizon software engineering tests: SWE-Marathon at 57.5 vs. Opus 5.5’s 46.3, and DeepSWE 1.1 at 79.0 vs. 74.2 for Opus 5.5. Legal agent work also favors Wonder on Harvey’s LAB-AA benchmark (13.1 vs. 8.6 for GPT-6 Astra vs. 4.2 for Opus 5.5), per Neowin’s coverage.
- Claude Opus 5.5 leads 6 rows.
- GPT-6 Astra leads 1 row.
- The standard Wonder beats Opus 5.5 on only 5 of the 14 rows — the rest of the family’s wins come from Wonder Full.
- Two of Wonder’s 7 wins are on Meta’s own internal benchmarks, which nobody outside Meta has run.
I’ve seen enough launch-day charts to know the drill: vendor-reported numbers are a starting point, not a verdict. Meta picked the tests, the prompts, the effort settings, and two of the tests are its own inventions. None of that makes the numbers fake, but it means you should treat the cross-vendor deltas as directional until independent runs land. Several launch-day observers flagged possible benchmark optimization, and one widely shared comparison noted the evaluation setups between vendors “aren’t identical.”
The independent scorecard: Artificial Analysis
Artificial Analysis runs the same tests on every model, which makes its Intelligence Index one of the few outside scores worth quoting. Its v4.3 results for the new models:
- Claude Opus 5.5: 58 — still the top score
- Muse Wonder Full: 56 — ties Claude Sonnet 5.5
- GPT-6 Astra: 53 and Gemini 4 Argon: 53
- Muse Wonder: 52 — one point behind Astra
- Muse Spark 1.3 (Meta’s previous model): 48
Two things jump out. First, Opus 5.5 — the model that costs exactly what Wonder costs — still leads by 6 points. Second, the trajectory: Wonder Full is an 8-point jump from Muse Spark 1.3 in about six weeks. That’s a steep curve for one vendor, and it’s the most impressive number in this whole launch. Meta went from “the cheap option” to “in the top tier” in a month and a half. As one analysis put it, having shipped a model this close to Opus 5.5, Meta will now be measured against that standard in a way it wasn’t before.
The Artificial Analysis writeup adds texture that matters for your wallet:
- Cost per task for Muse Wonder (max) on their index averages $4.52 — cheaper than Opus 5.5 ($5.98), Sonnet 5.5 ($5.46), and Claude Fable 5.1 ($7.63), but more expensive than GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99), GPT-6.1 Sol ($0.72), and Meta’s own Muse Spark 1.3 ($1.60). List price is only half the story; what you pay per finished task depends on how many tokens the model burns thinking.
- Wonder Full is a token hog. It used roughly 140 million output tokens across the index runs, against a median of 88 million. It’s also more efficient than plain Wonder in one respect — 16% fewer input tokens per task (6.44M vs. 7.69M) — but the headline is that the best version reasons at length. That’s exactly the usage pattern a flat $50–$200/month subscription favors.
- On the Coding Agents Index, Wonder Full scores 68 — tying Claude Code with Sonnet 5.5 and edging Opus 5.5 (66). Standard Wonder scores 63; Muse Spark 1.3 managed 54.
- Speed-wise, Artificial Analysis measured Wonder’s output at about 85 tokens per second with a time to first token near 37 seconds. Built for long reasoning, so short prompts can feel slow.
One caution worth repeating: a viral chart showing different rankings circulated right after launch, and analysts flagged it as likely fan-made. Only trust scores you can trace to the tester.
The dissent
Not everyone buys the framing. Abacus.AI CEO Bindu Reddy said on X that Wonder is “nowhere near Fable or Astra level” — that’s one critic’s claim, with no counter-scores published, so take it as a signal of skepticism in the market rather than a measured result. Early-access users, rounded up by Latent Space, were more positive but nuanced: @theo ranks Wonder S-tier alongside Opus 5.5 and uses Opus to call Wonder for code review inside his own orchestrator; @kimmonismus found it resistant to trick questions with precise instruction-following, though “a bit dry” at creative writing; @QuixiAI found Wonder Full better than Astra but below Claude’s Fable and Opus tiers, while praising its speed; @clairevo called it strong at agentic tasks, coding, and emails, but weak at frontend design, SVG, video, and creative work.
