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 has introduced Muse Ember, a model designed for repeated agent tasks such as tool use, coding subtasks, routing and document parsing. It is available through the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens; Meta’s performance and cost comparisons rely on published evaluations and estimated task costs.
Meta has introduced Muse Ember, a model it describes as its smallest and most efficient, aimed at tasks that AI agents repeat frequently, including tool calls, coding subtasks and long-document parsing. The company says the model is available today through the Meta Model API, with listed prices of $0.10 per million input tokens and $0.50 per million output tokens.
Meta says Muse Ember is designed for complex reasoning execution and structured, repeated work, rather than only one-off conversations. It highlights agent workflows that gather records, call tools in sequence and update information across applications. The model has a 1 million-token context window and can produce up to 512,000 output tokens, according to the company.
Meta’s launch report compares Muse Ember with Claude Haiku 5.5 and GPT-6 Luna using Artificial Analysis results dated November 9, 2026. At the listed reasoning settings, the report gives Muse Ember task-cost estimates of $0.116 for a score of 33.9 and $0.136 for a score of 35.8. It describes those results as more intelligence per dollar than either comparison model. The costs are estimates based on proposed pricing, and the scores reflect the named benchmark rather than every possible use.
In the report’s maximum-reasoning results, Muse Ember scored 37.8% on AutomationBench-AA, compared with 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5. Meta also reports 61.6% on Agents’ Last Exam, against 57.1% and 57.0%, respectively. The company lists results across coding, computer-use and document-related evaluations, but says the tests use different published or internal evaluation setups.
Lower-Cost Models for Agent Work
The launch targets a cost pressure that can grow when an agent makes many model calls for a single workflow. A model used for repeated, bounded steps—such as extracting information, checking a clause or applying a small code fix—may be chosen on speed and per-task cost as well as raw capability. Meta is positioning Ember as a way to handle that volume without sending every step to a more expensive, larger model.
If the reported results hold up in independent use, the combination of low listed token prices and agent-focused benchmark scores could make the model relevant to businesses building automated workflows. But benchmark performance is not a guarantee of accuracy in a particular company’s systems, and the total cost of a workflow also depends on how many tokens it uses, how often it retries and what tools or human review it requires.
The announcement also adds a new competitor to a market where providers are differentiating models by task and operating cost, not only by broad reasoning scores. For buyers, the practical question is whether Ember performs reliably on their own data and integrations at the stated price. Meta’s figures provide an initial comparison, not a substitute for that testing.
Products Worth Considering
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Meta’s Agent-Focused Model Strategy
Meta’s announcement describes agents as systems that perform work through a sequence of model outputs and external actions. In that design, a model may be called repeatedly to interpret a request, choose a tool, process a result and decide what to do next. The company says those small, structured steps account for much of an agent’s token use, which is why it built Ember for repeated execution.
The launch report includes results from third-party benchmark material attributed to Artificial Analysis, alongside internal Meta evaluations and other listed tests. Those categories should not be treated as interchangeable: benchmark scores depend on their task definitions and evaluation methods. The report identifies some results as internal and notes, for example, that Agents’ Last Exam uses a Linux subset. Meta also says its model comparisons use each model at specified reasoning settings, while the cost chart is an estimate.
Meta cites early partner testing as an example of intended enterprise use. Box said it tested Ember on contract review and candidate screening against detailed criteria. That is a partner account of early evaluation, not a published independent assessment of performance across customers or deployments.
“Muse Ember is our smallest and most efficient model.”
— Meta
Independent Results and Deployment Costs
The announcement does not establish how Muse Ember performs across independent evaluations or production deployments. Meta’s report provides benchmark results, but some are internal, and the material supplied does not include enough detail to independently verify every comparison, harness or scoring decision. Performance may differ on a customer’s own tasks, data and tools.
It is also not clear from the announcement what the estimated per-task costs include beyond the stated model pricing, or how they change with longer prompts, output length, retries and tool use. The company’s claim that Ember offers more intelligence per dollar than the named competitors is based on its reported evaluations and cost estimates; it should not be read as a general finding across all workloads. The announcement does not specify broader availability terms, service limits or whether prices may change.
Customer Testing Will Test the Claims
Developers can access Muse Ember through the Meta Model API, according to the launch announcement. The next useful evidence will come from customer testing and independent comparisons that report task performance alongside real usage costs, including input and output tokens and any repeated calls.
Meta’s report points to additional details in its evaluation material, but the supplied announcement does not give a date for further results or identify a next product milestone. Prospective users will need to assess the model against their own agent workflows, including accuracy, failure handling, latency and the cost of human review.
Key Questions
What is Muse Ember?
Muse Ember is a model from Meta designed for repeated agent tasks such as tool calls, coding subtasks, routing, extraction and long-document parsing.
How much does Muse Ember cost?
Meta lists a price of $0.10 per million input tokens and $0.50 per million output tokens through the Meta Model API. Its report’s per-task cost figures are estimates, not fixed prices for every workload.
How large is Muse Ember’s context window?
Meta says the model has a 1 million-token context window and supports up to 512,000 output tokens.
Has Muse Ember’s performance been independently confirmed?
The announcement reports results from Artificial Analysis and other evaluations, including some internal tests. The information provided does not establish independent confirmation across all reported benchmarks or performance in customer deployments.
Source: research.meta.ai
