Muse Ember: API pricing, agent tasks and model comparisons

Meta positions Muse Ember for repeated, bounded work: tool calls, routing, extraction, document parsing and coding subtasks. A larger model such as Muse Spark or Muse Wonder can plan a workflow and delegate the many small jobs to Ember. That division is useful when a task has a clear input, a checkable output and enough repetition for lower per-call costs to matter.

Muse Ember is available through Meta Model API and Muse Code. Meta documents the hosted model ID muse-ember-1.0; its current availability guide lists API pay-as-you-go and Muse Code plans. The documented API accepts text, image, video, audio and PDF input and returns text, with a 1,048,576-token context window and up to 524,288 output tokens. A tool-using agent still needs an application to call tools, check results and decide what to do next.

Create a Meta Model API key, set it as MODEL_API_KEY, and send a request to the Responses endpoint. Existing OpenAI-compatible clients can use https://api.meta.ai/v1 and the same model ID. Its reasoning effort can be set from minimal to max, with max as the documented default.

Start with a small, repeatable task and inspect the returned usage. For an agent, record the tokens and tool calls for the full run, including retries, then divide the cost by tasks that pass your checks.

Meta’s listed USD API prices depend on the total input tokens in a request, including cached tokens and any conversation history carried forward. Up to 200,000 input tokens, uncached input costs $0.10 per million, cached input $0.004 per million and output $0.50 per million. If a request exceeds 200,000 input tokens, the higher rates apply to the whole request: $0.30 input, $0.012 cached input and $1.50 output per million.

For 80,000 uncached input tokens, 20,000 cache-hit tokens and 10,000 output tokens below the threshold, the estimate is $0.008 + $0.00008 + $0.005 = $0.01308. If the request instead has 250,000 input tokens, including 50,000 cached tokens, plus 10,000 output tokens, the long-context estimate is $0.060 + $0.0006 + $0.015 = $0.0756. These examples cover one API request; add every agent turn, search query and retry in a complete workflow.

A cache discount applies only to matching cached-prefix tokens. Check cached_tokens in the usage response. Meta’s web search grounding tool adds $2.50 per 1,000 queries on top of token charges. Team rate limits are shared across API keys; the current table lists 3,000 requests and 4,000,000 tokens per minute for muse-ember-1.0.

Choose Ember for high-volume steps with a narrow brief and a checkable result. Meta describes Spark as a model for coding and agentic workflows and Wonder for harder planning and reasoning. For hard single-shot coding problems, knowledge-heavy questions or visual coding, Meta recommends Spark or Wonder. The practical pattern is to keep the larger model on the planning loop and use Ember for repeated, well-scoped work.

The API price difference is substantial: Muse Spark 1.3 Standard is $1.25 input and $4.25 output per million tokens, while Muse Wonder is $4 and $20. For 100,000 uncached input tokens plus 10,000 output tokens, that is $0.17 on Spark 1.3, $0.60 on Wonder and $0.015 on Ember, before any search queries. Compare the accepted result and total agent run; one extra repair loop can erase a per-call saving.

Artificial Analysis currently displays Muse Ember at 36 on Intelligence Index v5.0.1, with a listed cost of $0.14 per index task. In Meta’s November 9 max-effort chart, Ember scored 35.8 at $0.136 per task, Claude Haiku 5.5 scored 36.2 at $0.411, and GPT-6 Luna scored 27.8 at $0.152. Artificial Analysis rounds Ember and Haiku to 36.

The evaluations use different tasks and setups. Meta’s methodology describes internally run competitor results as best-effort. For Terminal-Bench 4.0, its comparison uses mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna. Those harnesses affect results, so the scores do not isolate model quality alone.

In a November 9 report, AI for Mortals staff describe five everyday jobs run twice through OpenRouter at Ember’s default max effort: a 48-page NIST framework Q&A, spreadsheet cleanup, email reply, ten-slide HTML deck and tip calculator. The two runs of each job cost about $0.017–$0.018 total and took 47 seconds of cumulative waiting. Their checks found complete NIST answers and spreadsheet output. One email omitted the agreed fee and contradicted its laptop instructions; the calculator’s rounded per-person amounts did not add up to its displayed total.

On the same prompts with Claude Haiku 5.5, the max-effort run cost $0.116; two of five tasks hit a 60,000-token cap without returning an answer. At high effort, all five completed for $0.024. The effort setting and output cap differed between those runs.