Infrastructure · NOVEMBER 10, 2026
Meta’s Muse Ember targets repetitive agent work at low token cost
Meta’s small model posts 37.8% strict pass@1 on AutomationBench-AA. For small teams, that supports a bounded customer-follow-up pilot—not unattended outreach.
Meta released Muse Ember through its Meta Model API on November 9, listing prices of $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Meta describes its smallest model as designed for repeated agent steps: tool calls, routing, extraction, coding subtasks and long-document parsing.
Meta reports 37.8% strict pass@1 on AutomationBench-AA at maximum reasoning effort, versus 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5. Those self-reported results concern business-workflow execution, not evidence that Ember improves customer outreach.
What the benchmark measures
Meta’s evaluation-methodology summary describes 657 tasks across simulated business applications, including Gmail, Salesforce and Slack. Workflows execute through application APIs, with deterministic checks inspecting the final system state. Strict pass@1 measures whether an entire task succeeds on its first attempt without violating guardrails.
VentureBeat reports a separate objective-level score that credits completed portions of workflows. That measure shouldn’t be confused with strict first-attempt success: completing part of a workflow isn't the same as passing the whole task.
Meta also cautions that competitor runs are best-effort. Prompts, tools and runtime may not be optimally tuned for proprietary models. The comparisons therefore carry a qualification beyond their status as vendor-reported results.
Context capacity isn't a cost estimate
Meta lists a 1M-token context window, but pricing changes before that limit. Prompts above 200,000 tokens cost $0.30 per million input tokens and $1.50 per million output tokens.
Meta’s Intelligence Index chart separately estimates maximum-effort task costs of $0.136 for Ember and $0.411 for Haiku, with respective scores of 35.8 and 36.2. These are benchmark-specific estimates, distinct from both the published token rates and the cost of producing a usable customer follow-up.
The practical implication: measure cost per accepted draft, not merely cost per token.
A bounded follow-up pilot
For an owner running a small B2B or consumer-facing team, the pricing suggests a narrow test of organizing prospective-customer details and preparing personalized follow-up drafts under human review. The 37.8% strict pass@1 result doesn’t establish improved outreach accuracy, time savings or customer conversion.
A hypothetical pilot could extract customer needs and supporting details from inquiry records, then prepare an unsent follow-up for approval. B2B teams could test prospect-account notes; consumer-facing teams could test individual service inquiries. Neither example is a documented deployment.
The owner’s evaluation should compare field accuracy, unsupported personalization, review time, total API use and retries against the existing manual process. That would test whether the agent reduces effort rather than assuming low token prices translate into useful work.
Sending should remain a separate approval step. Expansion would depend on observed results, not on treating a business-workflow benchmark as a prediction of more customers.