---
type: "post"
id: "meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e"
slug: "meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e"
title: "Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls"
description: "Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls, coding subtasks, routing and long-document parsing."
retrieval_nugget: "Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls, coding subtasks, routing and long-document parsing. Its published token prices rise above a 200,000-token prompt threshold, while Artificial Analysis comparisons cited by Simon Willison put its task cost at about $0.12 in xhigh and $0.14 in max."
published_at: "2026-11-10"
updated_at: "2026-11-10"
record_date: "2026-11-10"
date_kind: "published_at"
topics: ["claude"]
entities: ["Meta","Muse Ember","Artificial Analysis","Claude Haiku 5.5","GPT-6 Luna","OpenCode"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_angle"
what_changed: "The launch is framed as a detailed price comparison with GPT-6 Luna and Claude Haiku 5.5, highlighting how long-prompt thresholds and reasoning-token usage complicate per-token comparisons."
prior_refs: ["https://newruntime.com/posts/meta-released-muse-ember-its-smallest-reasoning-model-98fb00/"]
source_url: "https://simonwillison.net/2026/Nov/10/muse-ember"
source_urls: ["https://simonwillison.net/2026/Nov/10/muse-ember"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e/","markdown":"https://newruntime.com/posts/meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e.md","json":"https://newruntime.com/posts/meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e.json"}
markdown_url: "https://newruntime.com/posts/meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e.md"
---

# Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls

## Retrieval answer

Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls, coding subtasks, routing and long-document parsing. Its published token prices rise above a 200,000-token prompt threshold, while Artificial Analysis comparisons cited by Simon Willison put its task cost at about $0.12 in xhigh and $0.14 in max.

## Summary
Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls, coding subtasks, routing and long-document parsing. Its published token prices rise above a 200,000-token prompt threshold, while Artificial Analysis comparisons cited by Simon Willison put its task cost at about $0.12 in xhigh and $0.14 in max.

## What happened
Meta describes Muse Ember as its smallest and most efficient model for complex reasoning execution and recurring agent tasks, including tool calls, coding subtasks, routing, extraction and long-document parsing. Simon Willison’s account links to Meta’s announcement and reproduces its claim that Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task. These are Meta’s characterization and a benchmark-based comparison, not a guarantee for every workload.

The pricing schedule reported for Muse Ember 1.0 depends on prompt length. At up to 200,000 input tokens, the listed rate is $0.10 per million input tokens and $0.50 per million output tokens; cached input is $0.004 per million. Above 200,000 input tokens, input rises to $0.30 per million, cached input to $0.012, and output to $1.50 per million. The page compares these thresholds with GPT-6 Luna, whose input rate rises above 272,000 tokens, and Claude Haiku 5.5, whose rate rises above 100,000. All three are described as supporting up to one million input tokens.

The Artificial Analysis chart reproduced in the account estimates cost per task at about $0.12 for Ember with xhigh reasoning and $0.14 with max. The comparison plots these costs against Intelligence Index scores: Ember’s xhigh point is approximately 34 and max approximately 36. The page also shows Claude Haiku 5.5 at about $0.044 in low and $0.25 in xhigh, and GPT-6 Luna at about $0.07 in non-reasoning and $0.10 in xhigh. These figures are chart readings, not a universal production bill; the account explicitly warns that token price alone is not enough because reasoning models can use different numbers of tokens on the same problem.

Ember offers six reasoning levels: minimal, low, medium, high, xhigh and max. In a small SVG-generation test described by Willison, the high setting did not return a complete SVG, and the other outputs did not show a clean progression in visual quality with reasoning level. That test is an illustration, not a systematic evaluation. The linked account does not establish deployment availability, service limits beyond the context and price tiers, or performance across production tasks.

## What the archive adds
New Runtime's earlier coverage on [November 9](https://newruntime.com/posts/meta-launched-muse-ember-its-smallest-model-built-4baf24/) introduced Ember and summarized Meta's efficiency claims. A [November 10 note](https://newruntime.com/posts/meta-released-muse-ember-its-smallest-reasoning-model-98fb00/) reported Artificial Analysis rankings and average evaluation cost, while [another November 10 note](https://newruntime.com/posts/meta-muse-ember-is-listed-as-available-in-3010fc/) covered a provider listing and its context limit. This is a new pricing and task-cost detail on a known story.

## Trend signal
This supports [agent economics moving to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/): the useful comparison is cost per task at a stated reasoning setting, not just cost per token. The evidence is a vendor release alongside a benchmark chart, so it does not establish what teams will pay for verified production outcomes.

## Why now
The release arrives as model comparisons increasingly account for reasoning-token usage rather than treating token rates as a direct proxy for workload cost. The cited chart makes that distinction concrete by comparing estimated cost per task across reasoning settings.

## Connections to other events
New Runtime's earlier coverage also noted that Muse Ember was listed through the OpenCode provider, alongside a listed context limit and token prices. That makes provider availability a separate part of the same adoption question: benchmark economics are only useful to a team if the model is accessible in its chosen stack.

## Why it matters
For teams building agents, a cheaper token is not necessarily a cheaper completed job. Reasoning settings can change both output quality and token consumption, while long prompts can cross Ember's 200,000-token pricing threshold and materially raise the bill. The task-cost figures offer a more relevant starting point than headline token rates, but they remain benchmark estimates rather than a promise about a team's own workload.

A useful evaluation should replay representative tasks, record the prompt size and reasoning level, and score successful completion against human review or an existing acceptance test. Compare total cost and failure rate with the model already in production. Include cached-input behavior where applicable, and test prompts on both sides of the threshold. This turns the release's efficiency claim into a decision grounded in the team's actual agent loop.

## Limits and uncertainty
The cost and intelligence figures are estimates shown in Artificial Analysis comparisons reproduced by Willison, not independently verified production results for a particular team. The small SVG test does not measure general coding, extraction or tool-use reliability, and its result cautions against assuming higher reasoning settings yield uniformly better outputs. The announcement and comparison do not establish latency, throughput, service availability for every deployment, or total costs once retries, tool calls and verification are included.

## Skill to build now
Build a small cost-per-completion harness for one agent task your team runs repeatedly. Replay the same inputs at Ember's reasoning levels and a current baseline; log prompt length, cached tokens, output tokens, retries, successful completion and reviewer corrections. Include examples just below and above 200,000 input tokens so the pricing step is visible. A stable test set makes the published task-cost comparison useful without mistaking it for your own result.

## Reader takeaway
Muse Ember's relevant claim is not simply a low token price: it combines agent-focused positioning with task-cost comparisons at different reasoning levels. For a real workload, the answer depends on prompt length, reasoning setting and whether the agent finishes correctly. Measure the cost of accepted completions before moving a production route.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)