Meta Muse Ember: A $0.10 AI Model for Agents and Its 200K Catch

Highlights

Meta released Muse Ember on Nov. 9, 2026, through the Meta Model API and Muse Code.
It costs $0.10 per million input tokens and $0.50 per million output tokens.
Ember matches Claude Haiku 5.5's score of 36 on Artificial Analysis at about a third of the cost per task.
It trails Haiku 5.5 on the FrontierCode 2.0-AA coding test.
Requests above 200,000 input tokens are billed at triple the rate on every token.

Meta has released Muse Ember, the smallest and most cost-efficient model from Meta Superintelligence Labs. It is designed for the repetitive work AI agents run all day, with prices starting at $0.10 per million input tokens.

Ember is the fourth model in the Muse family, following Spark in April, Glimmer in August and Wonder in October. The launch shows that competition in AI is no longer only about the smartest model, but also about which one is cheapest to run at scale.

Quick answer: Muse Ember is Meta's smallest AI model, built for agent tasks such as tool calls, coding subtasks and long document parsing. It launched on Nov. 9, 2026, on the Meta Model API and Muse Code. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1 million token context window.

What Muse Ember Is Built For

Meta says Ember is built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing.

The model is designed to be paired with a larger one. Meta sums it up in one line: "Muse Wonder plans, Muse Ember handles the volume." TechCrunch gave the example of a customer support agent, where a heavier model sets the rules for an issue and a lighter model applies them across tens of thousands of cases.

Ember natively understands text, images, audio and video. Agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video without a separate model. It has a 1 million token context window and supports up to 512,000 output tokens.

Muse Ember Pricing and the 200K Token Line

The Meta Model API documentation lists two price tiers, and this is the part developers should read carefully. The tier is set by the number of input tokens in a single request, including cached tokens.

Token type	Up to 200K input tokens	Above 200K input tokens
Input	$0.10 per million	$0.30 per million
Cached input	$0.004 per million	$0.012 per million
Output	$0.50 per million	$1.50 per million

Source: Meta Model API pricing documentation. Model muse-ember-1.0, Standard mode only.

The long-context rate applies to every token in the request, not just the excess. A request with 200,000 input tokens and 20,000 output tokens costs $0.03. Raise the input to 201,000 tokens and the cost jumps to about $0.09. Ember is also not offered in Contributor mode, the discounted option that requires permission to use customer content.

For comparison, Muse Spark costs $1.25 for input and $4.25 for output per million tokens, while Muse Wonder costs $4 and $20.

How Muse Ember Scores Against Haiku 5.5 and GPT-6 Luna

Meta compared Ember directly with small models from Anthropic and OpenAI at their highest reasoning settings. Some of the tests were run by Artificial Analysis, while others were reported by Meta.

Benchmark	Muse Ember	Claude Haiku 5.5	GPT-6 Luna
Artificial Analysis Index	36	36	28
AutomationBench-AA	37.8	13.9	20.2
Agents' Last Exam	61.6	57.0	57.1
Terminal-Bench 4.0	40.9	39.2	16.4
Coding Agent Index-AA	46	37	31
FrontierCode 2.0-AA	27.2	34.9	25.2

Source: Meta announcement, Nov. 9, 2026, at maximum reasoning settings.

The clearest weakness shows up on FrontierCode 2.0-AA, which judges whether code changes are fit to merge into real projects. Haiku 5.5 leads there, 34.9 to 27.2. For teams using a small model to write production code, that number matters more than a composite score.

Cost Per Task Is the Main Selling Point

Meta emphasizes cost per task rather than token price alone. In its chart based on Artificial Analysis data, Ember at maximum reasoning scored 35.8 at an estimated $0.136 per task. Haiku 5.5 scored 36.2 at $0.411, and Luna scored 27.8 at $0.152.

The Decoder also noted that the larger Muse Spark 1.3 costs $2.39 per task for a score of 37, nearly the same as Ember. Meta's chart labels Ember's costs as estimates at proposed pricing, so actual spending will still depend on usage patterns.

Partner Feedback and Availability

Several companies tried Ember before launch. Cloudflare said it ran up to 30% faster at the tail than the comparable model it tested. You.com reported a p95 latency of 2.2 seconds in its coding tests, and Box used it to score contracts against a company's negotiation playbook. These comments come from Meta's launch materials.

Ember is available on the Meta Model API as muse-ember-1.0 and in Muse Code. The announcement does not list availability by country, so developers should check access and payment options in their Meta Model API dashboard.

FAQ (Frequently Asked Questions)

1. What is Muse Ember?

Muse Ember is Meta's smallest and most efficient AI model. It is built for the repetitive tasks that AI agents perform.

2. How much does Muse Ember cost?

It costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 200,000 input tokens. Above that line, rates triple.

3. Is Muse Ember better than Claude Haiku 5.5?

Both score 36 on the Artificial Analysis index, and Ember is cheaper per task. Haiku 5.5 still leads on the FrontierCode 2.0-AA coding test.

4. Where can I use Muse Ember?

Ember is available on the Meta Model API and in Muse Code under the model name muse-ember-1.0.

5. Can Muse Ember process audio and video?

Yes. Meta says Ember understands text, images, audio and video without a separate model.

Conclusion

Muse Ember brings aggressive pricing to high-volume agent work and puts pressure on Haiku 5.5 and GPT-6 Luna. Its composite score is attractive, but weaker coding results show it will not suit every task.

Developers who want to try it should keep requests under 200,000 input tokens and test Ember on their own workloads.