Meta launches Muse Ember for high-volume agent work

Meta's smallest Muse model costs $0.10 per million input tokens and matches Claude Haiku 5.5 on one index at a third of the cost per task.
What it means for founders
- Price your agent by the task. Ember's advantage depends on short outputs and cheap cache hits, so a pipeline that reuses a long system prompt gains the most. Replay a day of real traces through it before you move traffic.
- Keep requests under 200K tokens. Agents that load whole repositories or long stored conversations pay triple on the entire request, and a `previous_response_id` chain counts toward the limit. Trim context or split jobs.
- Leave shipping code to a stronger model. The FrontierCode gap points Ember toward reading files, routing and extraction, with Haiku or a larger model writing pull requests.
- Weigh the lock-in. Ember runs on Meta's servers with no self-hosting option, so a startup that builds on it, or on Meta AI more broadly, depends on Meta keeping its pricing and terms.
The story
Meta released Muse Ember on Monday, November 9, 2026, the smallest model in its Muse family, built for the repetitive steps inside AI agents such as tool calls, routing, extraction and coding subtasks. Developers can call `muse-ember-1.0` through the Meta Model API or use it in Muse Code, at $0.10 input and $0.50 output per million tokens. Meta is keeping the weights private, The New Stack reports.
A worker for a bigger planner
Meta wants teams to split agent work by difficulty: Muse Wonder or Muse Spark plans, and Ember runs the many small calls that follow. "Muse Wonder plans, Muse Ember handles the volume," Meta wrote, as TechCrunch reports. Meta says Ember reads text, images, audio and video and answers in text. The New Stack puts its context window at 1M tokens with up to 512K output tokens.
Same list price, cheaper per task
TechCrunch describes Ember as significantly cheaper than frontier rivals such as Anthropic's Claude Haiku and OpenAI's GPT-6 Luna. The New Stack and Artificial Analysis both say its input and output rates equal those two models at standard rates. The savings come from cached input, which Meta bills at $0.004 per million tokens against $0.01 at its two rivals, and from fewer output tokens: Artificial Analysis counted roughly a third as many as Haiku 5.5 produced.
On Artificial Analysis's Intelligence Index, Ember at max effort scored 36 for about $0.14 per task. Haiku 5.5 (max) reached the same rounded score, 36, at $0.41 and sits slightly ahead before rounding, and Luna scored 28 at $0.15, per The New Stack.
Meta's pricing page adds a condition. Any request carrying over 200K input tokens, cached ones counted, bills the whole request at long-context rates: $0.30 input, $0.012 cached input and $1.50 output per million tokens. Ember runs only in Standard mode, where Meta says it does not train on your content.
Where it wins and where it trails
From Artificial Analysis's own runs:
- Terminal-Bench 4.0: Ember 40.9%, Haiku 5.5 32.8%.
- FrontierCode 2.0, which asks whether a maintainer would merge the change: Ember 27.2%, Haiku 34.9%.
- Coding agents: Ember in Muse Code scored 46 on the firm's Coding Agent Index, ahead of Haiku in Claude Code (37) and Luna in Codex (31).
- Knowledge: Ember declined to answer 64% of AA-Omniscience questions, which left it the lowest accuracy among peers but a hallucination rate half of Haiku's.
What we don't know yet
- Whether Meta will publish Ember weights, as it did with Glimmer.
- How Ember performs on production workloads, since Meta ran some launch comparisons itself.
- Whether hosts beyond Meta will serve it. OpenRouter lists the model with one provider today.
Sources
Tools in this story
We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.
Meta's free AI assistant across its apps and glasses
More in Models & Labs
- Google releases Gemini 4 Carbon, previews /swarm

For founders: Budget for the December price. Cost models built on the $2 and $10 rate break after November 30.
Google DeepMind · 8d ago - Anthropic releases Fable 5.2 and Mythos 5.2

For founders: Price the switch by task cost. At twice the Opus 5.5 rate, Fable needs to finish jobs with fewer tokens or retries to save money.
The Verge · 8d ago - OpenAI's GPT-6.1 Terra halves per-token API rates

For founders: Rerun your agent costs. At $1 in and $5 out per million tokens, high-volume chores like invoice entry and form filling get cheaper.
OpenAI · 23d ago - Block unveils Dessa, voice AI trained only on speech

For founders: A simpler voice stack may be coming. Many voice products today chain transcription, a language model and text to speech from separate vendors such as…
The Verge · 25d ago - Recursive launches Flint, a coding model for laptops

For founders: A local coding model with no token bill. Teams that cannot send code to a cloud API, or that watch inference spend, have a free option worth trialling on a…
TechCrunch · 27d ago