Muse Wonder: Meta's Frontier Model, Benchmarks and Price

Muse Wonder is Meta's bid for the top of the AI model race. On October 12, 2026, Meta released Muse Wonder and a stronger Muse Wonder Full and said they put the company at the frontier, a goal it set when it created Meta Superintelligence Labs last year.

The launch also has a business job to do. Meta wants the new models to make Muse, the personal agent it launched in September, smart enough to keep users coming back. That won't be easy. This article covers what Muse Wonder is, what the benchmarks really show, what it costs, and how to test it before you move any work.

Key Takeaways

• Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, and calls them its most capable models so far.
• Meta's own charts show Muse Wonder Full ahead of rivals on several agent and coding tests, such as 57.5% on SWE-Marathon against 46.3% for Claude Opus 5.5.
• Independent tests are closer: Artificial Analysis scores Muse Wonder Full at 56, behind Claude Opus 5.5 at 58 and ahead of GPT-6 Astra at 53.
• Muse Wonder costs $4 per million input tokens and $20 per million output tokens, the same list price as Claude Opus 5.5.
• The bigger test is Muse itself, whose download growth has slowed since its September launch, according to the Wall Street Journal.

Muse Wonder at a Glance

• Released: October 12, 2026.
• Versions: Muse Wonder for developers, and Muse Wonder Full for Muse Code subscribers.
• Price: $4 per million input tokens and $20 per million output tokens for the standard model.
• Context window: about 1 million tokens, which the prompt and the response share, so very long inputs leave less room for the answer.
• Best independent score: 56 on the Artificial Analysis Intelligence Index, for the Full version.
• Headline features: Active Memory and Active Focus.
• Weights: closed.
• Safety report: promised by Meta but not yet published, so the safety claims can't yet be checked against full test results.

What Is Muse Wonder?

Muse Wonder is a family of two AI models from Meta Superintelligence Labs, the unit Meta set up in 2025 to reach the AI frontier. Meta says they are its most capable models to date and that they lead on personal and professional agent tasks.

Meta's examples are everyday and expert jobs: handling insurance claims, understanding a lease, and building games or business apps. The company says the models are strongest in law, finance, health and coding.

The models were trained at Prometheus, Meta's gigawatt-scale data centre, on "larger, stronger base models" than its earlier Muse Spark, according to Meta's research blog. Muse Spark arrived in April 2026, so this is Meta's second big model step in six months.

Who Built It?

Meta Superintelligence Labs, a unit Mark Zuckerberg set up in mid-2025 after Meta's Llama 4 models drew a lukewarm response. To staff it, Meta invested about $14.3 billion in data-labeling firm Scale AI and brought in its CEO, Alexandr Wang, to lead the effort.

Meta then spent heavily on researchers and compute. Zuckerberg laid out the goal in a July 2025 memo on "personal superintelligence": AI that helps each person with their own goals, rather than one that automates all work.

Muse is the product version of that idea. Muse Spark in April was the lab's first big model, Muse launched as a personal agent in September, and this Meta AI model is the lab's first serious claim to the top of the leaderboards.

Muse Wonder vs Muse Wonder Full: What's the Difference?

There are two versions, and they're sold in different places. Here's how they compare, based on Meta's announcement:

Feature	Muse Wonder	Muse Wonder Full
Who can use it	Developers via the Meta Model API and Muse Code	Muse Code subscribers
Public API price	$4 input, $20 output per million tokens	Not announced
Artificial Analysis score	52	56, at maximum reasoning effort
In the Muse app	Coming soon, subscribers first	Paid subscribers first, no date given

Think of Muse Wonder as the one you can build with today. Muse Wonder Full is the stronger model, but for now you can only reach it through Meta's own coding tool.

What Do Meta's Muse Wonder Benchmarks Show?

