Meta AI released Muse Ember on Monday, Nov. 9, calling it the company's smallest and most efficient model. Meta built Ember for the repetitive work inside agent systems: tool calls, coding subtasks, routing, data extraction and long-document parsing. Pricing starts at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens.
Ember carries a 1-million-token context window and can return up to 512,000 output tokens. It natively handles text, images, audio and video. Prompts longer than 200,000 tokens cost $0.30 per million input tokens and $1.50 per million output tokens, and cached input runs $0.004 or $0.012 per million depending on prompt length. The model, listed as muse-ember-1.0, is available now on Muse Code and the Meta Model API and supports the Responses API.
Meta's pitch is intelligence per dollar. The company says Ember beats Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna on that measure at every reasoning setting, matching Haiku's best score at about a third of the cost. On the Artificial Analysis Intelligence Index v5, published Nov. 9, Ember scores 36, tied with Haiku 5.5 and ahead of Luna at 28. Meta puts Ember's cost at about $0.14 per Artificial Analysis task, compared with $2.39 for its own Muse Spark 1.3, which scores 37.
Key facts
- 01Meta released Muse Ember on Nov. 9 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens.
- 02Ember has a 1M-token context window, up to 512K output tokens, and accepts text, image, audio and video input.
- 03Ember scores 36 on the Artificial Analysis Intelligence Index v5, tying Claude Haiku 5.5 and ahead of GPT-6 Luna at 28.
- 04Meta puts Ember's cost at about $0.14 per Artificial Analysis task, versus $2.39 for Muse Spark 1.3, which scores 37.
- 05Ember is live on Muse Code and the Meta Model API, and the Muse family is available on Google Cloud and Oracle Cloud.
The widest gaps show up on agent and automation benchmarks. Ember scored 37.8% on AutomationBench-AA, a strict pass@1 test of business workflows, against 13.9% for Haiku 5.5 and 20.2% for Luna. On Terminal-Bench 4.0, using each model's best published harness, Ember posted 40.9% to Haiku's 39.2% and Luna's 16.4%. Its Coding Agent Index-AA score was 46, versus 37 and 31.
“Muse Ember is purpose-built for, and excels at, subagent work like coding and document analysis.”— Saoud Rizwan, Founder of Cline
Meta's own table does not show a clean sweep. Haiku 5.5 leads FrontierCode 2.0-AA with 34.9, ahead of Ember at 27.2 and Luna at 25.2. Ember and Luna tie at 22.8% on GDP.pdf-AA. Several other results Meta cites, including Agents' Last Exam, its Personal Superintelligence Index and PerceptionBench, are internal Meta evaluations rather than third-party tests.
Ember is the first model Meta trained using Muse Wonder Full, and it fills out a three-tier lineup. Meta positions Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Ember for repeatable tasks at scale. Ember's list price is a fraction of Muse Spark 1.3's $1.25 per million input tokens and $4.25 per million output tokens. All three Muse models are available on Google Cloud's Model Garden and on Oracle Cloud.
Meta published early results from partners including Box, Cline, Oracle, Cloudflare and You.com. Cloudflare CTO Dane Knecht said Ember ran up to 30% faster at the tail than the comparable model his team tested. Chak Pothina, inference lead at You.com, said Ember completed every coding task in its testing with a p95 latency of 2.2 seconds, the lowest of any model it tried.
“What stood out in our testing was speed.”— Dane Knecht, CTO of Cloudflare
On safety, Meta said it evaluated Ember under its Superintelligence Scaling Framework, which covers chemical and biological risk, cyber capability and loss of control, and that the model stayed below the framework's high-risk thresholds. Meta also tested agent behavior, including whether Ember resists instructions hidden in web pages or tool outputs.
The launch landed the same day Anthropic released Claude Fable 5.2 and Mythos 5.2 and cut Fable's API price to $8 per million input tokens and $40 per million output tokens. Meta's numbers put Ember at the opposite end of the market: at roughly $0.14 per Artificial Analysis task, the company says the model is cheap enough to run in every product.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
