Meta's Muse Wonder Full Has No API: Distribution Is the Lock-In

The Short Answer

Meta's Muse Wonder Full has no API and no published token price — it ships only inside a Muse Code seat that defaults to letting Meta train on your content. On ibl.ai you own all the code and the data, so the model, the gateway and the memory stay yours.

The lock-in in this launch is not in the memory layer. Meta documents Muse Code memory as Markdown files you hold. It is in distribution and in the data default.

What did Meta announce on October 12, 2026, and what is Active Memory?

Meta announced two models, Muse Wonder and Muse Wonder Full, in a post dated October 12, 2026 on its corporate newsroom. Meta calls them "our most capable models to date, built to deliver state-of-the-art performance on personal and professional agentic tasks."

Two named capabilities arrived with them. Active Memory is, in Meta's words, "the ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand."

Active Focus is "the ability to keep working toward your goal when a task gets hard, try other approaches, recover from errors, and follow through, asking for clarification when needed and checking in before following through on anything important."

Meta says the models "bring that capability to expert domains including law, finance, and health, along with frontier coding performance." The announcement is the source for all of the above.

Note what Meta does not claim there: the announcement names no price, no context length and no benchmark figure.

Those appear on Meta's developer blog and model page instead, which list a 1M-token context window for both models and the model ids muse-wonder-1.0 and muse-wonder-full-1.0.

One widely repeated framing of this launch deserves correcting before anything else: Active Memory remembering "your client's risk tolerance, tax windows and portfolio preferences" is an extrapolation, not a Meta claim.

Meta says "finance" and "relevant information." The wealth-management example is the industry's, including ours in this post.

Where does Muse Code store an agent's memory, and can you take it with you?

In Markdown files you hold, and yes. This is the part of the launch that deserves credit rather than suspicion, and getting it wrong is the easiest way to lose a buyer's trust in everything else on the page.

Meta's Muse Code configuration documentation describes three memory scopes, all of them files. Personal project memory, the default, is "stored on your machine outside the repo, private to you, scoped to this project."

Project memory is "committed to the repo under <repo>/.agents/memory/," shared with everyone who clones it. Personal memory is machine-wide notes across all projects.

The retrieval mechanism is documented too. At session start Muse Code "injects an index of your memory: MEMORY.md plus a list of the other Markdown files (their paths, not their contents), up to 48 files. The agent reads individual files on demand."

Meta's own authoring guidance is to "keep an index in MEMORY.md and one Markdown file per topic."

Read that plainly: on this surface, an agent's accumulated context is version-controllable text in your repository, diffable in a pull request, greppable, and portable to any other tool that reads Markdown.

That is more portable than most managed offerings, and more portable than a row in a vendor's database.

The consumer surface is documented as well. Meta's help center article on managing AI memory documents downloading your Muse data, editing or deleting Memory.md directly, a forget skill for removing specific items, and a toggle for "Help improve our AI models."

Two caveats, both from Meta's own pages rather than from us.

The first is a security property the configuration page names itself: Muse Code "reads committed project memory into the model's context even in an untrusted workspace," and Meta's guidance is to "treat a repo's MEMORY.md as a prompt-injection surface, and review it on checkouts you don't control."

That is the correct warning, and it generalizes: any file that is injected into a model's context is an input channel, so a memory file that arrives with a cloned repository is untrusted input until a human reads it.

The second is that portability of the memory files is not portability of the system that uses them. The files tell you what the agent knows.

They do not give you the model, the gateway, the routing policy, or relief from the per-seat bill — which is where this launch's real switching cost sits.

Why does it matter that Muse Wonder Full is not available through any API?

Because the stronger of the two models has no published price and no programmatic access at all. Meta's model page lists per-token rates for muse-wonder-1.0 only; muse-wonder-full-1.0 is marked "Available by subscription in Muse Code."

Meta's developer blog says Muse Wonder Full is "available exclusively by subscription in Muse Code," on the Power Usage and the new Maximum Usage plans.

Meta's own Muse Code product page prices those at $50.00/month and $200.00/month, alongside Everyday Usage at $5.00/month and High Usage at $15.00/month.

That is a distribution fact with an architecture consequence. A frontier capability you can only reach by putting named humans on a monthly subscription is a capability you cannot put behind your own gateway, cannot route per agent, and cannot meter against a budget cap.

It is also the clearest per-seat shape in the launch. Per Meta's subscriptions documentation, High, Power and Maximum are sold as "5x," "20x" and "80x more usage than the Everyday Usage plan" — usage sold in blocks, attached to people.

Nothing about that is a memory problem. It is a procurement problem, and it arrives before the first token is spent.

Does a Muse Code subscription use your content to train Meta's models by default?

Yes, unless someone turns it off. Meta's Muse Code subscriptions page states that "Muse Code subscriptions default to Contributor mode. In Contributor mode, you get more usage from your plan in exchange for letting Meta use your content to train future Meta models."

The same page says Contributor mode "is a setting in Muse Code, and you can turn it off at any time, at which point your future content is not used to train Meta models." So it is a default, reversible, and prospective — turning it off governs future content.

The default does not vary by plan. The same page says "Muse Wonder and Muse Wonder Full are each available in either mode on the plans that include them," which means the opt-out is a setting someone has to go and change, on every plan, including the $200.00/month one.

This is the second half of the lock-in, and it is the half that is hard to unwind. A model you stop paying for is a contract you end. Content that was already used to train a model is not retrievable.