Active Memory and Active Focus in Plain English
These are the two headline features, and they’re aimed squarely at people who run agents that do client work unattended. Here’s what they mean without the marketing:
Active Memory means the model is trained to carry important context across conversations — days, weeks, months — and pull back the relevant details when they matter. The practical version: you stop re-explaining your client’s brand voice, their budget constraints, or that weird edge case from last Tuesday every time you open a new chat. If the app it’s running in has a memory system, this works across sessions. Meta tests it with an internal benchmark called MemoryBench (hiding facts in 64K–100K token conversations, adding distractions, changing facts halfway through), where both Wonder versions score around 85.5 — higher than everything else Meta tested. That’s Meta’s own test, so treat it as a claim, not a fact.
Active Focus means the model is trained to sustain effort toward a hard goal: try another approach when something fails, work around blockers, adapt to new information — and, importantly, ask you when it’s actually unsure and check in before doing anything consequential. This is the “agent that doesn’t give up” pitch. For a freelancer, the difference is between an agent that books the first available option and stalls when the calendar is full, versus one that tries three more options and then comes back to you with a real question.
How does this compare to what you’re using now? Current Claude and GPT agents can already do multi-step work, but long-horizon reliability is the industry’s open wound — agents stall, loop, or quietly do the wrong thing after an hour. Meta’s bet is that training for persistence (rather than just bolting on retry logic around the model) closes that gap. Whether it does in practice is the thing nobody can verify yet, because independent long-horizon runs haven’t landed.
One honest downside, flagged by a launch-day technical review: a model that refuses to give up will also keep trying things you might not want it to try. Meta’s own model card recommends requiring human confirmation before any action that can’t be undone — refunds, deletions, sends, purchases. If you wire Wonder into client workflows, build that confirmation gate yourself. Don’t learn this the expensive way.
The $50–$200 Subscription Wall Is the Real Story
Here’s where I think the practitioner angle actually lives. Forget the benchmark horse race for a minute and look at the business model.
Meta’s best model is only sold as a subscription. You cannot buy Wonder Full per token on the API. For freelancers and small agencies, that creates three distinct situations:
If you’re a light user doing occasional coding help or drafting, the API-priced Wonder is your option — and at Opus-identical pricing with lower independent scores, it’s a hard sell unless you specifically want Meta’s ecosystem. You’d pick it for the agentic persistence features, not the price.
If you run long agentic jobs — multi-hour coding sessions, research pipelines, content operations — the subscription math gets interesting. Artificial Analysis found Wonder Full burns roughly 140 million output tokens on their index versus an 88-million median. Heavy reasoners like that are exactly where flat-rate pricing wins: $200/month for unlimited-ish use of a model that thinks at length is potentially a steal compared to per-token billing on a model with the same habits. This looks like the intended customer: the agency running client work through Muse Code all day.
If you need the best model inside your own product — your SaaS, your automation pipeline, your client deliverable system — you’re out of luck. Wonder Full is not on the API. That’s the most consequential part of this launch for builders: Meta’s frontier capability is fenced inside Meta’s tool. VentureBeat’s coverage put it bluntly in the headline: only the weaker model is available via API.
There’s also the lock-in question nobody at the launch answered. Subscription pricing means your costs are predictable but your leverage is zero — Meta can change the plan terms, the usage limits, or the model behavior, and your workflow is along for the ride. Per-token API pricing is the opposite: volatile bills, but you can switch vendors in an afternoon. Pick your poison based on how deep the model sits in your revenue path.
Meta Ends Its Open-Source Era
This one deserves its own section because it’s a genuine era change. Meta’s earlier Llama models helped popularize downloadable AI models that organizations could run on their own infrastructure. Neither Wonder model ships with publicly downloadable weights. And the open weights for Muse Spark — promised back in September — still haven’t shipped either.
What this means for your costs: self-hosting is off the table for Meta’s frontier tier, full stop. If you were hoping to run a frontier-class model on your own GPUs to dodge per-token fees, that door is closed. Your options are Meta’s API, Meta’s subscriptions, or a competitor. For small businesses that liked the Llama-era playbook — download the weights, run them cheap, keep data in-house — this launch is Meta saying that playbook is over for its best models. Data governance questions follow naturally: everything you send to Wonder goes to Meta’s servers, under Meta’s terms.
Should You Switch Your Workflow?
Honest answer: probably not this week. Here’s the decision framework I’d actually use.
