Meta's Muse Ember Prices Cached Agent Input at $0.004

By Nathan Brooks

Meta released Muse Ember on November 9 at $0.004 per million cached input tokens, one twenty-fifth of the model's $0.10 fresh-input rate. Meta's Superintelligence Lab calls Ember its smallest and most efficient model to date. Meta designed it for high-volume, repetitive work: tool calls, coding subtasks, routing, extraction and long-document parsing.

Meta describes Ember as the fourth Muse-family model, following Spark in April, Glimmer in August and frontier model Wonder in October.

Why does the cached-input rate matter?

Meta frames the product as the execution layer in a two-model agent stack. The company says: "Muse Wonder plans, Muse Ember handles the volume." Meta prices Wonder at $4 per million input tokens and $20 per million output tokens; the model decides what to do once, while Ember carries out the resulting steps thousands of times.

Meta's developer materials note that agent frameworks typically resend the same system prompt, tool schemas and reference material on every call. Caching that prefix can dramatically reduce the marginal cost of a call. Meta did not disclose cache-write fees or expiry rules, so actual savings still depend on unpublished implementation details.

Meta's developer pricing table sets a new tier above 200,000 tokens. Prompts longer than that cost $0.30 for input, $0.012 for cached input and $1.50 for output, exactly three times the headline rates. Meta's launch post quoted only the lower tier, leaving the catch outside its main announcement.

What do the benchmarks show?

Meta published 10 benchmark rows comparing Ember at max reasoning with Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna. The results mix independent Artificial Analysis runs with Meta's own evaluations.

On the Artificial Analysis Intelligence Index, Ember and Haiku both score 36, with Haiku slightly higher before rounding; Luna scores 28. Artificial Analysis estimates Ember's cost per task at $0.14, against $0.41 for Haiku and $0.15 for Luna.

Across the index run, Ember produced 115 million output tokens, versus Haiku's 298 million, at the same $0.50 output rate. Artificial Analysis attributes much of the cost gap to that output-volume difference. Meta's published comparisons against Haiku put Ember ahead on five benchmarks:

• AutomationBench-AA, covering business workflows: Ember 37.8%; Haiku 13.9%.
• Coding Agent Index-AA: Ember 46; Haiku 37.
• Terminal-Bench 4.0: Ember 40.9%; Haiku 39.2%.
• OSWorld 2.1 computer use: Ember 69.5% partial credit; Haiku 68.9%.
• PerceptionBench with tools: Ember 80.6%; Haiku 78.6%.
• GDP.pdf long-document parsing: Ember and Luna 22.8% each; Haiku 20.8%.

Haiku leads FrontierCode 2.0-AA, 34.9% to Ember's 27.2% on mergeable engineering tasks. Meta's report identifies that benchmark as the closest match to a production pull-request evaluation.

Comparability is limited. In Terminal-Bench 4.0, Ember's score came from an Artificial Analysis run in the mini-swe-agent harness, while Haiku's came from Anthropic's self-reported Claude Code result. Four of the 10 rows are Meta's own; Meta calls its competitor testing best-effort and warns that prompts and tooling may not be optimally tuned for proprietary models.

What can developers expect from Ember?

Meta says Ember natively processes text, images, audio and video. That lets agents handle screenshots, charts, scanned documents, voice memos, meeting recordings and video without a separate model. Artificial Analysis lists only text and image inputs, so the audio and video capability remains Meta's claim.

Meta made Ember available through the Meta Model API and Muse Code, its terminal coding agent, under the ID muse-ember-1.0. It is not in the consumer Meta AI app or the Muse Agent product. Meta lists a 1 million-token context window and 512,000-token maximum output.

Meta says Ember was evaluated under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks, and remained below high-risk thresholds. Agent-behavior evaluations covered respect for user intent, sensitive-information handling, resistance to hidden instructions and seeking human input before high-impact actions.

What do early users report?

Meta's announcement included five partner quotations. Cloudflare CTO Dane Knecht said Ember ran "up to 30% faster at the tail" and handled multi-repo investigations well. You.com inference lead Chak Pothina said the company measured 95th-percentile coding latency of 2.2 seconds, with 36% to 48% fewer output tokens than other tested models.

Secondary reports, rather than Meta's quote blocks, provided the names. One announcement quotation referenced the Box AI Complex Work Eval without identifying its speaker. Two further quotations carried no company or speaker identification in the available source material.

How large is the agent market?

McKinsey's 2025 State of AI survey found that 23% of organizations were scaling agentic systems and 39% were experimenting, roughly 62% combined. In technology, 24% reported using agents in software engineering and 22% in IT. Meta's volume pricing targets that existing base, according to its launch positioning.

What remains unverified?

Four material questions remain unanswered:

• No independent third party has yet replicated Meta's internal benchmark runs.
• Meta's launch materials do not specify cache-write fees, time-to-live or regional availability for the Meta Model API.
• No public test has established long-context recall quality across the full 1 million-token window.
• Meta says Ember was trained using Muse Wonder Full and describes this as distillation. The claim has not been independently verified.

What does this change for developers?

Ember gives agent teams a cheaper candidate for the execution layer, but its long-context behavior and complete cost remain unverified. Test it on a high-volume, easy-to-score task such as routing or field extraction against the current model. Measure accuracy, latency and cost per 1,000 calls, including retries.

For Meta's thesis to work, workloads must benefit from caching and avoid the premium tier above 200,000 tokens. Meta's execution-layer case rests on the complete bill, not the sticker price. If independent tests reproduce those per-task savings at scale, Ember could become the default low-cost worker inside Meta's two-model agent stack.