For a regulated team, that makes the Contributor-mode setting a day-one configuration item rather than a detail to revisit after the pilot — and it is worth confirming in writing which of your repositories were in scope while it was on.

What does Muse Wonder cost per token, and what does the same money look like per seat?

Per-seat pricing is not one option among several at enterprise scale — it is the wrong shape, because the bill tracks headcount rather than work. Meta's launch puts both shapes in one product and prices them, so the gap is unusually easy to read.

Every figure below is a published price, with the per-seat column multiplied out. The one row that needs a date is Google's: its current rates are promotional through November 30, 2026.

The last column is the whole argument. Five hundred people on the $200.00/month tier is $100,000 a month, or $1.2M a year, whether they open the tool or not.

At Muse Wonder's published $4.00 per million input tokens, that same $100,000 buys 25 billion input tokens a month. The question a per-seat bill never asks is whether your organization consumed anything close to that.

The cached-input rate makes the point sharper rather than softer.

Meta prices Muse Wonder's cached input at $0.20 per million tokens, a twentieth of its own standard input rate — so the workloads that reuse context heavily, which is exactly what a memory-carrying agent does, are the ones the metered path prices most cheaply.

The per-seat plans charge the same $200.00 either way.

One date to diary from Google's pricing page: Argon's input is "$2.00 through November 30, 2026. $4.00 starting December 1, 2026," with output going $10.00 to $20.00 and cached input $0.10 to $0.20. On December 1 it prices identically to Muse Wonder.

Did three AI labs just ship finance-first models?

No, and the accurate version is still interesting. The claim circulating after this launch — that three labs shipped finance-first AI in two weeks — does not survive a check against primary sources, on any of the three counts.

What is verifiable is narrower. Two frontier models were announced roughly 24 hours apart, and both name finance among their target expert domains. Meta's Muse Wonder post is dated October 12, 2026 and names "law, finance, and health."

Google's Gemini 4 Argon announcement is dated October 13, 2026 and says Argon "excels at deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science," and that it ranks "#1 on the Vals Index, which measures model performance across critical economic sectors like finance, coding, legal, and tax work."

Two is not three. "Names finance among five or six expert domains" is not "finance-first." And a general-purpose frontier model with a finance eval in its table is a different product from a finance-specific system.

The real signal is that professional-domain evals have moved into the launch material of general models.

That tells you where the labs expect revenue, and it tells a buyer that the differentiator will not be which model knows finance — all of them will claim it — but who controls the path between the model and the data.

What is the switching cost of a model you can only reach through a subscription?

It is a seat count, and seat counts are the hardest cost to unwind, because they are attached to people and to habits rather than to code.

Prompt engineering was the switching cost of the 2023–2025 era, and prompts are cheap to port: a few hundred strings, a weekend, and a regression suite. Fine-tunes raised the cost but capped it, because a fine-tune is a one-time artifact you can re-run.

A subscription-only model is different in kind. There is no endpoint to repoint, because there was never an endpoint — the integration surface is an application your people are logged into.

That changes the exit question from "which base URL do we swap" to "which 500 people change tools, and what else did they put in that tool while we were not watching."

It also removes the levers a gateway normally gives you.

With no API to call, there is nothing to route: you cannot A/B this model against another on the same traffic, fall back to a cheaper one under load, or meter its spend per agent, because the spend is not per agent — it is per person, per month, decided at purchase.

Memory, on both documented Meta surfaces, is the part of this you can take with you.

Our post on who owns an AI agent's memory when you switch models works the mechanism end to end, including the path built for migration: Anthropic's documented memory import, which asks the model you are leaving to recite what it remembers so you can paste the prose into Claude.

Worth noting for anyone building against the metered path: Meta's create response schema carries previous_response_id for server-managed state, so conversation state can live on Meta's side if you let it.

Whether it does is your design decision, and it is the one place in this architecture where you can choose to keep state in your own store instead.

How does ibl.ai keep the model, the gateway and the memory separately yours?

By shipping all three as things you own rather than things you reach.

On ibl.ai you own all the code and the data: the platform runs inside your perimeter under a perpetual source-code license, so the gateway, the routing policy and the memory store are yours to configure, query and audit.

The model is chosen per agent from the platform's provider list and can be switched at any time, because ibl.ai is model-agnostic — Claude, GPT, Gemini, Llama, an open-weight model on your own GPUs, or Muse Wonder through Meta's API if that is the right model for a given agent.

That ordering is the point. A model-agnostic platform can adopt Active Memory's underlying idea — carry what matters across sessions — and keep the right to move the model underneath it next quarter without renegotiating 500 seats.

Memory on the platform is also gated rather than ambient.

Capture and use are separate switches, scoped at the organization, the agent and the user, and private-mode conversations store nothing regardless of the agent's setting.

There is no Contributor-mode equivalent — no setting that trades extra usage for permission to train on your content.

For the schema and retrieval mechanics, see the agent memory documentation and our post on what turns an agent loop into a real AI agent.

And you pay with no per-seat pricing. Usage bills against a cap you set, so adding 500 people to a platform does not add a $100,000 monthly line before anyone has used it.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.

If you are evaluating Muse Wonder, Gemini 4 Argon or any frontier model for regulated work, the useful first question is not which one scores higher.

It is whether you can still change your mind in a year, and what it costs when you do. Agentic OS answers that one before the model question comes up.