AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Meta releases Muse Ember, a new lightweight model for agent execution

Collected Nov 9, 2026

Meta Superintelligence Lab released a new AI model on Monday called Muse Ember, describing it as the lab's lightest and most cost-efficient model to date. Meta frames Ember around agent execution: in its release it said the model is built for the repetitive work agents do continuously, listing tool calls, coding subtasks, routing, extraction and parsing long documents as target workloads.

Pricing sits at 10 cents per million input tokens. Meta positions that against frontier competitors including Anthropic's Haiku and OpenAI's Luna, and the release notes Ember is also cost-competitive with open-source options such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Rather than leaning on raw capability claims alone, Meta leaned on cost-per-task, using Artificial Analysis's composite benchmark score to argue Ember delivers more intelligence per dollar than those rivals.

Muse Ember is meant to run as part of a two-tier system rather than solo. Meta's pitch pairs it with Muse Wonder, the lab's frontier model, inside an agentic harness: the heavier model plans and sets policy, while Ember applies the plan at scale. Meta's own phrasing captures the split, with Wonder planning and Ember handling the volume. A customer-support agent is the illustrative case — a larger model defines how a particular issue should be handled, then the lighter model applies those rules across tens of thousands of cases.

That architecture is a familiar pattern in agent design. Most production agent stacks already separate an orchestrator or planner from smaller executor models that make many more calls per user request, because the executor's token volume dominates the bill. Ember's role is to be the cheap workhorse at the bottom of that stack. The practical consequence for developers is that the model's value depends less on how it performs on a single hard prompt than on whether it holds accuracy across millions of small, repetitive steps — and on how reliably it follows the instructions a planner hands down. The release does not detail context window size, output token pricing or throughput limits, so teams evaluating it will want their own measurements on their specific workloads.

Meta also built Ember to accept multiple media types, including text, images and audio. The stated aim is that agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, without needing to route to a separate model for each format. This suggests input breadth is being treated as a cost feature as much as a capability one: handling a scanned invoice or a recorded call in the same model that extracts fields from it removes a handoff, and handoffs in agent pipelines are where latency and error rates tend to accumulate. The flip side, as with most lightweight multimodal models, is that per-modality accuracy at small parameter counts often trails dedicated larger models, so teams should expect to validate parsing quality on their own document types.

Ember is the fourth model under the Muse brand. It follows Spark, a mid-weight model released in April; Glimmer, described as Spark's open-source twin, released in August; and the Wonder frontier model, released in October. The consumer-facing Muse Agent shipped in September and, per the report, found unexpected success, with Spaces and video chat features rolling out earlier this month. That sequence is notable: Meta built out a flagship, a mid-tier and an open variant before adding the cheap high-volume tier, which is the piece that makes a multi-model agent product economically viable at consumer scale.

Why it matters: Developers building agents have been constrained less by whether a model can perform a step than by what it costs to run that step millions of times, so a 10-cent-per-million-input-token executor with built-in image and audio input meaningfully widens the set of workloads that can be shipped profitably — my inference, based on the stated pricing and positioning. The competitive pressure lands on mid-tier and budget models from Anthropic, OpenAI and open-weight vendors, and product teams should now treat planner-plus-executor pairing as the default architecture to evaluate. The open question is accuracy at that price point on messy real-world documents and audio, which only per-workload testing will settle.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

At 10 cents per million input tokens, Ember is significantly cheaper than frontier competitors like Haiku and Luna.