Meta launches Muse Ember, a small model priced like GPT-6 Luna and Claude Haiku 5.5

Meta describes Muse Ember as its smallest and most efficient model, designed for complex reasoning and for the repeated work agents do, such as tool calls, coding subtasks, routing, extraction and long document parsing. It accepts up to a million input tokens. Pricing rises above a threshold: up to 200k tokens it costs $0.10 per million input tokens and $0.50 per million output, while longer prompts cost $0.30 input and $1.50 output.

Those shorter-prompt prices match GPT-6 Luna and Claude Haiku 5.5, though each of those rises at a different threshold. Simon Willison argues that price per token is a weak comparison because reasoning tokens vary between models, so he points to the Artificial Analysis Intelligence Index against cost per task. Muse Ember sits on that chart's Pareto line and on the efficiency frontier of the full leaderboard. Claude Fable 5.2, released the same day, costs $8 input and $40 output per million tokens.

Ember offers six reasoning levels: minimal, low, medium, high, xhigh and max. In Willison's pelican-riding-a-bicycle SVG test, the high level failed to return a complete SVG, and output quality did not map cleanly to the reasoning level. He says the Ember results are comparable to those from Muse Wonder, which costs 40 times as much.

For agent builders, Muse Ember is a candidate for high-volume steps such as tool calls, routing and extraction, where cost per task can matter more than the per-token price. Compare total cost on your own workload, since prices change above the 200k threshold and token use depends on the reasoning level. Test the reasoning levels yourself, because the high level failed to produce a complete output in one test.