Try Muse Wonder (API) if: you’re already building on the Meta Model API, you run long agentic coding or research tasks and want to test the persistence claims, or you want a second opinion model for code review (the @theo setup — Opus orchestrating, Wonder reviewing — is a genuinely interesting pattern worth stealing).
Consider the $50 Muse Code plan if: you live in a terminal, you run multi-hour agentic sessions, and you want to test whether Wonder Full’s reasoning-at-length actually finishes jobs that stall on other models. Note the plan name carefully: Power Usage vs. Maximum Usage is about usage caps, not model quality — both get Wonder Full.
Don’t bother yet if: your current stack works, you need API access to the best model, you self-host for cost or privacy reasons, or your work is frontend design, creative writing, or video — the early user reports are consistently weak there.
The 30-day test I’d run: pick one repeatable agentic workflow (a weekly research brief, a code review pass, a content repurposing pipeline), run it on your current model for two weeks while logging cost and failure rate, then run the same workflow on Wonder (API or Muse Code) for two weeks. Compare total spend, tasks completed without intervention, and output quality. Don’t switch on vibes; switch on your own numbers. The video below walks through the pricing tiers and benchmarks in more detail and is worth the watch before you spend anything:
What I’m Watching Next
Three things will decide whether this launch matters in six months:
- The safety report. Meta says a detailed evaluation under its Superintelligence Scaling Framework is coming, with no date. Until it lands, the safety claims are unchecked — relevant if you handle client data.
- Independent benchmark runs. The Artificial Analysis scores are a start, but independent long-horizon agent evaluations — the thing Wonder is actually for — haven’t landed yet. Meta’s internal wins need outside confirmation.
- The Muse app rollout. Meta’s consumer assistant is reportedly climbing the App Store charts, and Wonder is headed there for subscribers first (our earlier piece on the Muse app’s small-business connectors covers what’s already live). If Wonder-level agents reach hundreds of millions of consumers, the freelancer question flips from “should I use this model” to “how do I compete with everyone using it.”
The through-line: Meta is now a frontier lab that prices like Anthropic and distributes like a SaaS company. The 8-point jump in six weeks is real and impressive. Everything else — the persistence claims, the memory claims, the value of that subscription wall — is still waiting on proof. Spend accordingly.
Frequently Asked Questions
How much does Muse Wonder cost per token?
Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens through the Meta Model API — exactly matching Claude Opus 5.5’s list price. Context window is roughly 1 million tokens. (Sources: TechCrunch, VentureBeat, Artificial Analysis.)
What’s the difference between Muse Wonder and Muse Wonder Full?
Wonder Full is the larger, more capable model. Standard Muse Wonder is available via the Meta Model API and all Muse Code plans; Wonder Full is available only to Muse Code subscribers on the $50/month Power Usage or $200/month Maximum Usage plans. It is not sold per token and has no API access.
Is Muse Wonder better than Claude Opus 5.5?
It depends on who’s grading. On Meta’s own 14-row benchmark chart (vendor-reported), the Wonder family leads 7 rows, Opus 5.5 leads 6, and GPT-6 Astra leads 1 — but two of Wonder’s wins are on Meta’s internal benchmarks. On the independent Artificial Analysis Intelligence Index, Opus 5.5 scores 58, Wonder Full scores 56, and standard Wonder scores 52. So: competitive, not first.
Is Muse Wonder open source? Can I self-host it?
No. Neither Wonder model is released with downloadable weights, and the open weights Meta promised for Muse Spark in September still haven’t shipped. If you need a self-hostable frontier-class model, this launch doesn’t give you one.
What are Active Memory and Active Focus?
Active Memory is the model’s trained ability to carry important context across conversations over days, weeks, or months and retrieve relevant details when needed. Active Focus is its trained persistence on long-horizon tasks — trying alternative approaches, recovering from errors, and checking in with you before consequential actions. Both are training properties, not settings you toggle; the MemoryBench scores Meta cites come from its own internal test.
Should freelancers pay for the $50 or $200 Muse Code plan?
Only if you run enough long agentic sessions for flat-rate pricing to beat per-token billing. Artificial Analysis found Wonder Full burns well above the median in output tokens, which favors subscriptions for heavy users. If you only need occasional help, the API-priced Wonder — or your current model — is the cheaper experiment. Run a two-week cost comparison on a real workflow before committing.