Meta launches Muse Ember, a 10 cent model built for agent subtasks

Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens on the Meta Model API, where it is listed as muse-ember-1.0[1]. It has a 1 million token context window and can produce up to 512K output tokens, and it supports the Responses API.

Meta says that works out to about $0.14 per Artificial Analysis benchmark task, against $2.39 for its own Muse Spark 1.3, while scoring 36 on the Artificial Analysis Intelligence Index, one point behind Spark[1].

Meta's headline claim is that Ember matches Claude Haiku 5.5's best score at about a third of the cost and beats every reasoning setting of OpenAI's GPT-6 Luna[1]. At maximum reasoning effort, Meta reports:

• 37.8% on the AutomationBench business workflow test, against 13.9% for Haiku 5.5 and 20.2% for GPT-6 Luna.
• 40.9% on Terminal-Bench 4.0, narrowly ahead of Haiku's 39.2%.
• 69.5% on OSWorld 2.1 computer use with partial credit, against 68.9% for Haiku.

Ember does not win everywhere. Haiku 5.5 leads on FrontierCode 2.0, which judges whether code changes would be merged, at 34.9% to Ember's 27.2%. Several results, including OSWorld and Meta's Personal Superintelligence Index, are internal evaluations rather than independent ones.

Ember is the first model Meta trained using Muse Wonder Full, its largest model, as it said it would when Wonder launched[1]. Meta positions the line-up as Wonder for the hardest problems, Spark for coding and agent workflows, and Ember for repeatable tasks at scale. It also suggests pairing them, with Wonder planning and Ember handling the volume.

The model handles text, images, audio and video natively. Early partners quoted by Meta include Box, Cline, Oracle, Cloudflare and You.com, with Cloudflare reporting tail latency up to 30% faster than a comparable model it tested.

Agent workloads spend most of their tokens on small, structured steps, so the price of the cheapest capable model sets the cost of running agents at scale. A frontier lab pricing a 1 million token context model at 10 cents per million input tokens puts pressure on Anthropic, OpenAI and Google at the low end, where most agent traffic actually runs.