Meta’s Muse Ember targets lower-cost AI agent tasks

Originally reported by InfoQ on November 10, 2026.

Meta has released Muse Ember, a smaller AI model designed for the repeated work that can make AI agents expensive to run: tool calls, coding subtasks, task routing, information extraction and long-document analysis. It is available as muse-ember-1.0 through Meta Model API and Muse Code, with support for the Responses API.

Meta’s Muse Ember targets lower-cost AI agent tasks (2)

Meta lists prices of $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It says the model supports a one-million-token context window, up to 512,000 output tokens, and understanding of text, images, audio and video.

Meta’s Muse Ember targets lower-cost AI agent tasks (3)

Meta’s Muse Ember targets lower-cost AI agent tasks (4)

The cost comparison is central to Meta’s pitch, but it is tied to specific tests and settings. According to results Meta cited from Artificial Analysis, Muse Ember scored about 36 on its Intelligence Index at the highest reasoning setting—level with Claude Haiku 5.5 and above GPT-6 Luna’s 28. Meta put the cost per task at approximately $0.136 for Muse Ember at a near-maximum intelligence setting, versus $0.411 for Claude Haiku 5.5 at its highest setting.

Meta’s Muse Ember targets lower-cost AI agent tasks (5)

Meta’s reported coding results were mixed. Muse Ember scored 40.9% on Terminal-Bench 4.0, compared with 39.2% for Claude Haiku 5.5. On FrontierCode 2.0-AA, which assesses engineering tasks and whether code can be merged, Claude Haiku 5.5 led with 34.9% against Muse Ember’s 27.2%. Those results do not establish that the smaller model can replace a larger one across all coding work.

Meta’s Muse Ember targets lower-cost AI agent tasks (6)

Meta’s Muse Ember targets lower-cost AI agent tasks (7)

Early partner feedback also came from specific workloads. Cloudflare CTO Dane Knecht said its tests found tail latency up to 30% lower than for the comparable models it tested. You.com inference lead Chak Pothina reported a 2.2-second P95 latency on its tested coding tasks. Neither figure is a general latency guarantee.

Meta’s Muse Ember targets lower-cost AI agent tasks (8)

Meta describes a division of labor in which Muse Wonder handles difficult planning while Muse Ember performs frequent execution steps. The practical question for developers is whether that combination can lower the accumulated cost of agent workflows without sacrificing the quality their particular tasks require.

Meta’s Muse Ember targets lower-cost AI agent tasks (9)

Source: infoq.cn (originally published 2026-11-10 15:22 UTC+8)