Meta
Meta prices Muse Ember at $0.10 per million input tokens
Meta made Muse Ember generally available at $0.10 per million input tokens, and says it matches Claude Haiku 5.5 on one index at a third of the cost.

Meta has made Muse Ember, its smallest and most efficient model, generally available on the Meta Model API and in Muse Code. It costs $0.10 per million input tokens, and Meta says it matches Claude Haiku 5.5 on the Artificial Analysis Intelligence Index at about a third of the cost.
Meta introduced the model Nov. 9 in a post on its research blog. It is built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing.
Muse Ember Is Generally Available on the API and in Muse Code
Meta said Ember is “available today” on the Model API, which is general availability. It bills pay-as-you-go there and is included in Muse Code on every subscription plan. The model ID is muse-ember-1.0, and Meta lists it among its API-hosted models, not the open-weight ones.
Prices Triple for Prompts Above 200,000 Tokens
Meta charges three times as much per token once a prompt passes 200,000 input tokens. As of Nov. 9, Meta’s developer blog lists $0.10 per million input tokens and $0.50 per million output tokens up to that size, and $0.30 and $1.50 above it. On a Muse Code subscription, long requests use up plan usage faster.
Ember accepts the same request body as Muse Spark, so switching is mostly a change of model ID.
Meta Reports Ember Ahead of Haiku on Two Agent Benchmarks
In Meta’s chart of Artificial Analysis Intelligence Index data from Nov. 9, Ember scored 35.8 at $0.136 per task and Claude Haiku 5.5 at max scored 36.2 at $0.411. Meta’s chart notes the Ember cost figures are estimates. Reasoning effort, the setting for how long the model thinks before answering, runs from “minimal” to “max,” and the default is “max.”
On AutomationBench-AA, which tests end-to-end business workflows, Meta reported Ember at 37.8%, GPT-6 Luna at 20.2% and Haiku at 13.9%. On Terminal-Bench 4.0, which tests agentic terminal coding, Meta reported Ember at 40.9%, Haiku at 39.2% and Luna at 16.4%.
Haiku leads on FrontierCode 2.0-AA, which judges real engineering tasks on whether the result is mergeable: Haiku 34.9, Ember 27.2, Luna 25.2, per Meta’s table. All of these figures are Meta’s.
Meta Sends the Hardest Work to Spark and Wonder
Meta’s docs tell developers to use Muse Spark or Muse Wonder for hard single-shot coding, knowledge-heavy questions and visual coding. Meta’s recommended setup has Spark and Wonder plan and reason in an agent’s main loop while Ember takes the repeated subtasks.
Our View
We think the setting to watch is the default. Meta’s docs set reasoning effort to “max” unless a developer changes it, and Meta’s $0.136 per task for Ember was measured at max. Developers running agent subtasks who leave the setting alone get the behavior those figures describe.
Developers who lower the setting to save money step outside Meta’s numbers and have to measure the cost and the quality on their own work. Switching from Spark is mostly a change of model ID, so Ember can be tried on existing subtasks quickly.
What developers give up is hard coding. Meta sends hard single-shot coding to Spark or Wonder, and Haiku leads Ember on FrontierCode 2.0-AA, so a subtask that looks like real engineering belongs on the larger model. Muse Code subscribers have a second limit, because requests above 200,000 input tokens use up plan usage faster.
The low price pays off only while Ember gets the repeated subtasks Meta built it for and the prompts stay under 200,000 input tokens.
