Skip to main content
Product2026-11-10•7 min read

Muse Ember 1.0 Is Now on Multos — Meta's Smallest Model Beats Haiku at a Third of the Cost

Meta trained Ember using Muse Wonder Full and pointed it at the work agents repeat all day. It scores 36 on the Intelligence Index against Haiku 5.5's 28, at $0.10/$0.50. Live on Multos Lite and Free today.

O
Oghenekaro Oboido

Muse Ember 1.0 is Meta's smallest model and it is on Multos Lite and Free. It scores 36 on the Artificial Analysis Intelligence Index against 28 for Claude Haiku 5.5, at $0.10 per million input and $0.50 per million output. On Multos that is $0.043 per million blended, which is the cheapest competent model in the catalog.

What a Cheap Competent Model Unlocks

The builds Ember changes are the ones you would not have attempted because the token math did not work. A batch job that processes a thousand records overnight. A multi-agent app where five workers each handle a slice of a task. A code review bot that runs on every pull request. A test generator that reads a module and writes cases for it. A migration tool that walks a large codebase file by file. A scraper-and-summarizer that runs on a schedule. None of these are exotic. All of them are expensive on a frontier model because they make many small calls rather than a few large ones. Ember makes the arithmetic work.

The Efficiency Frontier Claim

Meta's claim is that Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task. The supporting numbers are specific. Ember reaches a score of 35.8 at $0.136 per task. Claude Haiku 5.5 needs $0.411 to reach 36.2, and GPT-6 Luna tops out at 27.8 for $0.152. That is the whole argument in one line: comparable intelligence, roughly a third of the cost. It is worth noting these are Meta's own estimates at proposed pricing, so treat the exact figures as directional rather than gospel.

What Partners Measured

The third-party numbers are more interesting than the vendor ones. You.com reported a p95 latency of 2.2 seconds on coding tasks, which they measured as 4x to 13x faster than leading open-weight models, with 36% to 48% fewer output tokens. Cloudflare saw it run up to 30% faster at the tail than a comparable model and called it a strong fit for high-throughput agentic workloads. Cline's founder described it as purpose-built for subagent work. On benchmarks, Ember scores 37.8 on AutomationBench against 20.2 for GPT-6 Luna and 13.9 for Haiku 5.5, and 61.6 on Agents' Last Exam against roughly 57 for both rivals.

Where It Does Not Shine

Ember is a small model and it behaves like one on open-ended work. It is strong at bounded tasks with clear structure: extraction, routing, tool calls, small edits, test generation. It is weaker when the task requires deciding what the structure should be. If you hand it an ambiguous feature request and expect it to design the approach, you will be disappointed. The right pattern is to use a stronger model for planning and Ember for execution, which is exactly how Auto mode already routes. Do not expect Ember to replace a frontier model on architecture work.

What It Costs on Multos

Ember is a Lite model, available on Free and Lite. At $0.10 input, $0.50 output, and $0.004 cached, the blend is $0.043 per million, a 0.09x token multiplier. That is not a typo. On Lite you get 30M tokens a month, and at this rate a build that would exhaust a frontier model's budget in a week can run all month. It is also the model that makes subagents practical on the lower tiers, where previously the cost of spawning workers made the pattern unaffordable.

The Case for Auto Mode

The most useful thing about Ember is not that it is cheap. It is that it is cheap and competent enough that routing to it is not a downgrade. Most agent turns are file reads, greps, small edits, and structured extraction, where a frontier model is wasted spend. Ember handles those at a fraction of the cost with meaningfully fewer output tokens. Auto mode classifies tasks and routes accordingly, and Ember gives that routing a much better floor to land on. If you have been avoiding multi-agent builds because the token cost made them impractical, this is the model that changes the arithmetic.

Ready to Get Started?

Muse Ember 1.0 is live on Multos Lite and Free. 36 on the Intelligence Index against Haiku 5.5's 28, at $0.043 per million blended. Open the model dropdown, find it under Meta, and try the batch job or multi-agent build you had written off as too expensive.

Start Building Free