An agent looking for one clause can send the likely pages or a stack of contracts. With Muse Ember, crossing 200,000 input tokens changes the price of the request. Meta released the small model in Muse Code and on its Model API for repeated work such as tool calls, code fixes and document extraction. It supports a context window of about one million tokens. Meta’s launch announcement
For prompts up to 200,000 tokens, Meta lists $0.10 per million uncached input tokens and $0.50 per million output tokens. Above that line, the rates are $0.30 and $1.50. A 190,000-token uncached prompt therefore costs about two cents in input, while a 210,000-token prompt costs about six cents before the model writes a word. If your agent sends the same long document on every turn, check what the provider counts as cached input as well. Meta’s pricing footnote
Meta’s launch table reproduces Artificial Analysis results that put Ember at about $0.14 per Intelligence Index task at maximum reasoning effort, compared with about $2.39 for Meta’s Muse Spark 1.3 at its maximum setting. Those are costs for a particular benchmark and token mix. An agent doing contract intake can add its own costs in retries, escalations and review, especially if the first extraction misses a field. Meta’s benchmark comparison
Try both sizes on the same work. Give Ember the likely pages for one run and the full document for another, then have a reviewer check the extracted fields and their source passages. If the shorter request misses an exception buried in an appendix, count the time a reviewer spends finding it by hand.