Meta's Muse Ember costs 10 cents per million input tokens, and the independent index puts it second in its class

Meta Superintelligence Labs released a small model on Monday, November 9, 2026, and the interesting part is the price. Muse Ember costs $0.10 per million input tokens, according to both TechCrunch and the independent leaderboard Artificial Analysis, which also lists $0.50 per million output tokens. It is the lab's lightest and most cost-efficient model so far.

For a lab whose stated goal is personal superintelligence, a cheap workhorse model is a strategic move more than a technical showpiece. Here is what is verified, and what is Meta's word.

What Meta says it built

TechCrunch reports that Meta describes Ember as "designed for complex reasoning execution and the work agents repeat all day." The idea is a division of labor. A larger model sets rules or plans a task, and Ember applies them at volume. Meta's own example, relayed by TechCrunch, is a heavier model defining the rules for a customer-support issue while a lighter one applies them across tens of thousands of cases. Its line for the pairing: "Muse Wonder plans, Muse Ember handles the volume."

Ember is multimodal. TechCrunch says it was trained on text, images and audio, and that Meta says "agents can read screenshots, charts, and scanned documents." We could not open a Meta page for the launch, so these descriptions are as relayed by TechCrunch.

What the independent index shows

Artificial Analysis, which runs its own tests, lists Muse Ember (Max) at an Intelligence Index of 36. That ranks second of 33 models in its class, where the median is 24. It measures 446 output tokens per second, 4.37 seconds to the first answer token and 3.2 minutes per Intelligence Index task, with a 1 million token context window. The listed release date is November 9, 2026.

Those are independent numbers, which is the right standard for a price and speed claim. The blended rate Artificial Analysis computes is about $0.073 per million tokens.

Two cautions apply. First, "in its class" is Artificial Analysis's grouping, and a second-place finish among small, cheap models says little about how Ember compares with the frontier. Second, the Intelligence Index is a composite of benchmarks, and composites can hide weaknesses on the specific task you care about.

What is only Meta's word

Meta compared Ember's cost per task and performance against Anthropic's Haiku and OpenAI's Luna, using the Artificial Analysis composite, TechCrunch reports. Those comparisons are Meta's, and the article does not give Haiku or Luna pricing. TechCrunch also describes Ember as cost-competitive with the open models DeepSeek V4.1 Flash and GLM 5.3 Flash without verified figures for either.

So the claim that Ember beats the small models from Anthropic and OpenAI on cost per task is plausible and unproven. Cost per task depends on how many tokens a model spends thinking, which is not visible from a price per million tokens.

Where it sits in Meta's lineup

TechCrunch lays out the Muse family. Muse Spark, a mid-weight model, came in April 2026. Glimmer, an open-source twin of Spark, followed in August. The Muse Agent consumer product launched in September, and TechCrunch later reported it as the number 2 app in the US. Muse Wonder, which Meta calls its first frontier model, arrived on October 12, 2026 at $4 per million input tokens and $20 per million output tokens, with Meta's model card putting it comparable to Fable 5.1 and GPT-6 Astra. We covered that launch in our piece on Muse Wonder and the missing benchmarks.

Ember at $0.10 is 40 times cheaper on input than Wonder at $4. That ratio is the business model. Agents burn tokens in loops, and a company that can run the loop cheaply and call the expensive model only for planning wins on cost.

Why it matters for the race

Frontier headlines go to the biggest models, but deployment volume goes to the cheapest adequate one. A model that scores 36 on an independent index at ten cents per million tokens puts pressure on every lab's small-model tier. It also tells us Meta's strategy is a full ladder: a frontier model, a mid-weight consumer model, an open-weight release and now a volume model.

Meta has not said what Ember cost to train or whether the weights will be open. Those are the questions that decide whether this is a product or a precedent.

What to test before you switch

Three practical checks follow from the numbers. First, measure tokens spent per task: a model at 446 output tokens per second that thinks for a long time can still cost more than a slower, terser one. Second, test the 1 million token context window on your own documents, since long-context quality varies far more than the headline length. Third, check where the model is served and under what terms. Meta says Wonder was trained at its Prometheus data center in central Ohio and calls the Wonder models "the next step on our scaling ladder and a major leap toward personal superintelligence," per TechCrunch. Ember is the commercial end of the same ladder, and Meta has published no safety results we could open for it.

Our take

Ember is a real, cheap, fast model by the one independent measure we could open. We would test it on a repetitive agent workflow before believing the Haiku and Luna comparison, and we would watch whether Meta publishes a model card with safety results. A lab pursuing superintelligence that sells its smallest model at ten cents is telling developers where it wants the volume to run.