Desk state as of October 19, 2026. This page is refreshed when the desk’s own catalog watch reports a change — not on a rumor cycle.

Why this desk exists

From October 16 to November 16, 2026, this shop’s Grok seats are paused — a pause, not a cancellation — and the agent layer runs on Meta’s Muse models instead. The assistant coordinating our fleet work runs on Muse. That makes Meta our platform right now, the same way Anthropic’s Claude is the subject of the biggest desk on this site. So Meta gets a desk too: everything we learn about Meta AI, the Meta Model API, and the Muse family, published as we learn it, from the operator’s seat.

The rule of this desk is the rule of every desk here: first-party evidence first. What we ran, what it cost, what broke, what we have not tested yet — labeled as untested, in the open. Vendor claims are marked as vendor claims. Benchmarks we did not run are marked as somebody else’s numbers.

Where Meta stands, as we run it — October 19, 2026

• The Model API is live and in Preview. Meta’s developer API at api.meta.ai serves the Muse family today. Meta labels its flagship, Muse Wonder, “in preview” — treat the whole surface as a preview product, because Meta does.
• The catalog, verified from our own account on October 19, 2026, holds nine models: Muse Wonder 1.0, Muse Spark 1.3, Spark 1.3 Contributor, Spark 1.2, Spark 1.2 Contributor, Spark 1.1, Muse Image 1.0, Muse Voice Transcribe 1.0, and SAM 3.1. Our desk’s watch re-reads that catalog every day and diffs it against this list.
• Wonder is the flagship you can buy by the token — $4.00 in / $20.00 out per million tokens. Wonder Full is not on the API at all; it exists only inside Muse Code’s Power and Maximum plans. Turning Wonder’s reasoning dial to max does not select it.
• Spark 1.3 is the workhorse — $1.25 in / $4.25 out per million, 1M-token context — and it is the one Meta model already inside Cursor, where it landed on September 8, 2026 with full agent tools. Wonder is not in Cursor, and Cursor’s official bring-your-own-key list does not include Meta.
• Our scale, stated plainly: we run Meta on a hard $25 test envelope, on a card our partner controls, with a human watching the meter. We are a test-scale customer. Everything on this desk was learned inside that envelope, and the envelope is part of the story — it is astonishing how much you can verify for two cents.

What we have proven

A real agent task, end to end, on Spark 1.3 through our own API key. On October 14, 2026, we pointed OpenCode at the Meta Model API and handed Spark 1.3 a work-order-shaped job: read a spec, write the module, write the tests. It finished in about 52 seconds, provisioned its own test environment when it found a piece missing, and passed 8 of 8 tests — which we then re-ran ourselves, with five extra cases, because the harness’s word is never the verdict here. The full account, including the embarrassing gotcha that made our first attempt look like an API failure when it was nothing of the kind, is in The Meta Model API, As We Run It.

The whole catalog answers. The same week, we probed every text model on the account with tiny live calls: all six responded, and the total probe bill came to under two cents. Muse Image generated at a flat cent. The voice and segmentation surfaces are catalogued but deliberately unprobed — documented, not claimed.

What we have not proven — labeled thin on purpose

• The three-way bake-off has not run. Composer in Cursor vs. Spark 1.3 in Cursor vs. Wonder through the API, on the same real work order, scored the same way — that is this desk’s first scheduled experiment, and it is unrun. Until it runs, this desk makes no claim about which seat wins.
• Wonder’s long-run reliability is unknown — to everyone. Independent benchmark data for Wonder was one day old when we researched it. There is no long-run independent study of Wonder in autonomous coding loops, ours included. Anyone who tells you otherwise is selling something.
• Muse Code is unrun by us. Meta’s own terminal agent looks complete on paper. We have not put it through a sitting, so this desk reports its plans and features as Meta’s claims, not our findings.
• Production volume is untested. A $25 envelope proves an API works. It does not prove rate limits, latency, or behavior at a real shop’s daily load, and we will not pretend two cents of probing says otherwise.

The desk’s pages

• The Meta Model API, As We Run It — the proof run, the gotcha, and what one API speaking three dialects means for a shop’s harness choices.
• Meta Model API Pricing & Limits — every tier priced, the Cursor split explained, the Muse Code plans labeled as reported, and the cheap tier that trains on your data, warned about in plain language. This is the desk’s drift-watch page: when Meta’s numbers move, this page is where the change lands.
• Wonder Full Stays in the Seat. The API Key Is a Different Promise. — the seat-versus-API decision: Wonder Full lives only in the Power and Maximum seats, the subscription credential works only in the Muse Code CLI, and the seat defaults to Contributor mode, which trains on the prompts, code, and files you send. Why the setting gets changed before a client repository is in the context.