New · November 9, 2026
Muse Ember is here — Meta's smallest model, $0.14 a task
A month after Muse Wonder, Meta filled the other end of the lineup: Ember, its smallest and most efficient model, built for the repetitive work agents do all day. It matches Claude Haiku 5.5 on the Artificial Analysis Intelligence Index — at about a third of the cost per task. Live November 9 on the Meta Model API and in Muse Code as muse-ember-1.0.
What shipped
💰 The cheap tier, priced to run everywhere
$0.10 per million input tokens, $0.004 cached, $0.50 output. Meta's headline comparison: Ember costs about $0.14 per Artificial Analysis task, versus $0.41 for Claude Haiku 5.5 and $2.39 for Muse Spark 1.3 — while scoring 36 on the Intelligence Index, one point behind Spark's 37. The per-token rates match Haiku and GPT-6 Luna; the savings come from cached input (about 60% cheaper than Haiku's, per Artificial Analysis) and shorter outputs. Per-task figures are benchmark estimates, not a real bill.
🔁 Built for the volume layer
Ember targets the work agents repeat hundreds of times a session — tool calls, coding subtasks, routing, extraction, long-document parsing. Meta's own line: agents spend most of their tokens on many small, structured steps, and Ember is built for exactly that. With a 1M-token context window and up to 512K output tokens, it can take in a full codebase or document set in one call. It natively understands text, images, audio and video — screenshots, charts, voice memos, meeting recordings — without a separate model.
🧠 First model trained by Wonder Full
Ember is the first model trained using Muse Wonder Full, Meta's largest model. When Wonder launched in October, Meta said it would use it to develop specialized models for everyone — Ember is the first. The intended team: "Muse Wonder plans, Muse Ember handles the volume." Meta says more efficient models are already in training. It is not open-weight: unlike Muse Spark Open, you cannot download it, and Meta has not disclosed a parameter count.
The benchmarks — where it wins, where it doesn't
Meta's table, Artificial Analysis data, November 9, 2026, all at maximum reasoning. Ember leads Haiku 5.5 on most agent and coding benchmarks — and loses clearly on real engineering tasks and general knowledge.
- AutomationBench-AA (business workflows): Ember 37.8 · Haiku 13.9 · Luna 20.2 — its biggest lead.
- Agents' Last Exam: 61.6 vs 57.0 (Haiku) vs 57.1 (Luna).
- Personal Superintelligence Index (Meta internal): 46.8 vs 36.1 vs 37.0.
- Terminal-Bench 4.0: 40.9 vs 39.2 vs 16.4. Coding Agent Index-AA: 46 vs 37 vs 31.
- FrontierCode 2.0 (real engineering, judged on mergeability): 27.2 vs Haiku 34.9 — Ember trails here.
- OSWorld 2.1 (computer use): 69.5 partial / 33.8 binary vs 68.9 / 27.8. PerceptionBench (visual, with tools): 80.6 vs 78.6 vs 62.8. GDP.pdf-AA (long documents): 22.8 vs 20.8 vs 22.8.
- Knowledge is the weak spot: Artificial Analysis puts Ember's Omniscience accuracy at 21.1% (Haiku 34.7%) and says it declines 64% of trivia-style questions. Use it for work you can check, not for answers you have to trust blind.
The honest read
- The story is cost per task, not intelligence. Ember's Intelligence Index score (36) is identical to Haiku's and one point below Spark 1.3's — the news is that the same score now costs $0.14 instead of $2.39. For agents that make hundreds of small calls per session, that changes what is worth automating.
- Wonder plans, Ember executes — and Meta is building a relay. The Wonder → Ember pipeline is now a product strategy, not a one-off: the frontier model trains the cheap model, the cheap model does the volume. Expect more specialized models on the same pattern — Meta says more are already in training.
- It is a developer launch, not an app launch. Ember arrived on the Model API and in Muse Code. Meta has not said it powers the consumer Muse app, and the app's November story remains Voice and Spaces. What reaches app users is the economics: cheap volume work is what lets a personal agent run hundreds of steps without the bill exploding.
- Safety is stated, not shown. Meta says Ember was evaluated under its Superintelligence Scaling Framework and stayed below the high-risk thresholds for chemical/biological, cybersecurity and loss-of-control risks, with the same core safety training as the bigger models — including resisting instructions hidden in web pages, documents and tool outputs, and seeking human input before high-impact actions. The evaluation report is Meta's own; treat the thresholds as its claim.
5 prompts to try Ember where it earns its keep
Ember is for the repeatable layer — use it in Muse Code or via the Model API as muse-ember-1.0, and let Wonder handle the planning. Keep a human check on anything high-impact.
The subagent router
You are the router for a coding agent. For each incoming task, decide: quick fix, test run, or file read — and hand it to the cheapest capable step. Do not plan architecture here; that goes to the planner model. Return a one-line routing decision plus the exact next action.
The contract screener
Score this contract against our negotiation playbook, clause by clause, and give one overall risk rating with the three clauses that moved it most. Flag anything ambiguous as 'needs human review' instead of guessing.
The long-document extractor
Read this document set (up to a million tokens) and extract every deadline, amount, and named owner into a table. Quote the source sentence for each row. If a value appears twice with different numbers, list both and flag the conflict.
The test-run fixer
Run the failing tests, read the errors, and make the smallest targeted fix for each one. Do not refactor anything that passes. Report each fix as: test name, root cause in one line, file and line changed.
The cost check
Before you run this workflow, estimate the token cost at Muse Ember pricing ($0.10 per million input, $0.004 cached, $0.50 output). If the same job on the planner model would cost more than 5x as much for the same output quality, use Ember and say why. If quality would drop, say so and use the planner.
Good to know
- Timeline: Announced Nov 9, 2026 (Meta AI Research, "Introducing Our Smallest, Most Efficient Model: Muse Ember") · available the same day on the Meta Model API and in Muse Code · also on Google Cloud, Oracle Cloud and OpenRouter.
- Price: $0.10 / $0.004 cached input / $0.50 output per million tokens · 1M-token context, 512K max output · Responses API supported · model ID
muse-ember-1.0. - Family: fourth Muse model — Spark (April), Glimmer (August), Wonder (October), Ember (November). Ember is the first trained using Wonder Full; Spark Open is the open-weight sibling, Ember is API-only.
- Early users: Box tested Ember on contract review and candidate screening — "score a contract against a company's negotiation playbook clause by clause and land on the right overall risk rating" (Yashodha Bhavnani, VP of AI Products, Box).
- New here? Set up Muse in 15 minutes → · Get 1B free tokens →
← Muse Wonder · Muse Spaces → · All news
Not affiliated with Meta. Sources: Meta AI Research, "Introducing Our Smallest, Most Efficient Model: Muse Ember" (Nov 9, 2026); Artificial Analysis Intelligence Index v5 (Nov 9, 2026); WSJ Tech, Media & Telecom Roundup (Nov 9, 2026); PYMNTS (Nov 9, 2026); CellCog and Unite.AI coverage (Nov 9, 2026). Benchmark figures are Meta's published table and Artificial Analysis data; per-task costs are estimates at list pricing.