Meta has released Muse Ember, its smallest and cheapest AI model yet. It costs 10 cents per million input tokens and is built to do the repetitive work AI agents handle all day, and Meta says it beats Anthropic's and OpenAI's small models on value.
Meta Superintelligence Labs released Muse Ember on Monday, November 9, describing it as the lab's "smallest and most efficient model." It is designed for what Meta calls "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing," according to Meta's announcement. The model is available now as muse-ember-1.0 through the Meta Model API and Meta's Muse Code tool. Meta also says its Muse models are now available on Google Cloud and Oracle Cloud.
The headline is price. At $0.10 per million input tokens and $0.50 per million output tokens, Ember is much cheaper than frontier rivals such as Anthropic's Claude Haiku and OpenAI's GPT-6 Luna, and it is competitive with low-cost open-source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash, TechCrunch reported.
What Muse Ember Can Do
According to Meta's developer documentation, Ember has a context window of about 1 million tokens (1,048,576) and can produce up to about 512,000 tokens of output. That is enough to take in an entire codebase or a large set of documents in a single request. It accepts text, images, video, audio and PDFs, and replies in text. Developers can choose a reasoning effort from "minimal" to "max," and the default is max.
Meta says the multimodal support matters for agents. They can "read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model." The model also supports computer use, tool calling, structured output and search grounding. It accepts the same requests as the larger Muse Spark, so in most cases switching is just a matter of changing the model name.
Prices go up for very long prompts. For requests over 200,000 input tokens, Ember costs $0.30 per million input tokens and $1.50 per million output tokens. Cached input is priced at $0.004 per million tokens at standard length.
Meta's Benchmark Claims
Meta is pitching Ember on cost per task rather than raw intelligence. Citing Artificial Analysis results published on November 9, the company says Ember "matches Haiku's best score at about a third of the cost and beats every Luna setting." On the Artificial Analysis Intelligence Index, Ember at max reasoning scored 36, level with Claude Haiku 5.5 and ahead of GPT-6 Luna's 28. Ember's cost per task was about $0.14, while Haiku's top setting cost about $0.41.
Meta's own charts show the biggest lead on agent workflows. On AutomationBench-AA, which tests end-to-end business workflows, Ember scored 37.8% against 20.2% for Luna and 13.9% for Haiku. It also led on Terminal-Bench 4.0 for terminal-based coding (40.9% vs. 39.2% for Haiku) and OSWorld 2.1 for computer use. It did not win everywhere: on FrontierCode 2.0, which grades real engineering tasks by whether they could be merged, Haiku scored 34.9% to Ember's 27.2%.
These numbers come with caveats. Several are Meta's own internal evaluations, each model was tested in a different harness on some benchmarks, and Ember's cost figures are estimates based on list pricing. Independent tests in real deployments will matter more than launch charts.
Built to Work Alongside Bigger Models
Meta is not presenting Ember as a model for everything. Its suggested setup pairs it with a larger model: "Muse Wonder plans, Muse Ember handles the volume." TechCrunch gave the example of a customer-support agent, where a heavier model decides how to handle an issue and Ember then applies those rules across tens of thousands of cases. For hard one-off coding problems or knowledge-heavy questions, Meta's documentation recommends Muse Spark or Muse Wonder instead.
Meta says Ember is the first model trained using Muse Wonder Full, its most capable system. It also compares Ember with its mid-sized Muse Spark 1.3. The two score almost the same on the Artificial Analysis index (36 vs. 37), but Ember costs about $0.14 per task compared with $2.39 for Spark.
What Early Partners Say
Meta published comments from partners who tested the model early. Cloudflare CTO Dane Knecht said Ember "ran up to 30% faster at the tail than the comparable model we tested." You.com's inference lead, Chak Pothina, said it finished every coding task in the company's tests with a p95 latency of 2.2 seconds, which he described as about 4x to 13x faster than leading open-weight models. Box and Cline described it as a strong fit for enterprise document work and sub-agent coding tasks. These are partner endorsements published by Meta, not independent reviews.
On safety, Meta says it tested Ember under its Superintelligence Scaling Framework for chemical, biological, cybersecurity and loss-of-control risks, and found it stayed below the framework's high-risk thresholds. It also says it tested whether Ember resists hidden instructions in web pages and documents, and whether it asks a human before taking high-impact actions.
Why It Matters
Ember is the fourth model under Meta's Muse brand, after the mid-sized Spark in April, its open-source twin Glimmer in August and the Wonder frontier model in October, TechCrunch noted. Unlike Glimmer, Ember is available only through Meta's API and Muse Code, not as downloadable weights.
Our take: Ember shows where the AI race is heading. As companies run agents that make thousands of small model calls, the cost of each call starts to matter as much as top-end capability. A model this cheap that comes close to its rivals on agent benchmarks puts pressure on Anthropic's and OpenAI's small models on price. Meta says it has "even more efficient models in training," so price competition at the low end is likely to continue.