Meta's Muse Ember Is a Tiny Model With a Big Price Cut
Meta launched Muse Ember on November 9, a small model priced at $0.10 per million input tokens that scores close to Meta's flagship on a major benchmark. Cheap models are quietly becoming the real story of 2026.

Meta Muse Ember launched on November 9, and it is the kind of release that barely makes headlines but quietly changes what developers build. Ember is the smallest model in Meta's Muse family, it runs through the Meta Model API and Muse Code, and it costs $0.10 per million input tokens and $0.50 per million output tokens. Above 200,000 tokens of context, the price rises to $0.30 in and $1.50 out. For a frontier lab, that is close to pocket change.
The number that matters is 36.0. That is Ember's score on Artificial Analysis's Intelligence Index, the benchmark that has become the closest thing the industry has to a shared scoreboard. Meta's flagship, Muse Spark 1.3, scores 36.6 on the same test. Ember gets within a rounding error of the flagship at about 6 percent of the cost per task, according to the same index's cost figures. If you run a lot of small jobs, that gap is the whole business case.
What Meta Muse Ember is actually for
Ember is not trying to win the benchmark crown. It arrived on the same day Anthropic's Claude Fable 5.2 took the top spot on that same index at 64.6, passing Google's Gemini 4 Carbon. The frontier race and the efficiency race are now two different sports, and Ember is playing the second one.
Small models earn their keep on volume work: classifying support tickets, summarizing documents, routing requests, drafting first passes that a bigger model checks later. At ten cents per million input tokens, a team can afford to point an AI at problems that were never worth a frontier model's price. The 200,000-token threshold is the catch to watch. Long-context work costs three times as much, so Ember is cheapest exactly where the job is short and repeatable.
Meta is also clearly thinking about its own developer tools. Ember is available in Muse Code, the company's coding assistant, where a cheap model that answers fast is worth more than a brilliant one that takes its time. Coding help is the use case where latency and cost per call decide whether a tool gets used a hundred times a day or once a week.
The cheap model wave is the real November story
November has already seen seven model launches, and the pattern is hard to miss. OpenAI split its image model in two and shipped a security model, GPT-6 Cyber, for approved organizations only. Tencent shipped Hy4 with vision and computer use, and Thinking Machines shipped Inkling 2 Small while promising open weights later. The frontier models keep getting bigger and more restricted. The small ones keep getting cheaper and more open.
That split is good news for anyone building products rather than demos. A year ago, the sensible default was to call the biggest model available and eat the bill. Now the sensible default is a small model for the routine work and a big one only when the task actually needs it. Ember's pricing makes that routing decision easy: the small model is not a compromise anymore, it is the default, and the flagship is the exception.
There are limits worth naming. Ember's benchmark score is strong for its size, but 36.0 is still a long way from the 60-plus scores the frontier models post. Hard reasoning, long documents, and genuinely novel problems still belong to the big models. And Meta has not published the full evaluation details that would let independent researchers verify the score the same way they can for open-weight releases.
But the direction is clear. The most interesting AI releases of late 2026 are not the biggest models. They are the smallest ones that are good enough, priced so low that using them stops being a decision at all. Muse Ember is the latest proof, and at ten cents a million tokens, it is the cheapest proof yet.


