Meta unveiled Muse Wonder and its larger Full sibling on October 12, its first frontier models, built to run personal and professional agents. The pitch rests on an active memory that carries a user’s context from one conversation to the next, priced at 4 dollars per million input tokens on the API.
Key Takeaways
- Muse Wonder is available through Meta’s API and Muse Code, while the Full version is limited to Muse Code subscribers.
- Artificial Analysis scores the Full version at 56 and the standard model at 52 on its Intelligence Index.
- Pricing climbs to 4 dollars per million input tokens and 20 dollars per million output tokens, close to three times the previous model’s cost per task.
Have an AI Sum Up This Article
ChatGPTMuse Wonder and its Full version, two models built for agents
Meta laid out the release in a research post introducing both new models, where it frames Muse Wonder as a frontier model that leads on personal and professional agents, the assistants that carry out whole tasks on a user’s behalf.
Two features carry the release. Active Memory keeps the important context about a user across conversations and pulls up the details that matter for the task at hand. Active Focus describes how the model keeps pushing toward a hard goal, trying alternative approaches, working through blockers and adapting when new information lands.

Meta research scientist Kavya Kopparapu put the goal plainly in comments to TechCrunch on launch day, saying the team worked hard to make sure the model remembers what matters to you over days, weeks and months.
Meta also claims domain expertise in law, finance and health, along with frontier-level coding. Its evaluations include Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon. Training ran at Prometheus, Meta’s gigawatt-scale data center in central Ohio, which came online earlier this year.
Access is gated for now. Muse Wonder runs on the Meta Model API and inside Muse Code, while the Full version stays reserved for Muse Code subscribers. The Muse app gets it next, starting with paying users on the 20 and 100 dollar monthly plans we covered when the Muse agent started sending emails and booking travel for its users.
The new model fills the top of a lineup whose other end already belongs to Muse Glimmer, the 30-billion-parameter agent model built to run on a single graphics card. Meta now covers both ends, with a top-tier model sold through its API and its subscriptions.
How Muse Wonder stacks up in independent scores
The first outside numbers come from Artificial Analysis. Its page on Meta’s new model puts it at 52 on the Intelligence Index, ninth out of 216 models, with a one million token context window and an output speed of 101.4 tokens per second.
The Full version reaches 56. That is a clear jump from Muse Spark 1.3, Meta’s previous model, which sat at 48. On the same group’s Coding Agents Index, Muse Wonder Full hits 68 and ties Claude Code running Sonnet 5.5, while the standard model climbs from 54 to 63.
Factual knowledge remains the weak spot. On AA-Omniscience, Muse Wonder Full scores 30 and the standard model 25, behind Claude Opus 5.5 at 46, GPT-6 Astra at 43 and Gemini 4 Argon at 42. The Full version’s hallucination rate still drops from 33% to 23%, second only to Gemini 4 Argon at 15%.
Meta says the models match the benchmark scores of Claude Fable 5.1 and GPT-6 Astra and beat rivals on MemoryBench, a memory test. Both claims come from the company itself and still wait for independent checks.
The price moves up a tier. Muse Wonder costs 4 dollars per million input tokens and 20 dollars per million output tokens, against 1.25 and 4.25 dollars for Muse Spark 1.3. Artificial Analysis puts the cost of one Intelligence Index task at 4.52 dollars, up from 1.60 dollars. Meta is leaving the bargain corner it held this summer with Muse Spark 1.1 and its cut-price entry into the frontier pack.
On that same cost per task, Claude Opus 5.5 comes in at 5.98 dollars, Claude Fable 5.1 at 7.63 dollars, GPT-6 Astra at 3.26 dollars and Gemini 4 Argon at 1.99 dollars, which places Muse Wonder in the middle of the premium tier.
More articles on Horizon
- ChatGPT Watermark Comes to Every Answer in the EU
- Claude for Startups Offers a Free Year of Claude Team
- Microsoft and Meta Are Pulling Back From Claude
What a model with long memory means for builders and rival labs
For a team building agents, the appeal is continuity. A model that holds a user’s context for weeks saves people from re-explaining preferences, open files and client constraints at every session.
The economics need a closer look. At 4.52 dollars per measured task, the new model costs close to three times its predecessor, and the strongest version is still off the API. Developers will need to test their own workloads before moving an agent into production.
For OpenAI, Anthropic and Google, the pressure shifts to memory. Meta has made persistent context its headline feature and claims the lead on MemoryBench, so the three labs will have to show their own assistants hold on to user details over similar stretches of time.
Safety gets real space in the launch. Meta calls Muse Wonder its most aligned model to date, with safeguards against chemical, biological, radiological, nuclear and cyber misuse, while keeping defensive cyber capabilities under its Meta Superintelligence Scaling Framework.
The consumer business is the open question. Meta was already weighing a 199.99 dollar monthly tier for Hatch, its premium agent, and the arrival of Muse Wonder in the Muse app, paying subscribers first, will show whether people will pay for an assistant that remembers them.
Follow the story on Horizon.


