Muse Ember has no chat-app entry point and no downloadable open weights — there is exactly one way to use it: through an API or through Muse Code. This guide covers the five channels, the six reasoning levels and three task families, with integration notes, effort-level choices and ways around the 200K price cliff.
| Channel | Live since | Notes |
|---|---|---|
| Meta Model API | Launch day | Model ID muse-ember-1.0, endpoint api.meta.ai/v1, Responses API supported, same request body as Muse Spark |
| Muse Code | Launch day | All plans; home of the "Wonder plans, Ember executes" two-model pattern |
| Google Cloud | Launch day | Pricing and integration detail per official pages; not itemised in this round |
| Oracle Cloud | Launch day | Pricing and integration detail per official pages; not itemised in this round; Oracle is also a quoted launch partner |
| Vercel AI Gateway | Launch day | Projects already on the gateway switch model IDs; pricing per the gateway's page |
How to choose: call the model directly through the Meta Model API; use the built-in Ember if you already work in Muse Code; use the matching cloud if you already deploy there; use the gateway if you are already on the Vercel AI Gateway. There is no free tier — budget every channel from $0.10/$0.50.
The model ID is muse-ember-1.0, the endpoint api.meta.ai/v1. Two things to note: first, the Responses API is supported and the request body is the same as Muse Spark — a project already wired to Spark runs by changing the model ID, with no request-structure changes; second, Standard mode only, no Contributor mode. Context is 1,048,576 tokens, max output 524,288 tokens; input supports text, image, audio, video and PDF (native multimodal), output is text only. The partner quotes in the launch post (Box, Cline, Oracle, Cloudflare, You.com) are Meta-published, not third-party verification — judge by the official docs and the AA evaluation, not by the quotes.
Ember is available on all Muse Code plans, positioned as "Wonder plans, Ember executes": Wonder / Spark plan, decompose and make high-value judgements, while Ember takes sub-agent tasks, tool calls, routing, extraction and long-document parsing. In practice it runs on two layers: planning on Wonder / Spark, execution batched to Ember for repeatable work; Ember itself was trained with Muse Wonder Full, which stays locked inside Muse Code and is not offered externally. The split only pays if the price gap is real — a 17× per-task cost gap (Spark $2.39 vs Ember $0.14) makes "plan once, execute a thousand times" pencil out.
Six effort levels: minimal, low, medium, high, xhigh, max; default max. Meta published benchmark results only at max effort — lower-effort quality is officially unverified. Simon Willison's hands-on test (Nov 10) is the only measured warning: after trying all six levels, quality did not rise steadily with effort, and one request at high failed to return a complete drawing. A single-person, single-task observation, not a benchmark — but the practical takeaway stands: don't assume higher is always better, and don't assume dropping a level saves money for free. The working method: get critical tasks running on the default max first, then step repeatable tasks down one level at a time while spot-checking output quality, until you find your own sweet spot; spot-check drawing outputs for completeness in particular.
Requests over 200,000 input tokens bill the whole request at 3× ($0.30/$0.012/$1.50); input up 25% can mean the bill up 233% (see the pricing guide's worked example). Three ways around it: first, split the request — chunk long documents so each stays under 200K; second, shrink the context — send only the passages the task needs, using summaries or retrieval instead of the full text; third, raise cache hits — route unchanging long context through cached input ($0.004, ~60% cheaper than Haiku) — but note cached input also triples above 200K ($0.012), so the cache is no exemption from the cliff. Long-document parsing is one of Ember's named use cases and exactly the kind of work most likely to cross the line — count tokens before sending.
| Task | Fit | Basis |
|---|---|---|
| Sub-agent tasks, tool calls, routing, extraction | Good fit | Meta's named use cases; AA Index 36 ties Haiku at a third of the cost per task |
| Business-process automation | Good fit | AutomationBench-AA 37.8% vs Haiku 13.9% |
| Long-document parsing (under 200K) | Good fit | 1M context, native multimodal input; mind the 200K cliff |
| Terminal tasks, computer use | Worth trying | TB 4.0 leads Haiku by just 1.7 points, OSWorld by just 0.6 — a small edge |
| Code-heavy work | Not the first pick | FrontierCode 27.2% vs Haiku 34.9% |
| Long-context work over 200K input | Do the math first | Whole request at 3×; split or shrink the context before using it |
| Needs open weights / local deployment | Not a fit | Ember has no open weights; the open option is Glimmer / Spark Open |
Meta says Ember was evaluated under the Superintelligence Scaling Framework (chem/bio, cyber, loss-of-control), with no category exceeding the high-risk threshold; it was also evaluated for resisting instructions hidden in web pages, documents and tool outputs, and for seeking human input before high-impact actions. These are Meta's claims, not third-party verified — for high-impact actions (deleting data, moving money, messaging externally), don't rely on the model's self-restraint alone; add human confirmation or tightened permissions in the workflow. The partner quotes at launch (Box, Cline, Oracle, Cloudflare, You.com) are likewise Meta-published, not independent verification.
Meta Model API: model ID muse-ember-1.0, endpoint api.meta.ai/v1, Responses API supported, same request body as Muse Spark. Or use it directly in Muse Code on all plans; Google Cloud, Oracle Cloud and the Vercel AI Gateway also carry it.
Default max. Meta published results at max effort only, so lower-effort quality is officially unverified; Simon Willison's test found quality does not rise steadily with effort. For repeatable tasks, step down one level at a time while spot-checking quality to find your own sweet spot; spot-check drawing outputs for completeness.
Split requests and chunk long documents; shrink the context to the passages the task needs; raise cache hits. Note cached input also triples above 200K, so the cache is no exemption. Count tokens before sending.
No. Ember ships no open weights and is API-only. For local deployment or open weights, use Meta's Glimmer / Spark Open — a different model.
Reference: Meta research blog — channels, reasoning tiers, safety evaluation, partner quotes · Meta Model API docs (muse-ember-1.0, api.meta.ai/v1, Responses API) · Simon Willison hands-on test (six reasoning levels, pelican drawing test) · Artificial Analysis — Index 36, $0.14 per task (rewritten and organized by ModelBrief; facts per the Meta launch post and API docs)