Meta’s Muse Ember targets high-volume agents, with a long-context pricing catch

What Meta released

Meta introduced Muse Ember on November 9, 2026, positioning its smallest model for the repeated work inside agents: tool calls, routing, extraction and coding subtasks. It is available in Muse Code and through Meta Model API under the model ID muse-ember-1.0, with Responses API support.

The research announcement lists a one-million-token context window and up to 512,000 output tokens. Those ceilings describe capacity; they do not establish that filling the context is the most efficient way to solve a task.

Public record: Meta research: Introducing Muse Ember — November 9, 2026

The headline price has a context threshold

Meta’s model page lists standard rates of $0.10 per million input tokens and $0.50 per million output tokens for the lower context tier. Cached input costs $0.004 per million tokens. Above 200,000 input tokens, the listed rates rise to $0.30 input, $1.50 output and $0.012 cached input per million tokens.

The developer announcement explicitly describes the lower tier as applying to prompts up to 200,000 input tokens. Budget for the higher tier when a prompt exceeds that threshold. The model’s million-token capacity and its cheapest processing tier are different numbers.

For an illustrative call with 100,000 uncached input tokens and 10,000 output tokens, those lower rates imply $0.015 in model token charges. That is our arithmetic, excluding tools and other services, and it assumes the stated token counts.

Public record: Muse Ember model and pricing — checked November 11, 2026 · Meta: Build with Muse Ember — checked November 11, 2026

Where a smaller worker fits

Meta’s developer guide describes text, image, audio and video inputs, and says Ember is included in all Muse Code subscription plans. Its suggested architecture pairs Ember’s bounded execution with Muse Wonder or Muse Spark for planning and review.

That division still needs an application to manage it. The guide explains that a direct API call returns model output; the surrounding application executes requested tools and returns their results. Choosing a model does not supply that orchestration by itself.

Public record: Meta: Build with Muse Ember — checked November 11, 2026

Measure the work you can accept

Our suggested first trial is a small document-extraction queue with known answers. Give each item the same required fields, retain its source reference, and count how many results are correct without manual repair. Include unreadable and missing fields so a plausible guess cannot pass as success.

Compare total spend and elapsed time across the whole queue. Track retries, larger-model escalations and the minutes someone spends correcting output. Repeat the trial with only the relevant pages supplied, then with a larger document bundle, to see whether extra context earns its cost.

This is an evaluation method, not a claim that Promptwire has tested Ember. The release is interesting because inexpensive repeated calls can change workflow economics. The decision should follow a measured cost per accepted result on your own workload.