Meta Puts a 10-Cent Model on Agent Busywork: Muse Ember Undercuts Haiku and Luna on Cost per Task

Meta Superintelligence Lab released a new model on Monday called Muse Ember, and the headline number is the price: 10 cents per million input tokens, 50 cents per million output tokens, and just four-tenths of a cent per million for cached input. Meta describes Ember as the lightest and most cost-efficient model it has ever shipped, and it is priced to compete not only with frontier small models like Anthropic's Haiku and OpenAI's Luna but with open-source workhorses such as DeepSeek V4.1 Flash and GLM 5.3 Flash.

Ember is not trying to be a general-purpose brain. Meta says it was "designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." Developers get it as muse-ember-1.0 through Muse Code and the Meta Model API, with support for the Responses API. The weights stay closed, in line with Meta's other Muse releases.

Where Ember breaks from most small models is input flexibility. It natively handles text, images, audio and video, with a one-million-token context window and up to 512,000 tokens of output. An agent built on Ember can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, without routing to a separate multimodal model.

The release is explicitly built around a two-tier architecture. Meta's suggested pattern pairs Ember with its frontier Muse Wonder model: "Muse Wonder plans, Muse Ember handles the volume." In a customer-support scenario, the heavier model would establish the rules for handling an issue and Ember would apply them across tens of thousands of cases. Meta also says Ember is the first model it trained using Muse Wonder Full, its own training pipeline.

The benchmark claims, which are Meta's own and have not been independently verified, lean on Artificial Analysis's composite index. There Ember scores about 36, the same rounded score as Haiku, at roughly $0.14 per task where Haiku costs nearly three times as much; Luna sits close on cost but scores 28. On AutomationBench-AA, a test of end-to-end business workflows, Ember scored 37.8 percent against 20.2 for GPT-6 Luna and 13.9 for Claude Haiku 5.5. On OSWorld 2.1, a computer-use test, Ember took 69.5 percent partial credit and finished 33.8 percent of tasks outright, edging Haiku's 68.9 and 27.8.

Coding is where the story gets less tidy. Ember beat Haiku on Terminal-Bench 4.0 at 40.9 percent, but on FrontierCode 2.0-AA, which asks whether a maintainer would merge an agent's code, Ember managed 27.2 percent behind Haiku's 34.9 and DeepSeek's 28.5. Meta also published no numerical results for its tests of everyday deployment behavior, such as resisting hidden instructions or seeking human input before high-impact actions, saying only that Ember stayed below its high-risk thresholds in the categories it assessed.

Ember is the fourth model under the Muse brand in seven months, following Spark in April, the open-source Glimmer in August and the frontier Wonder in October. The consumer-facing Muse Agent app launched in September and quickly became the number two app in the US, with collaboration Spaces and video chat rolling out earlier this month. The cadence makes clear that Meta is building a full model portfolio rather than chasing a single flagship.

The competitive signal is aimed squarely at the cost side of the agent market. If a major lab prices a purpose-built execution model at the same level as open-source alternatives while keeping commercial support and a closed stack, the price-per-task line becomes the battleground for the next wave of agent deployments, and every frontier lab now has to answer for what its smallest model costs.