Meta Launches Muse Ember, a 10-Cent AI Model Built for the Work Agents Repeat All Day

AI
Written by Methodical Millions
Colourful editorial illustration of a glowing amber ember crystal at the centre of an electric blue agent network, linked by a bright beam to a blue planning orb on a dark navy background

Meta's newest model isn't trying to be the smartest in the room. It's trying to be the cheapest one that can still do the job thousands of times a day, and that's where the AI agent money is heading.

By Methodical Jon | Nov 9, 2026

Meta Platforms (NASDAQ: META) released Muse Ember on Monday, November 9th 2026, calling it the "smallest and most efficient" model from Meta Superintelligence Labs. It costs $0.10 USD per million input tokens and $0.50 USD per million output tokens, according to Meta's announcement, and Meta claims it delivers "more intelligence per dollar" than Anthropic's Claude Haiku 5.5 or OpenAI's GPT-6 Luna "at any reasoning setting," as The Verge reported.

What happened

Meta says Ember is "designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." In other words, it isn't the brain that plans a task. It's the worker that carries out the hundreds of small steps an AI agent takes to finish one.

The pitch is built around cost per task. Meta used scores from the independent benchmarking firm Artificial Analysis to show Ember as cheaper and smarter than Haiku and Luna, TechCrunch reported. Meta says Ember matches Haiku's best score "at about a third of the cost." On Meta's chart, Ember and Haiku 5.5 both score 36 on the Artificial Analysis Intelligence Index at maximum reasoning, with Luna at 28. Ember's cost per task is about $0.14 USD, versus roughly $0.41 USD for Haiku at its top setting.

TechCrunch noted that Ember's price is also competitive with open-source rivals like DeepSeek V4.1 Flash and GLM 5.3 Flash.

Meta didn't claim a clean sweep. On its own results table, Ember led on agent workflow tests like AutomationBench-AA (37.8 vs. 13.9 for Haiku and 20.2 for Luna), but Haiku 5.5 beat it on FrontierCode 2.0, a test of real engineering tasks judged on whether the code could be merged (34.9 vs. 27.2).

How it works

Ember is meant to run alongside a bigger model, not replace it. "Muse Wonder plans, Muse Ember handles the volume," Meta wrote. TechCrunch gave the example of a customer-support agent: a heavier model sets the rules for handling an issue, then a lighter model applies those rules across tens of thousands of cases.

The model natively handles text, images, audio and video, so agents "can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model," Meta said. It has a 1 million-token context window and can produce up to 512,000 output tokens in one response, per Meta's developer documentation.

Ember is available now as muse-ember-1.0 on the Meta Model API and in Muse Code. Meta's pricing page lists a higher long-context tier for requests over 200,000 input tokens: $0.30 USD per million input tokens and $1.50 USD per million output tokens.

Meta says Ember is the first model trained using Muse Wonder Full. It's the fourth model under the Muse brand, after Spark in April, its open-source twin Glimmer in August and the Wonder frontier model in October, according to TechCrunch.

Why it matters

The AI race is splitting in two: the smartest frontier model, and the cheapest model good enough to run inside every product, all day. Agents make the second race bigger, because a single request can trigger dozens of model calls. A few cents per task adds up fast at that volume.

Meta has a lot riding on bringing that cost down. It spent $34.91 billion USD on capital expenditures in the third quarter alone, and its free cash flow was negative $2.12 billion USD, according to its Q3 2026 results. It expects 2026 capex of $135 billion to $145 billion USD. On the earnings call, Meta said it expects to improve compute cost efficiency by more than tenfold over the next year, partly through "more efficient models," the AP reported. Ember is a direct piece of that plan.

What to watch next

  • Independent testing. The head-to-head numbers come from Meta's charts, and some benchmarks are internal. Watch how Ember holds up in developer hands.
  • Price responses. If Anthropic and OpenAI cut prices on Haiku and Luna, Meta's cost edge shrinks quickly.
  • Meta's 2027 budget. Meta said it will share more on next year's spending at its fourth-quarter earnings call.

The money lesson

1. Use the right tool for the job. Meta isn't asking its best model to do everything. The expensive model plans and the cheap one executes. In business, don't pay senior rates for repeat work.

2. Unit cost decides who scales. At high volume, a small difference in cost per task becomes the difference between a profitable product and a money-losing one. Know your cost per unit before you scale.

3. Read the fine print on "cheapest." Ember's price triples for very long requests, and it trails Haiku on one coding test. Headline numbers rarely tell the whole story.

For a founder's view of building a business on top of AI, listen to Episode 87 - Spotlight: Konstantin Bayandin - founder of Tomi.ai. It's a conversation with someone who builds an AI product for a living, and cheaper models like Ember change the math for every company in that position.

At the time of writing (about 5:00 p.m. ET, Nov 9, 2026), Muse Ember is live on the Meta Model API and Muse Code. This article is for informational purposes only and is not financial advice.

Sources:

Thanks for reading another Methodical Minute. To read more just like this, please subscribe below and share with someone you know.

This article may be AI-assisted and is subject to errors and omissions.

Methodical Millions https://methodicalmillions.com
Next

Bitcoin ETFs Flip Negative for 2026 After a $1.28 Billion Outflow Week as BTC Lags the Stock Rebound