Muse Ember cuts agent costs, but long context triples the bill

Meta positions Muse Ember as a small model for tool calls, coding subtasks, routing, extraction, and long document processing. Short requests cost $0.10 per million input tokens, $0.004 for cached input, and $0.50 per million output tokens.

The low price ends at 200,000 tokens

The model has a 1 million token context window and can return up to 512,000 tokens. Once a request exceeds 200,000 input tokens, Meta charges the higher rate for the whole request: $0.30 for input, $0.012 for cache, and $1.50 for output. Ember is available through Meta Model API and Muse Code as muse-ember-1.0.

Meta says the model reaches 36 on the Artificial Analysis Intelligence Index at maximum reasoning effort, at roughly $0.14 per task. Simon Willison notes that token price alone is inadequate because models can spend very different numbers of reasoning tokens on the same problem.

A cheap executor changes multi-agent economics

For teams that use one strong model to plan and smaller models to perform hundreds of steps, Ember is most interesting as the executor. Savings compound in routing, extraction, and tool use, where an agent repeats short, bounded operations.

The practical consequence is less glamorous than another benchmark record. Agent designers can separate expensive decisions from cheap execution and measure quality at the boundaries between models.

The million-token window invites expensive convenience

Marketing highlights 1 million tokens, but the pricing threshold punishes applications that attach an entire repository or history to every request. The evidence also comes mainly from Meta and Artificial Analysis. Production evals for each workflow still need to establish real tool-use reliability.

Cost per completed task will settle the argument

Teams should track retries, failed tool calls, reasoning-token use, and the share of requests above 200,000 tokens. Ember can reduce agent operating costs if it finishes more tasks without correction and context stays below the threshold. Frequent escalation to a larger model would erase the attractive list price.

Lilith's verdict

Ember is a cheap runner in the agent relay. Load the full archive onto its back for every leg, and it reaches the finish line carrying a triple bill.