Zuckerberg's Best AI Model Isn't in Meta's API. You Need a $50 Muse Code Plan to Use It

#discuss #ai #llm #devtools

Hermes IDE: Download free · Star it on GitHub — one console for every coding agent, on macOS, Windows and Linux
Book: AI Agents Pocket Guide · AI in TypeScript — the 5-book series
Me: xgabriel.com | GitHub

Meta shipped two frontier models on Monday, October 12. You can call
one of them from your code.

The other one, Muse Wonder Full, is the model behind the best numbers
on Meta's own scorecard. It is not served through the Meta Model API.
TechCrunch
reports it's for Muse Code subscribers only. Muse Code is Meta's
coding agent. Its
product page puts Wonder
Full in the Power Usage plan at $50 a month, and in the Maximum Usage
plan at $200.

What the API gets is Muse Wonder. At launch it lists at $4 per
million input tokens and $20 per million output
(TechCrunch),
the same list price as Claude Opus 5.5
(Artificial Analysis).
On the independent
Artificial Analysis Intelligence Index
it scores 52. Opus 5.5 scores 58.

Two models, two doors

Here is the split on launch day, from TechCrunch, the Muse Code page,
Meta's
model card
and
Artificial Analysis:

	Muse Wonder	Muse Wonder Full
Meta Model API	Yes, as `muse-wonder-1.0`, $4 / $20 per 1M tokens	No
Muse Code	Yes	Yes, Power Usage ($50/month) or Maximum Usage ($200/month)
Artificial Analysis Index	52	56
Meta's Personal Superintelligence Index	54.3	59.1

That last row is Meta's internal composite, the number at the top of
its own scorecard. Wonder Full leads every model there at 59.1. The
model you can call scores 54.3, behind Opus 5.5 (55.2) and GPT-6
Astra (54.5).

Where each Muse model is available on October 12, 2026. Only one of
the two has an API route.

Put the API model next to Opus 5.5, at an identical price, across the
14 rows of Meta's own scorecard: Muse Wonder wins 5 and loses 9.
Wonder Full, which you can't buy per token, wins 8 and loses 6. The
row-by-row reading is in
the full scorecard breakdown.

None of this makes Muse Wonder a bad model. It beats Opus 5.5 on
DeepSWE 1.1 (78.8 vs 74.2) and on Harvey's legal agent benchmark,
LAB-AA (10.0 vs 4.2). It also leads SWE-Marathon (53.1 vs 46.3), a
row Meta itself flags as not a common-harness comparison. At launch,
Artificial Analysis ranks it #9 of 219 models. It is a frontier-class
model, sold at the price of the model that beats it on the
independent index.
Handy AI
reports that Meta's Alexandr Wang said API pricing would be "meaningfully
below" rivals. That holds against Claude Fable 5.1 at its
$10 / $50 list price. It doesn't
hold against Opus 5.5.

The model is becoming a feature of the agent

The deal used to be simple. A lab ships a model, the model goes into
the API, and you pick whatever tool you like to call it. Your editor,
your agent and your scripts were all doors into the same room.

Wonder Full doesn't go into the API. At launch, Meta's strongest
model is a feature of Meta's agent. You don't choose the model. You
choose Muse Code, pay for the plan, and the model comes with it.

Claude Code already comes with Claude's plans, and Codex with
ChatGPT's. I went through those plans and prices in
You Already Pay for Claude Code and Codex. Agents
already arrive with a vendor plan, a usage pool and a sign-in. Meta
goes a step further: for now, its strongest model lives in the bundle
and nowhere else.

So picking an agent now means picking a lot more than an interface:

which models you can reach at all
which plan, and whose usage limits
which sandbox and approval model you work inside
which bill the work lands on

There is a plausible reason a lab would want this. Meta's model card
says performance "can vary with quantization, serving configuration,
prompts, tools, and the agent harness." If the harness moves the
score, shipping your best model only inside your own harness is one
way to control the result people see. That's my reading. Meta hasn't
given it as the reason.

Meta's strongest Muse model at launch: inside the agent, with the API
cable unplugged.

Muse Code may be worth the $50

Be fair to the product before you dismiss it. Meta's
Muse Code page lists
multi-agent coordination built in, parallel workers, background
reviewers, a cookbook for subagents in isolated worktrees, and "every
action is auditable." Meta also describes two model features: Active
Memory, which carries relevant context across conversations, and
Active Focus, which keeps working on a hard goal, tries other
approaches and checks in before important actions.

