Meta's Muse Ember: a $0.10/Mtok agent model that beats Haiku 5.5

Meta's lightest Muse is multimodal with a 1M-token window, but the $0.10 rate triples past 200K. Plus e-means hits FAISS and Gemini's Drive-scanning flap.

Nowline NOV 10 4:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Muse Ember undercuts Haiku and Luna on agent work

    Meta Superintelligence Labs shipped Muse Ember, its “lightest and most cost-efficient model yet,” at $0.10 per million input tokens and $0.50 output — multimodal (text, image, audio), with a 1M-token context and up to 512K-token outputs, served through Meta's Model API. It's the executor half of a planner/executor pair with the bigger Muse Wonder. On Artificial Analysis it beats Claude Haiku 5.5 and OpenAI's Luna on agentic tasks: AutomationBench 37.8 (vs 13.9 / 20.2), Coding Agent Index 46 (vs 37 / 31), Terminal-Bench 4.0 40.9 (vs 39.2 / 16.4), and it ties Haiku at 36 on AA Intelligence Index v5. For high-volume agent loops, that's roughly Haiku-grade behavior at a fraction of the price.

  • The 200K catch: the cheap rate doesn't last

    The headline $0.10 input price only holds under 200K tokens. Above that, Ember roughly triples to $0.30 input / $1.50 output (cached input $0.012), per Meta's own Muse Code pricing. What it means for you: the 1M-token window is real, but long-context agent runs quietly cross into a pricier tier, so budget for the jump before you point Ember at a whole repo or a giant transcript.

  • e-means: a ~10x faster k-means, already in FAISS

    Meta FAIR dropped e-means, a drop-in replacement for Lloyd's algorithm that uses mini-batch EMA updates with auto-tuned batch and step sizes. It reaches lower-than-Lloyd's loss in 1.2–3 epochs instead of 25 — about 10x fewer epochs and wall-clock, with 10–45x GPU speedups over Flash-KMeans and 8–15x on CPU over Faiss. It ships now in FAISS: `from faiss.contrib.e_means import Emeans`. If you cluster embeddings, it's a one-line swap you can test this weekend.

  • Gemini reportedly reads Drive files without clear consent

    Users report Google's Gemini scanning and summarizing Google Drive files — PDFs included — without explicit permission and with no obvious off switch; Google says the access isn't unauthorized. What it means for you: if you've connected Workspace to Gemini, audit what it can see before trusting it with client or private docs, and double-check your Workspace and Labs data settings.