TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta introduced Muse Ember, a model it describes as its smallest and most efficient, aimed at repeated agent work such as tool calls, coding subtasks and document processing. Meta says it is available now through the Meta Model API, with listed prices of $0.10 per million input tokens and $0.50 per million output tokens; its performance comparisons are based on evaluations reported by Meta and Artificial Analysis.
Meta has introduced Muse Ember, a model it describes as its smallest and most efficient, designed for repeated agent tasks such as tool calls, coding subtasks, routing and parsing long documents. The model is available through the Meta Model API at listed prices of $0.10 per million input tokens and $0.50 per million output tokens, Meta said.
Meta’s announcement positions Ember for workloads made up of many smaller, structured steps rather than a single long response. The company says it can execute business workflows by retrieving records, calling tools in sequence and updating systems across apps. It lists a 1 million-token context window and support for up to 512,000 output tokens, which Meta says can allow a full codebase or document set to be handled in one call.
Meta’s published comparison reports strong results against Claude Haiku 5.5 and GPT-6 Luna on selected agent and coding evaluations. At maximum reasoning, the company reports Ember scores of 37.8% on AutomationBench-AA and 61.6% on Agents’ Last Exam. It reports 13.9% and 57.0% for Haiku, and 20.2% and 57.1% for Luna, respectively. Meta labels the AutomationBench-AA result strict pass@1 and describes its Agents’ Last Exam figure as covering the Linux subset.
The report also presents Artificial Analysis comparisons of intelligence scores against estimated cost per task. Its table gives Ember scores of 33.9 at an estimated $0.116 per task and 35.8 at $0.136. Meta says Ember matches Haiku’s best score at about a third of the cost and exceeds every Luna setting in that comparison. Those are company-reported comparisons; the report says Ember cost figures are estimates at proposed pricing, while API token prices are listed separately.
Lower-Cost Steps for AI Agents
For organizations running agents at scale, repeated calls can make the cost of each individual model response matter. Meta is pitching Ember as a way to handle frequent, bounded tasks—such as extracting information, checking code or calling tools—without using a larger model for every step. If its reported performance and cost hold in customers’ own workloads, teams could assign routine agent actions to a lower-priced model and reserve more capable systems for harder work.
The announcement also reflects competition over the economics of agent systems, not only peak benchmark scores. A model that performs well on workflows involving tools, software and long documents could be useful in business products where reliability across multiple steps matters. But the evaluations do not establish how Ember will perform across every real-world deployment, and the published cost-per-task estimates should not be treated as a guaranteed bill for a particular customer.
Products Worth Considering
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Raspberry
As an affiliate, we earn on qualifying purchases.
How Meta Evaluated Ember
Meta’s report compares Ember with two named models across a set of agent, coding, multimodal and general evaluations. The supplied results include Terminal-Bench 4.0, where Meta reports Ember at 40.9%, Haiku at 39.2% and Luna at 16.4%, with each model tested in its best published harness. On the Coding Agent Index-AA, the report lists scores of 46, 37 and 31, respectively. Different tests measure different capabilities, so these scores should be read within each benchmark rather than as a single universal measure of model quality.
The report dates the Artificial Analysis cost-and-intelligence comparison to November 9, 2026. Meta also cites early partner testing: Box said Ember performed on contract review and candidate screening in its own complex-work evaluation. That is a partner account rather than an independent finding. Meta’s materials describe internal evaluations for some measures, including the Personal Superintelligence Index, which the company characterizes as a composite index of personal and knowledge work.
““Our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day.””
— Meta
Independent Performance Still Unclear
The announcement does not establish how Ember performs across a broad range of independent customer deployments. The supplied material includes a mixture of Artificial Analysis results, Meta internal evaluations, benchmark results and partner testing, but it does not provide enough detail here to verify every testing condition or establish that results are directly comparable across all tasks. Benchmark scores can depend on the harness, reasoning setting and scoring method.
Meta’s cost-per-task figures are estimates, and the report says those calculations use proposed pricing. The relationship between those estimates and a customer’s actual spending will depend on usage, task design and the number of input and output tokens. The announcement also does not provide details about the model’s training, availability outside the Meta Model API, or a release schedule for additional versions.
Testing in Customer Workflows
Meta says Muse Ember is available now through the Meta Model API. The next practical test will be how it performs when developers and businesses use it for their own agent workflows, including whether it can complete multi-step tasks reliably at the advertised token prices. Meta’s report points to early partner use, but it does not give a wider customer rollout timetable or promise particular results for individual deployments.
Further evidence may come from independent evaluations, customer reports and more detail about the benchmarks and pricing estimates. Until then, Meta’s reported comparisons offer an initial picture of the model’s intended strengths, rather than a complete account of its performance in production.
Key Questions
What is Muse Ember?
Muse Ember is a Meta model designed for repeated agent tasks, including tool calls, coding subtasks, routing, extraction and long-document processing.
How much does Muse Ember cost?
Meta lists API pricing of $0.10 per million input tokens and $0.50 per million output tokens. Its separate cost-per-task comparison uses estimated costs and should not be confused with a guaranteed customer bill.
Where can developers access it?
Meta says Muse Ember is available now through the Meta Model API. The announcement does not specify additional access routes.
How did it compare with Claude Haiku 5.5 and GPT-6 Luna?
Meta reports that Ember scored higher than both models on several named evaluations, including AutomationBench-AA and Agents’ Last Exam at maximum reasoning. Results vary by benchmark, and the reported comparisons do not independently establish performance on every real-world task.
What remains uncertain about the model?
Independent results across customer workloads, the full conditions behind each comparison, and how estimated task costs translate into actual deployment spending remain unclear from the announcement.
Source: research.meta.ai