If your work looks like the rows Wonder Full wins, that matters.
Long-horizon coding is its strongest showing: SWE-Marathon 57.5
against Opus 5.5's 46.3, and DeepSWE 1.1 at 79.0 against 74.2. The
model card itself warns that SWE-Marathon "is not a controlled
common-harness comparison," and that Wonder Full's number used xhigh
reasoning effort. Treat that row with care.

Early reports are mixed. In the
Hacker News thread,
one commenter said Muse Code's sandbox "never worked across Mac or
Linux". Another said paying for Muse Code kept throwing errors. Others
were enthusiastic. They're launch-day comments from a handful of
people, and I haven't run Muse Wonder or Muse Code myself.

Don't marry one agent

If models are going to sit behind agents, the defense is to keep
agents swappable:

Keep each agent on a separate sign-in and plan.
Run them side by side on the same repo, one git worktree per task, so trying a new one costs you a branch.
Switch when one hits its usage limit, or when a better model lands behind a different agent.

There are two routes to Muse today. Meta's own agent installs a
command named muse into ~/.local/bin. Read the script before you
pipe it into a shell:

curl -fsSL https://dev.meta.ai/install.sh | bash

Or point Claude Code at Meta's API. Meta's
coding-agents docs show
this with muse-spark-1.3. Swap in muse-wonder-1.0, which the docs
list as a valid Model API ID:

export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="muse-wonder-1.0"
export CLAUDE_CODE_SUBAGENT_MODEL="muse-wonder-1.0"
claude

The last four exports follow the docs' advice to set the Opus,
Sonnet, Haiku and subagent model variables to the same ID "so no path
falls back to a Claude model." This route gets you Muse Wonder, never
Wonder Full. For a full same-ticket comparison against Opus 5.5,
follow the bake-off how-to.

Running `muse` next to Claude Code and Codex

I build Hermes IDE, so weigh this section with that in mind. Hermes
runs each agent in a real terminal, so a CLI it has never heard of
still works.

Press ⌘N (Ctrl+Shift+N on Windows and Linux), pick "Custom agent",
and give it a Name (Muse Code) and a Command (muse). The hint
under the field says what happens: "Hermes starts this command in the
session's terminal, as if you typed it." Whatever sign-in muse
needs happens inside that terminal. Hermes never asks for your
tokens. Download it if you want to
follow along.

Claude Code and Codex sit in the same list as built-in agents, with
their accounts. Every task gets a hermes/<task> branch in a
separate worktree, so the three agents never touch each other's
files. When one agent hits its usage limit, you can continue the task
in another.

Because Hermes types the command into a real shell, you can enter the
API route as a Custom agent too. Put the seven exports in a script
that ends with exec claude "$@", name it claude-muse, and set the
Command to claude-muse. The bake-off how-to has the
full script. I haven't run it against Meta's API myself.

From hermes-ide.com: a recreation of the 2.0 app with demo data. Four
vendors' agents in one sidebar, each on its own branch. A Custom
agent like muse joins the same list, without the status signal.

Hermes doesn't know a custom command's approval model, and it has no
status signals for one. The attention inbox can tell you when Claude
Code or Codex is waiting on you. It can't tell you when muse is, so
you watch that terminal yourself. Built-in agents keep full status,
accounts and model and effort pickers. Hermes is also young: 2.0
shipped on October 2, 2026, and 2.2.10 on October 12. There are no
cloud agents and no Docker isolation. Hermes is not affiliated with
Meta, Anthropic, OpenAI or any other agent vendor named here.

So, would you pay?

If the strongest models start landing in one agent and nowhere else,
"which agent do you use" turns into "which subscriptions do you
carry."

Would you pay $50 a month for an agent to reach a model you can't
call from your own code? Or is a model that only lives inside one
vendor's agent a dealbreaker for you? Tell me in the comments.

Try it with the agents you already use

I build Hermes IDE. It's free, source-available, and runs on macOS,
Windows and Linux. Download it,
press ⌘N (Ctrl+Shift+N on Windows and Linux), and start the same task
in Claude Code and in Codex. If you have a Muse Code plan, add muse
as a Custom agent and give it the same ticket on its own branch.

If keeping agents swappable matters to you,
star it on GitHub. It's
built in the open, and a star helps other developers who mix agents
find it. If your agent is missing or something breaks,
open an issue.
Hermes IDE is not affiliated with Nous Research.

If you build agents yourself, the
AI Agents Pocket Guide covers
the patterns behind tools like these.