# Meta releases Muse Ember 1.0, smallest model targeting agent tool calls and coding

> **Key takeaway:** Meta AI's Muse Ember 1.0 enters the efficient-model category with six configurable reasoning levels and competitive per-task pricing, though long-context costs exceed rivals GPT-6 Luna and Claude Haiku 5.5.

**Published:** 2026-11-10T07:28:21+00:00 
**Source:** Simon Willison Web & Open Source 
**Category:** software-engineering 
**Canonical URL:** https://fosswire.org/news/meta-releases-muse-ember-10-smallest-model-targeting-agent-t.html 

## Summary
Meta AI has launched Muse Ember 1.0, its entry in the lightweight large language model segment alongside OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5. The model is positioned for agent workloads including tool calls, coding subtasks, routing, extraction, and long document parsing, with pricing at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 200,000 tokens. Beyond that threshold, costs rise to $0.30/$0.012/$1.50, placing it between Luna's 272,000-token threshold and Haiku's 100,000-token cutoff in terms of long-context economics.

The model offers six reasoning levels from minimal through max, allowing developers to trade computation depth against latency and cost. Independent benchmarking by Artificial Analysis places Muse Ember on the Pareto efficiency frontier for intelligence versus cost per task, suggesting competitive practical economics despite token-level pricing that appears less favorable than Luna for longer contexts. Meta explicitly targets agentic repetition rather than frontier reasoning, distinguishing Ember from its larger Muse Wonder sibling priced 40 times higher.

## Pricing architecture and reasoning configuration
- **Tiered pricing structure:** Muse Ember 1.0 uses a two-tier input pricing model with a 200,000-token inflection point, below which it matches GPT-6 Luna and Claude Haiku 5.5 at $0.10/M input and $0.50/M output, but above which it costs $0.30/M input and $1.50/M output.
- **Cached input economics:** Cached input pricing at $0.004/M for short prompts and $0.012/M for long prompts undercuts both Luna ($0.01/$0.02) and Haiku ($0.01/$0.05), potentially benefiting applications with repeated context windows.
- **Six reasoning levels:** The model exposes minimal, low, medium, high, xhigh, and max reasoning settings, providing explicit control over inference-time compute allocation rather than fixed behavior.
- **Context window:** Supports up to one million input tokens, matching Luna and Haiku, though practical costs at that scale differ substantially across the three models.
- **Agent-targeted design:** Meta explicitly optimizes for tool calls, coding subtasks, routing, extraction, and long document parsing rather than open-ended creative or scientific reasoning.

## Why it matters
For developers building open-source agent frameworks and self-hosted automation pipelines, Muse Ember's configurable reasoning levels and competitive cached-input pricing offer a new optimization target for cost-sensitive workloads. The model's placement on the Artificial Analysis Pareto frontier suggests it may displace Luna and Haiku in tool-heavy applications where per-task cost matters more than per-token pricing, though the 200,000-token threshold penalizes workflows requiring longer context without caching.