Meta introduced Muse Ember on November 9, aiming to lower the expense of repeated tasks inside AI agents, including calling software tools, extracting information and completing small coding jobs. The model is available through Muse Code and the Meta Model API.
Meta lists starting prices of $0.10 per million input tokens and $0.50 per million output tokens. Those rates apply to prompts of up to 200,000 tokens. Longer prompts cost $0.30 for input and $1.50 for output per million tokens, an important qualification for customers processing large collections of documents.
A smaller model with a specific role
The company positions Ember alongside its larger Muse models rather than as a replacement for all of them. It describes a division of labor in which Muse Wonder handles difficult planning while Ember executes repeated work. An AI agent is software that uses a model to choose and carry out steps, often by operating other tools.
The commercial question is not simply whether a model can answer a difficult question. In a workflow that calls a model many times, the cost and reliability of each small operation can affect the total bill. Retries, verification and failed actions need to be included in any comparison.
Benchmark comparisons need their methodology
Meta’s evaluation report describes a mixture of external results and internal tests. For controlled internal comparisons, it says models use a shared test framework. External coding results retain the software framework used by the reporting provider, so not every score represents an identical application setup.
The report also cautions that internally generated competitor results are best-effort comparisons. Prompts, tools and runtime settings may not be tuned optimally for proprietary rivals. That caveat limits how confidently a customer can translate a leaderboard position into an expected advantage in its own software.
For Agents’ Last Exam, Meta used the Linux portion of the evaluation and excluded five tasks because of storage limits or missing scores. Its report says the official leaderboard compares complete agent products, whereas its own runs used a common internal agent loop. Those are related but different measurements.
Cloud access has its own restrictions
Google Cloud’s documentation lists Muse Ember 1.0 as a preview release dated November 9. It supports text and image inputs and produces text, with a maximum input allowance of 1,048,576 tokens and output allowance of 524,288 tokens. Customers should use those service specifications when building on Google Cloud.
Google lists reasoning, function calling and structured output as supported preview features. Batch predictions and provisioned throughput are not supported. Preview status also means the offering may have limited support, a material consideration for businesses planning production services.
The launch gives developers another option for distributing work between models of different sizes. The next practical test is whether Ember completes a customer’s repetitive tasks at lower total cost without adding unacceptable errors. Published prices and benchmark reports provide a starting point, but they do not establish that outcome for an individual deployment.
