Meta Prices Muse Ember at $0.10 per Million Input Tokens, Says Nothing on Weights

Meta released Muse Ember on Nov. 9, a model it calls its smallest and most efficient, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. The launch post on Meta's research site does not give a parameter count, a licence or any statement on whether the weights can be downloaded.

Muse Ember is available as muse-ember-1.0 through Muse Code and the Meta Model API, and on Google Cloud and Oracle Cloud. It takes text, images, audio and video, has a 1 million token context window and can return up to 512,000 tokens in one response.

What Muse Ember costs

Prompts above 200,000 tokens are billed at $0.30 per million input tokens and $1.50 per million output tokens. Cached input costs $0.004 per million tokens at the lower tier and $0.012 at the higher one.

Meta said a task on the Artificial Analysis suite costs about $0.14 to run on Ember, against $2.39 on Muse Spark 1.3, its earlier model. Those costs come from Meta's post, not from a page Artificial Analysis published that The Terminal could read.

Which scores Meta ran and which an outside firm ran

The post lists 10 benchmarks, all at maximum reasoning effort, and labels who ran each. Artificial Analysis, an independent testing firm, ran five. Meta ran four internally. The tenth, Terminal-Bench 4.0, is credited to the best published harness without a named runner.

Benchmark	Run by	Ember	Haiku 5.5
Intelligence Index	Artificial Analysis	36	36
Coding Agent Index	Artificial Analysis	46	37
FrontierCode 2.0	Artificial Analysis	27.2%	34.9%
OSWorld 2.1, partial credit	Meta	69.5%	68.9%

Ember ties Anthropic's Claude Haiku 5.5 on the Intelligence Index and trails Muse Spark 1.3, which scores 37. On FrontierCode 2.0, Ember's 27.2 percent is 7.7 percentage points below Haiku's 34.9 percent, the one coding test in the post where Meta's model loses to the Anthropic one. GPT-6 Luna scores 25.2 percent on the same test.

1. Muse Ember46 points
2. Claude Haiku 5.537 points
3. GPT-6 Luna31 points

Source: Meta, Muse Ember launch post, accessed 2026-11-10 (scores attributed by Meta to Artificial Analysis)

The biggest gap is on Artificial Analysis's AutomationBench: 37.8 percent for Ember, 13.9 percent for Haiku and 20.2 percent for GPT-6 Luna. Meta-run tests, such as Personal Superintelligence Index (46.8 percent against Haiku's 36.1 percent), are vendor-supplied and should be read that way. Meta also said the model stayed below the high-risk thresholds in its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks.

Why the weights question matters

Artificial Analysis wrote in April that Muse Spark was Meta's first frontier model not released as open weights, and that it scored 52 on the index version then in use. That version differs from the one behind Ember's 36, so the two numbers cannot be compared. Meta's Ember post lists no download link, and the model is offered only through paid APIs and partner clouds. That leaves the licence and the weights unconfirmed.

What the launch partners said

The post quotes partner testers. Dane Knecht, chief technology officer of Cloudflare, said: "What stood out in our testing was speed." Cloudflare reported Ember ran up to 30 percent faster at the tail than a comparable model it tested. Chak Pothina, inference lead at You.com, said Ember had a p95 latency of 2.2 seconds, the lowest of any model You.com tested.

These are quotes selected by Meta for its own launch page. The Terminal has not found a measurement of Ember's speed from a party that Meta did not choose. Compare the pricing approach with the one in the Terminal's report on Claude Fable 5.2 and Mythos 5.2, and see how coding-agent tooling is moving in the agent skills repositories story.

Meta's post gives no date for a licence or weights announcement.