Meta aims to reduce the costs of operating AI agents. On November 9, the company released Muse Ember, a lightweight model that would perform the repetitive tasks that keep AI agents occupied. Unlike others, Ember does not aim at solving complicated tasks, but rather focuses on doing regular work at the lowest possible cost.Built to handle volume while Muse Wonder plansMeta built Ember to handle the routine work that keeps AI agents running. It can call other software tools, direct requests to the right system, extract information from documents, and process lengthy files. Ember can also handle smaller coding jobs, such as running tests and making targeted fixes.In order to make more complex decisions, Meta uses a combination of Muse Wonder, which is a superior reasoning model, and Ember. Wonder decides how to tackle a task, while Ember executes the smaller actions necessary to complete the task. In the words of Meta:“Muse Wonder plans, Muse Ember handles the volume.”Ember is the fourth iteration in the Muse lineup, following Muse Spark, Muse Glimmer, and Muse Wonder. It is accessible to developers through the Meta Model API and Muse Code. Under its pricing plan, Meta charges $0.1 per million input tokens, as well as $0.004 per million cached input tokens and $0.5 per million output tokens.Ember comes with a one-million-token context window that allows it to process larger documents and codebases. It is also capable of handling images, audio, and videos.What Artificial Analysis found when it tested the claimsIndependent tests clearly demonstrate the extent to which Ember’s cost advantage exists. According to Artificial Analysis, Ember’s Intelligence Index score of 36 is the same as that of Claude Haiku 5.5. However, while Ember costs only $0.14 per benchmark task, Claude Haiku costs $0.41. The figures quoted are estimated costs for the tests, not fixed charges paid by customers. Additionally, Ember scored 68.1% on AutomationBench-AA, putting it on par with Meta’s more capable model, Muse Wonder.However, its coding test results were inconsistent. Ember had a better performance than Haiku in Terminal-Bench 4.0 with scores of 40.9% and 32.8%, respectively. But Haiku performed better than Ember when it came to FrontierCode 2.0, with scores of 34.9% and 27.2%.Another area of weakness of Ember is general knowledge. The model scored only 21.1% accuracy in AA-Omniscience and skipped answering 64% of queries. Thanks to its careful stance, Ember was able to keep the hallucination score to 17.4%.Meta Muse Ember’s $0.14 Agent Task: Cost and Performance vs Claude Haiku 5.5Why cost per task is the fight that matters nowEmber’s lower pricing comes at a time when companies reconsider their spending on AI technologies. Gartner predicts that demand for AI agents and assistants will push spending to $29.2 billion in 2026 and is part of a $2.7 trillion AI market forecast. As the use of AI technologies increases, BCG encourages software companies to explore charging for completed tasks or results rather than simply for each licensed user. However, it cautions against abandoning traditional subscription pricing too quickly.The bigger question is whether cheaper AI delivers better returns. McKinsey argues that token costs alone cannot measure business value. Lower prices may also encourage companies to use more AI. IDC found that 67% of enterprises exceeded their agent budgets. The research firm expects costs per agent action to fall 95% by 2030, even as total token-delivery costs rise 487-fold.The ROI, reliability, and oversight questionsMeta’s early partners have highlighted Ember’s speed. Cloudflare CTO Dane Knecht said the model ran:“up to 30% faster”You.com inference lead Chak Pothina reported:“a p95 of 2.2 seconds”Box executive Yashodha Bhavnani said Ember could assess contracts:“against a company’s negotiation playbook clause by clause”These results are promising, but they come from partner testing. Businesses still need to assess Ember’s performance in their own operations.Reliability also goes beyond speed. The FII Institute cites survey findings showing that 73% of respondents are concerned about AI data privacy and security. Meanwhile, a McKinsey survey found that only about 30% of organizations reached higher maturity levels in agentic AI governance.Cost pressures extend beyond the models themselves. Meta’s recent $15 billion bond sale highlights the expense of building AI infrastructure. Cryptopolitan has also reported on Meta’s expanding data-center investments.If you're reading this, you’re already ahead. Stay there with our newsletter.