Meta ships Muse Ember: Haiku-level agents at a third the cost

Live on the Meta Model API at $0.10/Mtok with 1M context and native video; it tops Haiku on agent tasks but trails on coding. Plus: Copilot Chat's new caps.

Nowline NOV 10 1:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Haiku-level agents at $0.10 per million tokens

    Meta's smallest model, `muse-ember-1.0`, is live today on the Meta Model API and in Muse Code: $0.10/M input ($0.004 cached), $0.50/M output, a 1M-token context window with up to 512K output, native text/image/audio/video, and computer use from screenshots. It's built for the high-volume work agents repeat all day — tool calls, routing, extraction, long-doc parsing, and small coding subtasks like reading files and running tests — at about $0.14 per Artificial Analysis task versus ~$2.39 for the larger Muse Spark.

  • It ties Haiku's score — but read the coding numbers

    Ember matches Claude Haiku 5.5 on the Artificial Analysis Intelligence Index (36, vs GPT-6 Luna's 28) at roughly a third of Haiku's cost per task, and leads Meta's own AutomationBench (37.8% vs 13.9%), Terminal-Bench 4.0 (40.9% vs 39.2%) and OSWorld. But Haiku still wins the real-world FrontierCode 2.0 coding test (34.9% vs 27.2%), many wins come from Meta's internal benchmarks, and the weights are private — so benchmark your own workload before you swap out your cheap model.

  • Microsoft Copilot Chat gets secret usage caps

    As of Nov 9, Microsoft applies dynamic, unpublished fair-use limits to Copilot Chat for users without a Microsoft 365 Copilot license in commercial tenants over 2,000 seats — across web, desktop and mobile, including image generation. The numbers aren't published and admins can't view or change them, so if you've been leaning on free Copilot Chat, it's now throttled on terms you can't see.