Meta's charts are strong. These numbers come from Meta's launch materials, as reported by AlphaSignal, and haven't been checked by outside testers:

Test	Muse Wonder result	Rival result
SWE-Marathon, long coding projects	57.5% (Full)	46.3% (Claude Opus 5.5)
MemoryBench, recall across chats	About 85.5%	81.5% (Claude Opus 5.5)
Personal Superintelligence Index	59.1 (Full)	55.2 (Claude Opus 5.5), 54.5 (GPT-6 Astra)
Prompt attack success, lower is better	10.5%	19.7% (GPT-6 Astra)

Two cautions apply. Some of these tests, like the Personal Superintelligence Index, are new and closely tied to what Meta is building, so they may suit its models.

Vendor charts also depend on setup. Model version, prompts, tools and how long each model is allowed to think can all move a score by several points.

How Does It Score in Independent Tests?

Closer to the pack, but still near the top. Artificial Analysis runs the same set of tests on every model, and its Intelligence Index is one of the most widely cited outside scores.

Here's how it ranks the leading models, as reported by Trending Topics and Artificial Analysis:

• Claude Opus 5.5: about 58, the top score.
• Muse Wonder Full: 56, at maximum effort, and in the top three on six of the ten tests in the index.
• GPT-6 Astra: about 53.
• Muse Wonder: 52.
• Muse Spark 1.3, Meta's previous best: 48.

By our math, Muse Wonder Full is 8 points ahead of Meta's last model. That's a real jump in six months, and it puts Meta in the same top tier as Anthropic, OpenAI and Google on that index.

One more caveat: a viral chart showing different rankings circulated after the launch, and AGI Hunt flagged it as likely fan-made. Stick to scores you can trace to the tester.

How Much Does Muse Wonder Cost?

Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with a context window of about 1 million tokens, according to Vercel's AI Gateway listing. Here's how that compares:

Model	Input per 1M tokens	Output per 1M tokens	Our math: 1M in, 200k out
Muse Wonder	$4	$20	$8
Claude Opus 5.5	$4	$20	$8
GPT-6 Astra	$10	$50	$20
Muse Spark 1.1	$1.25	$4.25	$2.10

Claude Opus 5.5 prices come from Anthropic's pricing page, and GPT-6 Astra prices are as listed on OpenRouter. List price is only half the story, though.

What you really pay depends on cost per finished task. A model that uses twice as many tokens to think, or needs two tries to get a task right, can cost more even at the same rate. Artificial Analysis noted that Muse Wonder Full used about 140 million output tokens to run its index, against a median of 88 million.

What Are Active Memory and Active Focus?

They're the two agent features Meta is selling hardest. Both aim at the same problem: agents that forget things or give up halfway through a long job.

• Active Memory carries useful details from past conversations into the current task, so you don't have to repeat yourself.
• Active Focus keeps the model working toward a goal when a task gets hard. Meta says it tries other approaches, recovers from errors, asks for clarification and checks in before important actions.

That "checks in" part matters. Agents that take real actions, like paying a bill or sending a message, need a clear point where a person approves the step. We covered why in our guide to AI agent permissions.

Why Does Meta Need Muse Wonder Now?

Because Muse growth is cooling. Muse launched on September 8 and drew millions of downloads in its first weeks, helped by heavy promotion and Meta's invite code and token offers.

The pace has since slowed. The Wall Street Journal reported that global downloads have slowed and that the launch is a test of whether Muse Wonder can lift retention.

Outside data points the same way. Hunterbrook, using TickerTrends data, said average daily downloads fell 8.1% in the week to early October, to about 165,000 a day. Hunterbrook's fund disclosed it was short Meta stock, so weigh that source with care.

Meta hasn't published Muse's retention or revenue. A smarter model is the obvious lever to pull, since people keep using an agent only if it gets things right.

What Has Meta Said About Safety?

Meta calls Muse Wonder its "most aligned model to date." Its main safety claims are:

• Misuse: its strongest safeguards so far against help with chemical, biological, nuclear and cyber weapons.
• Privacy: training to avoid sharing private information without permission.
• Cyber defence: staying useful for security teams that find flaws in their own code, while refusing to help attackers.
• Prompt attacks: better resistance to hidden instructions planted in web pages or files.

