What Muse Ember, Meta’s model for AI agents, costs per task

Built for repetitive AI agent tasks, Muse Ember has a million-token context window. API rates triple above 200,000 input tokens.

At the same base rates for input and output tokens, Muse Ember and Claude Haiku 5.5 produced different bills. At their highest reasoning settings, the two models scored 35.8 and 36.2, respectively, on the Artificial Analysis Intelligence Index. Meta’s published chart puts their estimated average costs at $0.136 and $0.411 per benchmark task. At its highest reasoning setting, GPT-6 Luna scored 27.8 and cost $0.152 per task.

According to Artificial Analysis, Meta’s model generated about a third as many output tokens as Haiku across the benchmark. Reusing context also cost less. At the base rate, a million cached input tokens costs $0.004 for Ember, compared with $0.01 for Haiku. Cached input accounts for a substantial share of the bill in tests where agents carry out a series of actions.

The overall scores were close, but results varied by task. On FrontierCode 2.0, which evaluates whether code changes are ready to merge into a project, Haiku scored 34.9% and Ember scored 27.2%. A million-token context window and two pricing tiers

The Meta Model API documentation lists a context window of 1,048,576 tokens and a maximum output of 524,288 tokens. The model accepts text, images, audio, video, and PDFs, and returns text. An agent could use it to extract fields from a scanned document or analyze a screenshot attached to a bug report.

According to Meta’s API pricing, requests with up to 200,000 input tokens are billed at $0.10 per million uncached input tokens and $0.50 per million output tokens. Above that threshold, those rates rise to $0.30 and $1.50, while cached input rises to $0.012 per million tokens. The higher rates apply to the entire request, including its output. Cached input and any conversation history carried forward through the API also count toward the threshold. A model for carrying out subtasks

Available in Muse Code and through the Meta Model API as `muse-ember-1.0`, Ember is designed for operations an agent repeats during a session, such as reading files, running tests, and making targeted fixes. In a code migration example published by Meta, Muse Wonder drew up the plan and assigned 16 Ember agents to rewrite Python lexers, which split source code into tokens, in Rust.

Meta says Ember processed 158 million input tokens in that run at an estimated cost of $1.84. Wonder handled planning and oversight, processing 4.4 million input tokens at a cost of $2.12. Those figures reflect that specific run at Meta’s pay-as-you-go rates.