Cheap models rarely get the headlines, but they decide who makes money on agents. Meta Superintelligence Lab seems to get that. On Monday it released Muse Ember, its lightest and most cost-efficient model so far, pitched at the repetitive work agents grind through: tool calls, coding subtasks, routing, extraction, and long document parsing.
The pricing is the story. Ember costs 10 cents per million input tokens. Meta leaned on Artificial Analysis’s composite benchmark score to argue that Ember is both cheaper and smarter than Anthropic’s Haiku and OpenAI’s Luna. That’s a vendor-picked comparison, so treat it with some caution until independent testing lands. But the price gap alone is hard to ignore. Ember is also cost-competitive with open source options like DeepSeek V4.1 Flash and GLM 5.3 Flash, which is where most budget-conscious developers have been looking.
The design logic matters more than the benchmark chart. Meta expects Ember to sit beside a larger model, Muse Wonder, inside an agentic harness. Picture a customer-support agent: the heavy model works out the rules for a type of issue, then hands the lighter model the job of applying them across tens of thousands of cases. In Meta’s words, “Muse Wonder plans, Muse Ember handles the volume.” This planner and worker split is already common in agent stacks, and it’s where inference bills pile up. Cutting the worker tier’s cost is the most direct way to make large agent deployments affordable.
Ember is also multimodal. It handles text, images, and audio, so agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model. For developers, that means fewer moving parts. Anyone who has stitched together a transcription model, a vision model, and a text model knows how much that simplifies things.
Ember is the fourth model under the Muse brand. Spark, a mid-weight model, came in April. Glimmer, its open source twin, followed in August, and the Wonder frontier model arrived in October. The consumer-facing Muse Agent launched in September to unexpected success, with Spaces and video chat added earlier this month.
So is this noise? Not quite. Meta now has a full tier lineup, from frontier to lightweight, which is what Anthropic and OpenAI already offer. The open question is Ember’s licensing. Glimmer is open source, but Meta hasn’t said Ember is. If it stays closed, it competes with Haiku and Luna on price alone. If it opens up, DeepSeek and GLM should worry. Either way, the small-model price war just got another combatant.
