Meta Launches Muse Ember: $0.10/M Token Model Built for Agentic Subwork

Meta released Muse Ember on November 9, 2026 — the cheapest model in its Muse lineup, priced at $0.10 per million input tokens — directly targeting the high-volume, low-cost tier of the agentic AI market. The move puts Meta in direct price competition with Anthropic's Claude Haiku and OpenAI's Luna for the grunt work of AI pipelines: tool calls, routing, extraction, and long-document parsing.

What the Agentic Tier Actually Is

Most production AI systems aren't just one model. They're pipelines — a larger "reasoning" model at the top making decisions, and smaller "worker" models below handling repetitive subtasks: reading a database entry, extracting a field from a document, deciding which tool to call next. These subtasks run thousands of times per session and cost real money at scale.

This is the market Ember is built for. Meta calls it "agentic subwork" — the high-frequency, low-complexity layer of AI pipelines that needs to be fast and cheap, not powerful.

Ember's Specs

According to Meta's announcement, Muse Ember ships with:

• $0.10 per million input tokens — undercutting current small-model pricing from major competitors
• 1 million token context window — large enough to process full legal documents, codebases, or conversation histories in a single pass
• Optimized for tool calling, routing, extraction, and long-document parsing — the specific tasks that dominate agentic pipelines

Ember is the fourth model in the Muse family. It's designed to run alongside Muse Wonder — Meta's larger reasoning model — in what Meta calls an "agentic harness": Wonder handles planning and high-stakes decisions; Ember handles execution at scale.

The Price War Context

Small model pricing has been falling steadily throughout 2026. Anthropic cut Claude Haiku prices in October. OpenAI positioned Luna in a similar slot. Google's Flash tier has been the aggressive price leader in that range.

Ember at $0.10/M tokens is a direct shot across that competition's bow. For a business running 100 million agentic subtask calls per day — a realistic number for a mid-sized enterprise pipeline — the difference between $0.10 and $0.20 per million tokens is $10,000 per day, or roughly $3.6 million annually.

At that scale, model pricing is infrastructure pricing. And infrastructure pricing decisions are made by CFOs, not engineers.

What This Means for AI Builders

If you're building agentic systems today, Ember's release has three practical implications:

• Routing architecture just got cheaper. The standard pattern of "small model routes, large model reasons" becomes more economical with Ember at the base.
• Long-context parsing opens up. A 1M-token context window on a cheap model means you can feed an entire document corpus into a single Ember call rather than chunking and embedding. That simplifies pipelines significantly.
• Multi-vendor becomes realistic. Enterprises running mixed harnesses — Claude for reasoning, Ember for extraction, say — now have a credible sub-$0.15 option from a major lab with Meta's reliability track record.

The risk: Meta is a newer entrant to the managed API market. Enterprises will want to see uptime data and rate-limit guarantees before committing pipeline infrastructure to Ember.

What to Watch

Meta's pricing pressure will likely force a response from Anthropic and OpenAI within the next 60-90 days — either a price cut on existing small models or an announcement of next-generation efficiency improvements. Watch also for Ember's performance benchmarks on agentic tasks specifically; raw token pricing only matters if quality holds at the lower end.

Muse Ember is available now through Meta's AI API.