Meta Launches Muse Ember, Its Smallest Model, Built for Agent Subtasks
On November 9, 2026, Meta launched Muse Ember (muse-ember-1.0) on its Meta Model API and in Muse Code: $0.10 / $0.50 per million tokens, a one-million-token context window, and the first model trained using Muse Wonder Full.
Meta launched Muse Ember on November 9, 2026, introducing it on its research blog as its “smallest, most efficient model.” It is not aiming for the top of the leaderboards: it is built for the repetitive grind that agents churn through all day, such as tool calls, coding subtasks, query routing, data extraction, and analyzing long documents. Mark Zuckerberg announced it the same evening on X.
Availability and Pricing
Muse Ember is available starting today on the Meta Model API and in Muse Code, under the identifier muse-ember-1.0, with support for the Responses API. Meta advertises $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. The model features a one-million-token context window, can generate up to 512,000 tokens, and natively understands text, image, audio, and video.
According to Meta, it is also the first model trained using Muse Wonder Full. It puts into practice the strategy announced with the Wonder series, which was already evident in the specialized models deployed in Muse and Meta AI: a large frontier model is used to train smaller, cheaper-to-run models. The lineup now features three tiers: Muse Wonder for the hardest problems, Muse Spark for code and agentic workflows, and Muse Ember for raw volume.
What the Benchmarks Show
The most solid evidence comes from a third party. Artificial Analysis gives Muse Ember (max effort) a score of 36 on its Intelligence Index, matching Claude Haiku 5.5 (max), for roughly $0.14 per task versus $0.41. The cost difference does not stem from list pricing—which matches Claude Haiku 5.5 and GPT-6 Luna on input and output—but from significantly cheaper cached inputs and shorter completions: according to the testing firm, Ember generates about three times fewer output tokens than Claude Haiku 5.5. Muse Wonder (max) remains eight points higher (44), but at $6.95 per task.
In its own coding agent, Muse Code, Ember scores 46 on Artificial Analysis’s Coding Agent Index, compared to 37 for Claude Haiku 5.5 in Claude Code and 31 for GPT-6 Luna in Codex. The picture is not uniformly positive: on FrontierCode 2.0, which evaluates real-world engineering tasks based on whether they can be merged, Ember (27.2%) still lags behind Claude Haiku 5.5 (34.9%). Its world knowledge is also weak: 21.1% accuracy on AA-Omniscience, primarily because it refuses to answer 64% of questions. On the flip side, its hallucination rate (17.4%) is, according to Artificial Analysis, about half that of Claude Haiku 5.5.
The other figures cited in the post, such as Agents’ Last Exam (61.6%), OSWorld 2.1, and the Personal Superintelligence Index, are internal Meta evaluations, as are the speedups reported by test partners (Box, Cline, Oracle, Cloudflare, You.com).
Limitations
Meta does not disclose the model size or provide downloadable weights: Muse Ember is a closed-source model, accessible only through the API and Muse Code. On the safety front, Meta reports evaluating it under its internal framework (chemical and biological risks, cybersecurity, loss of control) and placing it below high-risk thresholds, which remains a self-assessment.
Our take
Muse Ember steps into today’s most fiercely contested category: small, inexpensive models that power sub-agents. Its selling point is not a lower per-token list price than Claude Haiku 5.5 or GPT-6 Luna, but a lower cost per task driven by aggressive prompt caching and terser responses—a critical factor for agents that repeatedly reread the same context. Independent benchmarks from Artificial Analysis lend credibility to the overall index claims, but they also expose its blind spots: limited general knowledge and a noticeable step back on more realistic coding tasks. For a developer, the most logical setup is the one Meta itself suggests: a large model handling planning while Ember executes the volume. One obstacle remains: you have to route calls through Meta’s API, which is still nascent compared to platforms from Anthropic, OpenAI, or Google.
Sources
- Meta AI Research — Introducing Our Smallest, Most Efficient Model: Muse Ember (November 9, 2026) Primary source
- Mark Zuckerberg (@finkd) on X — Muse Ember announcement (November 9, 2026) Primary source
- Artificial Analysis (@ArtificialAnlys) on X — Muse Ember evaluation (November 9, 2026) Primary source — Third-party benchmark: Intelligence Index, cost per task, Coding Agent Index, AA-Omniscience
- Neural Brew — Meta rolls out specialized models trained with Muse Wonder across Muse and Meta AI (November 6, 2026) Secondary source — Context: the strategy of specialized models trained with Muse Wonder
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
