Introducing Our Smallest, Most Efficient Model: Muse Ember

Meta's new small model, Muse Ember, has the same base list price as OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5: $0.10 per million input tokens and $0.50 per million output tokens. Simon Willison points out that the match only holds for shorter prompts. All three accept up to a million input tokens, and each charges more once a prompt passes a size threshold, but the thresholds and the higher rates differ.

Luna's higher rate starts above 272k tokens, Ember's above 200k and Haiku's above 100k. Past those points Luna charges $0.20 per million input tokens and $0.75 per million output, Ember $0.30 and $1.50, and Haiku $0.50 and $2.50. Cached input, the earlier part of a conversation that is re-sent and billed at a discount, costs $0.004 per million tokens for Ember below its threshold, against $0.01 for the other two, and $0.012 above it. Taking the thresholds into account, Willison places Ember's pricing between the other two.

Willison's larger point is that price per token says little on its own, because reasoning models can spend very different numbers of tokens on the same problem. He looks instead at Artificial Analysis's chart of its Intelligence Index against cost per task, where Ember sits on the Pareto line, the boundary at which one measure can only improve by giving up the other. Meta's announcement makes the related claim that no model on that leaderboard scores higher for less per task, and Artificial Analysis's own write-up explains why Ember's per-task cost is low despite matching list prices. Willison also ran his informal test of drawing a pelican riding a bicycle at Ember's six reasoning levels: the high setting failed to return a complete image, and the quality of the drawings did not track the reasoning level cleanly.