Meta launches Muse Ember; token rates triple above 200K input

Meta released Muse Ember on November 9 for the repeated execution steps inside AI agents: tool calls, coding subtasks, routing and document extraction. Available through Meta Model API and Muse Code, it is the first model trained using Muse Wonder Full, according to Meta. Its role is to handle the volume of smaller jobs within a larger workflow.

API requests with up to 200,000 input tokens cost $0.10 per million uncached input tokens and $0.50 per million output tokens. Above that threshold, rates triple to $0.30 and $1.50 respectively.

Conversation history and reasoning affect the bill

Cached tokens and stored conversation history count toward that threshold. The higher rates apply to the entire request. Separately, the API automatically reuses matching prompt prefixes to reduce input cost and time to first token. Its caching guide places stable instructions and examples before variable content to extend reuse. Responses report the amount of input served from cache.

Ember also defaults to max, the highest of its six reasoning settings. Meta's reasoning documentation says higher effort increases reasoning tokens, latency and cost. Internal reasoning consumes the same output budget as the visible answer and is billed at the output-token rate. The advertised output ceiling consequently covers both thinking and the final response.

The model supports a 1,048,576-token context window and up to 524,288 output tokens. It accepts text, images, video, audio and PDFs, returning text. Meta describes classification and extraction across large collections as supported workloads. Its documentation also lists structured output, search grounding and computer use among the supported capabilities.

Developers can select muse-ember-1.0 using the request format already used for Muse Spark. Supported endpoints include Responses, Chat Completions and Messages. Meta presents direct API calls for custom apps and agents alongside Muse Code, its ready-made terminal coding agent. Ember is available with pay-as-you-go access and every Muse Code subscription plan. Meta says requests above 200K input tokens consume subscription usage faster, without specifying a fixed multiplier.

Independent evaluator Artificial Analysis reported an Intelligence Index score of 36 at max reasoning, costing about $0.14 per index task. Claude Haiku 5.5 at max reached the same score for about $0.41. Those task costs reflect token consumption and pricing within that evaluation, not a fixed charge for arbitrary real-world jobs.

Results were mixed on specific tasks. On FrontierCode 2.0, Ember scored 27.2%, against Haiku's 34.9%. Artificial Analysis also reported 21.1% accuracy on its knowledge evaluation, with Ember declining 64% of questions.

Meta's model-selection guide assigns planning and long-running, complex development to larger models, with Ember handling bounded subtasks. It lists classification, repeated tool calls and delegated coding among Ember's workloads. For hard single-shot coding problems and knowledge-heavy questions, the guide directs developers to Muse Spark or Muse Wonder.

Access routes also have different data settings. Ember is Standard-mode-only on Model API. Muse Code subscriptions default to Contributor mode, which provides more usage in exchange for permission to train future Meta models on submitted content. The subscription documentation says disabling that setting applies from the next request; future content stays out of training unless the setting is enabled again. The preference is saved per computer and must be changed separately on each device.