Meta Muse Ember: 10-Cent AI That Powers Cheap Business Agents

Meta Muse Ember is a small, cheap AI model built for the repetitive work AI agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing. Launched on 9 November 2026, it costs $0.10 per million input tokens and $0.50 per million output tokens, and Meta claims it matches Claude Haiku 5.5 on its headline benchmark at about a third of the cost per task.

What Meta Muse Ember actually is

Ember is the smallest and most efficient model in Meta's Muse family, built for volume rather than brilliance. A flagship model thinks hard about one difficult problem; Ember does thousands of small jobs quickly: calling a tool, checking a record, classifying a message, or pulling a figure out of a long document. It takes text, images, audio and video, has a 1-million-token context window with up to 512,000 output tokens, and offers six reasoning levels, minimal to max. The model ID is muse-ember-1.0.

It is available now on the Meta Model API and in Muse Code, and is also listed on OpenRouter. Muse Wonder, Spark 1.3 and Ember are now on Google Cloud and Oracle Cloud. Early partners named by Meta include Box, Cline, Oracle, Cloudflare and You.com. If you are weighing the chatbot-versus-agent choice, Ember sits firmly on the agent side.

The price: ten cents in, fifty cents out — and the cliff

The headline price is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004 per million. The same day, Anthropic launched Claude Fable 5.2 at $8 per million input and $40 per million output tokens, so Ember's output token costs 80 times less. Fable is the heavy thinker for hard coding; Ember is the cheap worker for everything else — see our Claude Fable 5.2 coverage.

Now the catch. The $0.10/$0.50 price only applies up to 200,000 input tokens. Above that, the price triples to $0.30 input and $1.50 output — and it reprices the whole request, not just the tokens above the line. Keep prompts under the line, or budget for the higher tier.

Why the cost per task matters more than the sticker price

Ember's per-token prices are identical to Claude Haiku 5.5's and OpenAI's GPT-6 Luna's. Ember does not win on a lower sticker price; according to Meta, it wins by using fewer tokens per job. A model that writes shorter answers and calls tools more efficiently costs less per completed task at the same price per token. The saving is not guaranteed: a badly written prompt or a loop that retries too often will eat it. Our guide to AI calling agent costs makes the same point.

What Meta claims it can do

Meta's headline claim, on the Artificial Analysis Intelligence Index v5, is that Ember scores 36 — 35.8 at maximum reasoning — matching Claude Haiku 5.5's best score of 36.2 at about a third of the cost per task: $0.14 against Haiku's $0.41. Meta also says Ember beats OpenAI's GPT-6 Luna at every reasoning setting. These are Meta's and Artificial Analysis's claims, not independently verified facts.

Meta also says Ember leads on AutomationBench-AA and the Coding Agent Index-AA, which fits its design. But on FrontierCode 2.0, Ember trails Haiku 5.5 — 27.2% against 34.9% — so it is not better at everything. The sensible reading is that Ember is a specialist: excellent at routine agent work, not a replacement for a flagship on difficult problems. The same pattern showed up in the GDPx agent benchmark, where no model cleared roughly 12% on long-horizon tasks.

The team-of-models idea: Wonder plans, Ember does the volume

Ember is the first model trained using Muse Wonder Full, Meta's largest model, and Meta's framing is simple: "Muse Wonder plans, Muse Ember handles the volume." The big model works out the plan; the small model carries out the hundreds of small steps.

That is also the industry story of the week: the same day Ember launched, Anthropic launched Fable 5.2 at flagship prices. The direction is a team of models, each priced for its job. Meta's own Muse Voice follows the same logic, and LG's K-EXAONE 2.5 pushes in a similar direction with a free API tier.

Where cheap agents change the maths for a business

At $0.14 per task, jobs that were too expensive to automate start to make sense: a support inbox with 500 messages a day, a sales team that needs every lead qualified and routed, an operations team pulling figures out of long supplier documents. Our guide to AI agents for small business automation covers the same ground.

• Support triage: classify, prioritise and draft replies, then hand the hard cases to a person.
• Document extraction: pull dates, amounts and obligations out of contracts and long PDFs — while watching the 200,000-token price cliff.
• Coding subtasks: small fixes, test writing and code review, with a stronger model checking the result — see our how to build an AI voice agent guide.
• Routing: decide which team or workflow each message belongs to, and attach the right customer record first.

The saving is only real if the agent is reliable enough to run without constant supervision. A model that costs a tenth as much but needs a person to check every answer has saved nothing. For clinics and other regulated work, our guide to AI chatbots for clinics covers consent, data handling and escalation.

What Ember cannot do

No parameter count has been disclosed, the weights have not been released, and Ember is API-only — you cannot download it or self-host it. If your business needs a model on its own servers, for data or compliance reasons, Ember is not an option. That is the opposite of open-weight releases, and of Meituan's LongCat-2.5, aimed at long-range agent tasks. Ember was evaluated under Meta's Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks — a safety evaluation, not a guarantee. The model is days old and the tooling will change fast, so do not rebuild a core system around it. The same week, Nvidia showed GR00T 2, a robot model that learns from human video — another sign AI is moving from chat into physical work.

How to try it without betting the business

Start with one narrow job and a fixed budget. Pick a task you already measure — support triage is a good one — and run Ember beside your current setup for two weeks. Score it like a new hire: did it classify correctly, call the right tools, and hand over cleanly? Track the cost per completed task, not per token.

Wire it into your existing automation rather than building from scratch. If your workflows already run on n8n, Make or Zapier, Ember slots in as the cheap worker inside the steps you already have — our comparison of n8n vs Make vs Zapier shows how those tools differ. And if you want help scoping that pilot, talk to Coheser about an AI agent pilot or contact us.

Frequently asked questions

What is Meta Muse Ember?

Ember is Meta's smallest and most efficient Muse model, launched on 9 November 2026 for the repetitive work AI agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It is API-only, with a 1-million-token context window.

How much does Muse Ember cost?

$0.10 per million input tokens and $0.50 per million output tokens, up to 200,000 input tokens, with cached input at $0.004 per million. Above 200,000 input tokens the price triples to $0.30/$1.50 and reprices the whole request.

Is Ember really as good as Claude Haiku 5.5?

On Meta's headline claim, via the Artificial Analysis Intelligence Index v5, Ember scores 36 against Haiku 5.5's 36.2 at about a third of the cost per task. But it trails Haiku on FrontierCode 2.0, 27.2% against 34.9%. Test it on your own work before switching.

Can I download or self-host Ember?

No. The weights have not been released and no parameter count has been disclosed. Ember is available through the Meta Model API, in Muse Code, on OpenRouter, and on Google Cloud and Oracle Cloud.

Should my business switch to Ember now?

Not wholesale. Run a small pilot on one narrow, high-volume job, measure the cost per completed task and the accuracy on your own data, and keep a stronger model for the hard cases.