Meta's Muse Ember costs $0.14 a task, a third of Claude Haiku, as Anthropic prices its IPO
Meta's smallest model does the work agents repeat all day at $0.14 a task, matching Anthropic's Haiku at a third of the cost. Anthropic prices its IPO this week on about $81 billion of run-rate token revenue, and the floor under that business just moved.
Richard Tangthree-minute read
Meta+0.60% — Meta, up 0.60 percent today released Muse Ember yesterday, its smallest model, and the price is the news: $0.10 per million input tokens on the Meta Model API, aimed at the steps software agents repeat all day, from tool calls and coding subtasks to routing, extraction and long-document parsing 1.
The independent benchmarking service Artificial Analysis has already measured it: 36 on its intelligence index at $0.14 per task 8. On Meta's comparison, that matches Claude Haiku 5.5's best score at about a third of the cost per task and beats every GPT-6 Luna setting 1. The volume end of inference just got repriced.
Meta set the floor on cost per agent task
Ember costs a third of Haiku per task and a fraction of Meta's own Spark
Data
| Value | |
|---|---|
| Muse Spark 1.3 (Meta) | $2.39 |
| Claude Haiku 5.5 | $0.41 |
| GPT-6 Luna | $0.15 |
| Muse Ember | $0.14 |
A cheap executor changes how agent systems are built
Ember was trained with Muse Wonder Full, Meta's frontier trainer, and its stated job is the subagent slot: "Muse Wonder plans, Muse Ember handles the volume" 1. Agents spend most of their tokens on small structured steps, so routing those steps to a cheap executor changes the unit economics of every agent fleet 1.
The partner accounts in Meta's announcement point the same way: Cline's founder says teams run Ember as the subagent to cut costs, Box says it scores contracts against negotiation playbooks clause by clause, and Cloudflare+4.85% — Cloudflare, up 4.85 percent today and You.com report speed and token savings in early testing 1. TechCrunch notes it is cost-competitive with open-source flash models from DeepSeek and GLM 6.
Adoption started same-day: the developer Simon Willison had Ember wired into his own command-line tooling, driven from his phone, before the release was a day old 2. The limit is visible too: on FrontierCode 2.0, engineering work judged on whether it merges, Haiku still leads 1.
Anthropic prices its IPO on the opposite bet
Anthropic$350B — Anthropic, private, latest valuation $350B set its IPO range this morning at $70 to $78 a share and plans to price on November 18, trading on Nasdaq as ANT 4. The filing behind it carries roughly $81 billion in run-rate revenue, most of it tokens sold to businesses 4.
A draft circulated to banks on Monday showed a raise of nearly $70 billion before Anthropic retracted it as inaccurate 5. The obligations underneath are the point: nearly $900 billion in computing commitments, which token sales have to fund 5. Haiku is the volume workhorse in that machine, and Artificial Analysis lists Anthropic's new flagship, Claude Fable 5.2, at $26.54 per task against Ember's $0.14 8.
Anthropic's flagship costs about 190 times Ember per task
Data
| Cost per task | |
|---|---|
| Claude Fable 5.2 | $26.54 |
| Claude Haiku 5.5 | $0.41 |
| Muse Ember | $0.14 |
Meta can hold the floor because it is paying with advertising money: about 97% of its revenue is still ads 37, and Truist+0.14% — Truist, up 0.14 percent as of its last quote models only $5.2 billion of Meta enterprise revenue by 2030 against a $461 billion enterprise spend pool 3. Meta is buying the volume tier, not harvesting it.
At 14 cents a task, contract review stops being a pilot and becomes a line item. Whoever holds the volume tier when fleets make that switch holds the default, and the revenue that follows. Anthropic's IPO investors are underwriting the opposite bet.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
