Meta launches Muse Ember: a cheap agent model it says matches Haiku at a third of the price
Meta's smallest model is built for agents' routine work and costs $0.10 per million input tokens. Meta claims it delivers more intelligence per dollar than Claude Haiku 5.5 and GPT-6 Luna.


Meta is launching Muse Ember, its smallest and most efficient language model. The model is built for the small, repeated steps that AI agents spend most of their time on: tool calls, coding subtasks, routing, data extraction and reading long documents. It costs $0.10 per million input tokens and $0.50 per million output tokens, and Meta claims it delivers more intelligence per dollar than Claude Haiku 5.5 and GPT-6 Luna.
Meta announced the model on Monday 9 November 2026 on its research blog. The official @AIatMeta account posted the announcement on X at 19:04 Oslo time.
A small model for the agents' groundwork
Muse Ember is not meant to solve the hardest problems on its own. Meta describes it as an execution model that works alongside its larger Muse Spark and Muse Wonder models. The large model plans, and Muse Ember handles the volume. According to the developer documentation, Muse Ember accepts the same request body as Muse Spark, so switching is in most cases a matter of changing the model ID.
The model has a context window of 1 million tokens and can write up to 512,000 tokens in a single response. It understands text, images, audio and video, and Meta says it can read screenshots, charts and scanned documents, work from meeting recordings and operate a computer from screenshots.
Meta also writes that Muse Ember is the first model the company has trained using its large Muse Wonder Full model.
Price and performance according to Meta
The price applies to requests with up to 200,000 input tokens. Above that limit, the model costs $0.30 in and $1.50 out per million tokens. For comparison, Meta lists its own Muse Spark 1.3 at $1.25 and $4.25.
On the Artificial Analysis Intelligence Index, Muse Ember scores 36 points, the same as Claude Haiku 5.5 at its highest effort setting, and above GPT-6 Luna at 28. Meta says Muse Ember reaches this at roughly a third of the cost per task, about $0.14. Muse Spark 1.3 scores 37 but costs $2.39 per task. Meta itself notes that the Muse Ember cost figures are estimates based on proposed pricing.
On AutomationBench, a test of business workflows, Meta reports 37.8 percent for Muse Ember against 20.2 for GPT-6 Luna and 13.9 for Claude Haiku 5.5. On the Terminal-Bench 4.0 coding test the gap is small: 40.9 against 39.2 percent for Haiku. On FrontierCode 2.0, which measures whether code changes are good enough to merge, Haiku leads with 34.9 against 27.2 percent. Several of the measurements are Meta's own internal tests and have not been independently verified.
Early users and safety
Box, Cline, Oracle, Cloudflare and You.com tested the model before launch. Cloudflare says Muse Ember was up to 30 percent faster on its slowest requests than the model it compared against. Box describes the model scoring contracts clause by clause against a company's negotiation playbook.
Meta writes that the model was assessed under its own safety framework for chemical and biological risk, cybersecurity and loss of control, and that it stays below the framework's high-risk thresholds. The company also tested whether the model can be misled by hidden instructions in web pages, documents and tool outputs, and whether it asks a human before high-impact actions.
Muse Ember is available from today as muse-ember-1.0 through the Meta Model API and in the Muse Code coding tool, on all subscriptions. The announcement says nothing about where data is processed or whether European data processing is guaranteed.
What it means for leaders
- Agent costs can drop sharply. Organisations running many agent steps can save a lot by sending routine work to a cheap model and keeping the expensive models for planning. Ask teams to calculate cost per task, not just per token.
- Test before you switch. Meta's numbers show the model is strong on workflows but weaker than Haiku on code changes meant for merging. Run your own tests on real tasks.
- Clarify data and jurisdiction. The Meta Model API is a new supplier channel for many. Check terms, data retention and where data is processed before the model gets access to contracts, customer data or internal systems.
Sources and media
- Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember, 9 November 2026
- Meta for Developers: Start building with Muse Ember, 9 November 2026
- Meta: Muse Ember documentation
- AI at Meta on X, 9 November 2026
Illustration: AI-generated with Grok Bot / hogby.ai