Meta quietly added a fourth model to its
Muse family on Monday, November 9 — Muse Ember, its smallest and most efficient model yet. It is not a flagship. It is the workhorse: the model built to handle the thousands of small, repetitive jobs AI agents do all day — tool calls, coding subtasks, routing requests, pulling data out of long documents — at a price that undercuts its rivals. And Meta says it matches Anthropic's Claude Haiku 5.5 on benchmarks while doing it.
What Muse Ember Is
Ember is designed for the volume layer of agentic AI. When a big frontier model plans a task, it hands the small steps to cheaper subagents — and Ember is built to be that subagent. Meta's own pitch sums it up: "Muse Wonder plans, Muse Ember handles the volume." It is live now as muse-ember-1.0 on the Meta Model API and inside Muse Code, Meta's coding agent, with Responses API support. It is also listed on OpenRouter and now available through Google Cloud. It takes text, images, audio and video as input, with a 1 million-token context window and responses up to 512,000 tokens. One thing Meta did not do: release the weights or a parameter count. Like the earlier Muse releases, Ember is API-only.
What It Costs
The pricing is the headline. Ember costs $0.10 per million input tokens and $0.50 per million output tokens, with cached input at just $0.004 per million. On the Artificial Analysis benchmark, Meta puts the cost at about $0.14 per task — against $0.41 for Claude Haiku 5.5, roughly a third of the price. For teams running agents at scale, where thousands of small calls add up fast, that gap is the whole pitch.
How It Stacks Up
Meta evaluated Ember on the Artificial Analysis Intelligence Index, which aggregates nine benchmarks. The headline number: a score of 36, matching Claude Haiku 5.5's best result at maximum reasoning effort, and beating OpenAI's GPT-6 Luna at every reasoning setting. Meta says Ember outperformed both rivals on seven of the nine benchmarks, with its biggest edge in business automation — a 17% lead on AutomationBench-AA, a set of 657 tasks spanning apps like Gmail and Salesforce. One honest caveat: these are Meta's own numbers, from its evaluation report and Artificial Analysis data. Independent testing of the full suite is still to come. Ember is also the first model Meta says it trained using Muse Wonder Full, its largest model — the flagship teaching the workhorse. Early partners named by Meta include Box, Cline, Oracle, Cloudflare and You.com. Anthropic's own
Claude Fable 5.2 landed the same day.
Why It Matters
The model race is shifting. The question is no longer just "which model is smartest" but "which model is cheapest per job" — because real agents make thousands of small calls a day. A model that is nearly as capable at a third of the price wins the volume, and that is exactly the lane Ember is built for. Paired with Muse Wonder as the planner, it gives Meta a full two-tier agent stack to sell against Anthropic's Claude family and OpenAI's lineup. The verdict: not a flashy flagship, but the kind of launch that quietly changes what it costs to run AI agents. Whether developers actually switch will come down to reliability on real agent loops — pricing is only half the story. Meta bought
Conductor days earlier to manage coding agents side by side.
Comments
Post a Comment