The full safety report isn't out yet. Meta says it will be "available soon." Until it is, treat the safety claims as promises, not results, especially if you plan to give the model access to real accounts or data.

Is Muse Wonder Open Source?

No. Meta built its AI name on open Llama models, but Muse Wonder, like Muse Spark before it, is closed and available only through Meta's API and apps.

Meta hasn't said whether it will ever release the weights. If running a model on your own servers matters to you, look at open models instead, or at Meta's smaller Muse Glimmer, which was reported as a local agent model.

Which Workloads Suit It Best?

Meta's own pitch points to a few areas. Here's how we'd read it, based on where the model scored well and what's available today:

Workload	Fit today	Why
Long coding projects	Worth testing	Strong SWE-Marathon claims, and it's built into Muse Code
Multi-step agents	Worth testing	Active Memory and Active Focus target this directly
Legal and finance documents	Test with care	Meta cites legal and finance tests, but check data terms first
Health questions	Wait	High stakes, and the full safety report isn't out
Simple, high-volume tasks	Use a smaller model	Frontier prices rarely pay off on easy jobs

The pattern is simple. Test the new model where hard, long tasks are failing today, and leave easy work on cheaper models.

Should Your Team Test Muse Wonder?

Yes, if you run agent or coding workloads at scale. A new frontier model at the same price as the current leader gives you real bargaining room, even if you don't switch.

Be more careful in three cases:

• You need Muse Wonder Full. It isn't on the public API yet, so you can't build a product on it.
• You handle sensitive data. Read Meta's data and retention terms closely before sending client data, and wait for the full safety report.
• Your workloads are stable and cheap. If a smaller model already does the job, a frontier model may add cost without adding much value.

How Do You Run a Fair Model Test?

Use your own tasks, not Meta's benchmarks. Here's the process we'd follow:

• Pick 30 to 50 real tasks from your product, including a few that your current model gets wrong.
• Run each model the same way.
• Score the outcome, not the style: did the task finish correctly, and did a person have to fix it?
• Track cost per finished task, including retries and thinking tokens, not just the price per token.
• Measure speed, since a slower model can hurt a product even when it's more accurate, especially in chat or voice features where users wait on every reply.
• Keep a fallback, routing through a gateway so you can switch back in minutes if quality slips.

A week of testing like this tells you more than any launch chart. It also gives you numbers to take into price talks with your current provider.

What Is Still Unclear About Muse Wonder?

A few big questions remain open after the launch:

• When Muse Wonder Full will reach the public API, and at what price.
• When Muse Wonder will power the Muse app for all users, not just subscribers.
• What the full safety report shows.
• Whether Meta's lead on its own benchmarks holds up as more outside testers run them.

We'll update this page as Meta answers them.

What If You Already Use Muse Spark?

Then testing is easy. Muse Wonder sits on the same Meta Model API, so in most setups you change the model name and run your test set again.

Watch the bill, though. At $4 and $20 per million tokens, the new model costs about three to five times Muse Spark 1.1's list price, by our math, so move only the tasks where Spark falls short.

A mixed setup often works best. Keep Spark on routine work, send hard cases to the bigger model, and compare cost per finished task every month.

What Muse Wonder Means for the AI Model Race

The top tier now has one more serious player. Muse Wonder doesn't clearly beat the leader in independent tests, but it's close enough, and priced the same, that teams have a real new option.

For buyers, that's good news. More competition at the frontier usually means faster price cuts and better terms, and it makes a single-vendor lock-in harder to justify.

For Meta, the harder test is still ahead. The company has to turn strong scores into an agent that people use every day, and the download numbers suggest that work isn't done.

If you're choosing models for an agent product and want help running a fair Muse Wonder test against your current setup, that's the work we do. Our review of GPT-6 Astra covers another model in this race.