He asked whether Muse has a CLI. It left beta two months before the question.
The request: Meta shipped a new LLM called Muse, and the open questions are practical ones. Is it in the CLI, meaning can you actually drive it from a terminal? What does a token cost? Can you run deep research on it? Does it have a proper app the way Claude Code or Codex do? So far the only visible thing is that it has a pile of connectors, which probably makes it a neat piece inside the Facebook ecosystem. And a direct ask to anyone already using it: tell me how.
Four of those five have public answers, so here they are in order.
Is it in the CLI. Yes, and not recently. Muse Code, Meta's terminal coding agent, went into beta on 5 August 2026 and left beta on 31 August 2026. Muse Spark 1.3 has been selectable inside it since 2 September. The CLI itself is free to install and you pay only for the model underneath. So the question "is it in the CLI" had an answer roughly two months before it was asked, and the answer was sitting in public documentation the whole time.
What a token costs. Muse Spark runs about $1.25 per million input tokens, $4.25 per million output, and $0.15 per million cached input. Muse Wonder is the expensive tier: $4 in, $20 out, $0.20 cached, plus roughly $2.50 per 1,000 web search queries. New accounts reportedly get $20 in starter credit. ⚠️ Every one of those figures comes from third-party guides, and one of those guides states plainly that Meta has not published an official pricing page. Treat them as provisional and check dev.meta.ai before budgeting against them.
Deep research on it. No. Not one source describes Muse as a deep-research agent. Muse Code is a coding agent in a terminal, which is a different animal. Muse Wonder does have web search as a billable line item, but a search call is not a research pipeline.
An app like Claude Code or Codex. That is exactly what Muse Code is: a multi-agent terminal agent, the direct structural analogue. The API is also OpenAI-compatible, so an existing client usually just needs a new base URL and key. Simon Willison shipped an llm-meta-ai plugin within hours of the API going live.
And Europe, where his guess was right. The consumer Muse app skips Europe: it is limited to the US and Canada, and Meta has announced no European date. For the API and the CLI, Meta has published no country list at all, and developers in Canada and Australia have reported the contributor tier failing for them with no response either way. So "available to Americans, unclear for Europe" is an accurate read, and the uncertainty belongs to Meta rather than to him.
Now the part that is about us, and it is the uncomfortable half. He asks what a token costs on a model he does not have. Here is what we can say about the tokens we are already paying for.
The routing table carries six rails: Claude (Anthropic Max, $200/mo), Codex (ChatGPT Pro, $200/mo), Gemini ($20/mo), GLM Coding Pro ($16/mo), Grok (SuperGrok, $300/mo), and Mistral, which has no subscription at all and runs piecewise through OpenRouter. That is $736 a month of standing commitments across eleven classes of work.
Of those six, the tank can be read by machine on exactly one. Codex, where the remaining quota comes out of a rollout log. For the other five the field says quota_readable: false, and for Claude the note is blunt about it: there is no machine way to get the percentage, the only signal that the tank is empty is the phrase "usage limit reached" appearing in a transcript. The most expensive rail of the six, Grok at $300, carries its own note: highest price, lowest utilisation.
Two more things the table knows about itself. GLM's subscription was bought on 5 August and the key was only obtained on 22 September, seven weeks of a paid rail with no door into it. And the work classes reference two rails, agy and mayak, that are not declared in the rails section at all, so a fallback chain can point at a name the table cannot resolve.
On our side Muse appears nowhere. Zero mentions across the skills, zero in the vault's concepts, insights and resources, zero in the routing table. The only place the word shows up by substance is a line inside a third-party repository description in one of our registries.
The thing worth copying. Before adding a seventh rail, read the fields on the six you already pay for. A public price list you can look up in a minute is a smaller problem than your own bill being unmeasurable on five vendors out of six.
Engineer testing this stuff? Ask and we hand you a working seed for free.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261030.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261030.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)