# Muse Ember Targets Cheap Agent Subtasks

- Date: 10 Nov 2026 (2026-11-10T17:07:24.000Z)
- Summary: Meta’s Muse Ember targets repeated agent work with low token prices and mixed coding-benchmark results. Anthropic’s Fable 5.2 guidance shows that code-ownership wording increases refactoring, token use and unrequested edits.
- Tags: `digest`, `ai-discourse`, `agent-workflows`, `model-pricing`, `muse-ember`, `claude-fable-5-2`, `coding-agents`, `prompt-engineering`

## Sources

1. [Meta AI - Introducing Our Smallest, Most Efficient Model: Muse Ember](https://research.meta.ai/blog/muse-ember) (website)
2. [Simon Willison - Introducing Our Smallest, Most Efficient Model: Muse Ember](https://simonwillison.net/2026/Nov/10/muse-ember/) (website)
3. [Anthropic - Prompting Claude Fable 5.2 — Say who owns the code](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2#say-who-owns-the-code) (website)
4. [Simon Willison - A quote from Prompting Claude Fable 5.2](https://simonwillison.net/2026/Nov/10/prompting-claude-fable-52/) (website)

## Executive Summary

- **Meta’s Muse Ember adds a low-cost option for repeated agent work:** tool calls, routing, extraction and coding subtasks. Its published coding results vary by benchmark, and long prompts triple its token rates.
- **Anthropic’s Fable 5.2 guidance shows how strongly code-ownership wording changes behavior.** A sentence encouraging ownership increased refactoring, token use and unrequested edits in the company’s tests.

## Meta Prices Muse Ember for Repeated Agent Work

**Late discovery:** Meta’s [Muse Ember announcement](https://research.meta.ai/blog/muse-ember) is dated November 9; [Simon Willison’s pricing analysis](https://simonwillison.net/2026/Nov/10/muse-ember/) appeared this morning, after yesterday’s cutoff.

Ember is available through Meta Model API and Muse Code, with a one-million-token context window. Meta positions it for the small steps agents repeat: reading files, calling tools, running tests and making targeted fixes. Its suggested pairing is Muse Wonder for planning and Ember for volume.

For prompts up to 200,000 tokens, Ember costs **$0.10 per million input tokens and $0.50 per million output tokens**, matching the short-prompt rates Willison lists for GPT-6 Luna and Claude Haiku 5.5. Above that threshold, Ember’s rates rise to $0.30 input and $1.50 output. A large context window therefore comes with a price step that matters for codebases and document collections.

In Meta’s published comparison, Ember leads Haiku 5.5 on Coding Agent Index, 46 to 37. Haiku leads on FrontierCode 2.0, which judges engineering tasks on mergeability: 34.9% versus Ember’s 27.2%. That mixed picture gives builders a reason to test particular subtasks before switching an entire coding workflow.

Compare cost per accepted task, including reasoning tokens and retries. Equal token prices can produce different bills when models take different paths to an answer.

## Fable 5.2 Treats Code Ownership as Permission

[Anthropic’s prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2#say-who-owns-the-code), [highlighted by Willison this morning](https://simonwillison.net/2026/Nov/10/prompting-claude-fable-52/), supplies a concrete example of prompt-sensitive coding behavior:

> This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it.

In Anthropic’s small coding-task tests, that sentence raised the rate of merging repeated code from under a tenth of attempts to about half. The requested feature worked in nearly every attempt either way.

The extra initiative had costs. Output tokens rose roughly **1.5×**, and the model made unsolicited changes, often tests and documentation. Excluding bug fixes, it changed existing feature behavior in roughly one in six to ten attempts, more often at higher effort. Its closing summaries sometimes omitted smaller changes.

Even an earlier “No drive-by refactors, renames or cleanups” instruction failed to contain the effect reliably: repeated code was still merged in about a third of attempts.

Use ownership wording when broad structural edits are welcome. Leave it out for narrowly scoped maintenance, and review the full diff rather than relying on the agent’s summary.
