Meta Releases Muse Ember, Its Smallest High-Throughput Model for Agent Work

According to Meta’s AI research site and Meta Model API documentation (November 9, 2026), Meta released Muse Ember, framed as its smallest and most efficient model for the high-token jobs agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing. The formal model ID is muse-ember-1.0. The same posts say it is available on the Meta Model API in Standard mode and inside Muse Code.

What it is for

Meta positions Ember as the high-throughput execution layer inside larger agent stacks: running focused sub-agent coding steps, driving many tool calls where cost and latency add up, classifying and routing requests, extracting fields, and chewing through long document sets. Hard single-shot problems are still steered toward larger Muse models; Ember is the piece meant to absorb volume.

Pricing (per million tokens)

For requests with ≤200K input tokens, Meta’s listed standard rates are $0.10 input, $0.004 cached input, and $0.50 output per million tokens. When input exceeds 200K tokens, long-context rates apply: $0.30 input, $0.012 cached input, and $1.50 output per million tokens. Those figures come from Meta’s research post and the Muse Ember pricing table on the Model API site. Muse Code subscription plan prices are separate from these API token rates and are not treated here as Ember token pricing.

Context, modalities, and APIs

Documentation lists a context window of 1,048,576 tokens (about 1M) and a maximum output of 524,288 tokens (about 512K). Inputs may be text, image, video, audio, and PDF; output is text. Muse Ember supports the Responses API, with Chat Completions and Messages also listed among endpoints.

Wonder / Spark / Ember portfolio

Meta’s research post says Muse Ember is the first specialized model trained with Muse Wonder Full. In the same framing, Muse Wonder covers the hardest problems, Muse Spark covers coding and agentic workflows, and Muse Ember covers high-throughput, repeatable execution—larger models can plan while Ember absorbs volume on repeated subtasks. On Google Cloud the Spark listing is written as Muse Spark 1.3 (bare ID muse-spark-1.3); that hosted API channel is separate from the open-weight Muse Spark Open story in a neighbor piece.

Benchmark claims (Meta’s wording)

The research post places Muse Ember on the Artificial Analysis task cost-efficiency frontier and says it matches Claude Haiku 5.5’s best score at about one-third the cost. Those comparisons are Meta’s own blog claims, not independent retests.

Third-party availability

OpenRouter lists the model as meta/muse-ember. Separately, OpenCode has stated that Muse Ember is available in OpenCode.

Separately, @finkd said on 2026-11-10 05:49:49 Asia/Shanghai that Muse Ember, Muse Spark 1.3, and Muse Wonder are now available on Google Cloud. Google Cloud documentation lists them on the Gemini Enterprise Agent Platform, with bare model IDs muse-ember-1.0, muse-spark-1.3, and muse-wonder-1.0 (pass the bare model ID in the model field). That Google Cloud channel sits beside the Meta Model API and OpenRouter’s prefixed ID meta/muse-ember—the OpenRouter string is not the same as the GCP bare IDs. Docs also note the Responses API on the global endpoint under publishers/meta. We haven't signed in to Google Cloud or called these models.

Related reading (different stories)

For Muse family open-weight news—Muse Spark Open weights under Meta’s own license, not this Ember API model release and not this Google Cloud hosted channel—see Meta's Muse Spark Open Is an 816B Open-Weight Model With Its Own License. A weaker neighbor on Muse shared-link discovery (a security/privacy story, not a model launch or this channel note) is Researchers Say Shared Meta Muse Conversations Can Be Found Without the Link.