Meta's Muse Ember Targets the Cheap Execution Layer of AI Agents
Meta has released Muse Ember, a new AI model that the company describes as the lightest and most cost-efficient model it has shipped so far. TechCrunch reported the release on 9 November 2026. The model is priced at 10 cents per million input tokens.
Meta positions Ember for the high-volume, low-complexity work that surrounds an agent's core reasoning. The stated targets are tool calls (requests a model makes to external software such as a search engine or calendar), coding subtasks, routing, extraction, and long document parsing. That list describes the execution layer of an agent stack rather than a general-purpose flagship, and the pricing is set accordingly.
Meta presented Ember as cheaper and smarter than Anthropic's Haiku and OpenAI's Luna on cost-per-task, using Artificial Analysis's composite benchmark score as the basis for the comparison. Meta also says Ember is cost-competitive with the open-source models DeepSeek V4.1 Flash and GLM 5.3 Flash. Cost-per-task and composite scores depend heavily on the task mix and on how many tokens each model consumes to finish a job. Engineers evaluating the claim will want to rerun the comparison on their own workloads.
Ember is designed to sit alongside Muse Wonder inside an agentic harness, the scaffolding that manages a model's tool use, memory, and loop of actions. Wonder handles planning and Ember handles volume, a planner-executor split that many agent builders already use, with a larger model directing a cheaper one. The economics of that pattern depend on how often the planner must be invoked versus how often the executor runs, and Meta has not framed that ratio.
Ember accepts text, image, and audio inputs. Meta lists screenshots, charts, scanned documents, voice memos, meeting recordings, and video among the supported inputs, all handled without a separate model. A single checkpoint covering that range removes the need to route each modality to a specialised perception model and then stitch the outputs together. Whether quality holds evenly across every modality is a separate question, one that per-modality evaluation numbers would answer.
Ember is the fourth model released under the Muse brand, following Muse Spark, Glimmer, and Muse Wonder. The Muse name is also attached to a consumer-facing product. Reuters has reported that Muse is a personal AI agent that books trips, makes purchases, and places calls on a user's behalf, and that it has topped U.S. app download charts since launch. In September, Reuters reported that Meta was testing a “human concierge” for Muse, in which human contractors handle some of the calls placed through the agent. That detail matters for anyone building a comparable product, because it indicates that some agent tasks still route to people rather than to a model, at least during testing.
In my view, the human concierge test and the Ember launch point at the same engineering problem from different ends. A cheap, multimodal execution model lowers the cost of each step an agent takes. Human fallback covers the steps where the agent fails or where a live person is needed. The gap between a convincing agent demo and a dependable agent in production is largely the distance between those two things, and per-token pricing does not close it on its own.
The broader context here is a shift in where competition is happening. Frontier reasoning models compete on capability at the top of the range, while models like Ember compete on how cheaply an agent can execute its many small calls. For developers, that lowers the marginal cost of running agents over long documents, large meeting archives, and multi-step workflows. For Meta, a cost-efficient model that fits its own Muse agent offers a path to deploy agents at consumer scale without paying frontier prices for every step. How the market judges the price-performance claims will become clearer once independent evaluators and production users have run Ember against their own tasks.
