Meta launches Muse Ember, its cheapest model yet at 10 cents

Key Points

1. Meta Superintelligence Labs released Muse Ember, its lightest and most cost-efficient model, on Monday.
2. Ember costs 10 cents per million input tokens, undercutting Anthropic's Haiku and OpenAI's Luna.
3. The launch pushes Meta's competition with rivals toward price-per-task rather than frontier capability.

The latest:

Meta’s cheapest AI model to date went live Monday, priced at 10 cents per million input tokens and aimed squarely at the repetitive work that agents run at scale. Meta Superintelligence Labs said Muse Ember was built for complex reasoning execution — tool calls, coding subtasks, routing, extraction and long document parsing. It is the fourth model in the Muse line.

Details:

• The pricing claim: Meta said Ember undercuts frontier competitors including Anthropic’s Haiku and OpenAI’s Luna at 10 cents per million input tokens. The lab also positioned it as cost-competitive with open source rivals DeepSeek V4.1 Flash and GLM 5.3 Flash, a tier that usually wins on price alone.
• The benchmark used: To make the cost-per-task argument, Meta leaned on Artificial Analysis’s composite benchmark score, presenting Ember as both cheaper and smarter than its named competitors. The company did not publish a breakdown of how the composite score is weighted across individual tasks.
• The design logic: Ember is built to run inside an agentic harness alongside a heavier model such as Meta’s Muse Wonder. Meta’s framing of the split was blunt: “Muse Wonder plans, Muse Ember handles the volume.”
• The use case: Meta offered customer support as the illustration: a heavier model sets the rules for handling a given issue, then dispatches the lighter model to apply those rules across tens of thousands of cases. The appeal is unit economics, not raw capability.
• Multimodal inputs: With volume-heavy deployments in mind, Meta said it trained Ember across text, images and audio. According to the announcement, agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video without calling a separate model.
• The Muse lineup: Ember is the fourth release under the Muse brand, following the mid-weight Spark in April, its open source twin Glimmer in August, and the Wonder frontier model in October. The cadence puts four models inside roughly seven months.
• The consumer side: Meta’s flagship consumer product, the Muse Agent, launched in September to what the company described as unexpected success. New features including Spaces and video chat began rolling out earlier this month.
• Not disclosed: Meta did not announce availability terms, enterprise customers, or output token pricing alongside the input figure. No timeline was given for an open source counterpart to Ember, the treatment Spark received with Glimmer.

Background:

Meta Superintelligence Labs is the unit behind the Muse model family, which has expanded from a single mid-weight release in April to a four-model lineup spanning open source, frontier and lightweight tiers, plus a consumer agent product.

Between the lines:

The competitive framing has shifted. Meta benchmarked Ember against Haiku and Luna on cost-per-task rather than capability, and placed it in the same price band as DeepSeek and GLM’s open source flash models. Pairing Ember beneath Wonder suggests Meta expects agent deployments to be priced on volume economics, where the expensive model reasons once and the cheap one executes thousands of times.

What’s next

Watch whether Anthropic and OpenAI respond with price cuts on Haiku and Luna, whether Meta ships an open source twin to Ember as it did with Spark, and how Muse Agent’s Spaces and video chat rollout performs.