Name: GPT-6 Luna
Author: OpenAI

Mark Zuckerberg pitched Meta‘s newest AI model in a single line on X: “Releasing Muse Ember today — our smallest model that delivers higher intelligence per dollar than others in its class.” Much of the launch coverage read that as a price cut, Meta undercutting the small, cheap models that Anthropic and OpenAI sell to developers who build AI agents.

It is not a price cut, at least not on the price list. Muse Ember charges exactly what its two named rivals charge for every token that goes in and comes out. The saving Meta advertises is real, but it lives somewhere else: in how much the model writes before it reaches an answer. That is a more interesting claim than a discount, and a harder one for a buyer to check before the invoice arrives.

What Meta released

Muse Ember is the fourth model in Meta’s Muse family, after Spark, Glimmer and Wonder, and the smallest. Meta Superintelligence Labs built it for the repetitive steps inside an agent; in Zuckerberg’s words, “coding subtasks, routing, extraction, and long document parsing.” Meta’s own description of the division of labour is blunt: “Muse Wonder plans, Muse Ember handles the volume.” Alexandr Wang, who leads Meta Superintelligence Labs, said Ember is the first released model trained using Muse Wonder Full, Meta’s largest model.

Meta says Ember reads text, images, audio and video, holds a context window of one million tokens and can write up to 512,000 tokens in a single answer. It is closed. There are no downloadable weights, and it runs on Meta’s hosted Model API and in Muse Code, Meta’s coding agent, with Google Cloud and Oracle Cloud also listed as routes.

Where the “per dollar” comes from

Meta lists Ember at $0.10 per million input tokens and $0.50 per million output tokens. Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna list the same base rates, as both MarkTechPost and Fello AI point out. The one list-price edge is cached input, text the model has already seen and can reuse, which Ember bills at $0.004 per million against $0.01 for both rivals.

The gap opens per task. On the Artificial Analysis Intelligence Index, an independent benchmark suite, Ember scores 36, level with Haiku 5.5 at its highest setting, for roughly $0.14 per task against $0.41 for Haiku. Fello AI’s reading of the Artificial Analysis data explains why: across the full index, Ember produced about 115 million output tokens and Haiku about 298 million, at the same output rate. Same price per token, far fewer tokens. GPT-6 Luna costs about the same per task as Ember, $0.15, but scores 28.

That mechanism matters because token counts are printed on no price card. A developer comparing list prices sees three identical models. The difference shows up only once the agent runs, and it moves with the task, the harness and the reasoning setting. Meta’s comparison uses maximum reasoning effort, which its developer documentation also lists as Ember’s default.

The price that triples

Meta’s developer page carries a second price tier that the launch messaging leaves out. Above 200,000 input tokens, Ember costs $0.30 per million in and $1.50 per million out, three times the headline rate. Meta’s documentation warns that a request with a long conversation history can cross that line. MarkTechPost’s launch write-up stated that Meta lists no long-context tier for Ember; Meta’s own pricing table says otherwise.

That tier lands on the exact use Zuckerberg advertised. “Long document parsing” and a million-token window invite developers to feed in whole codebases and document sets, the requests most likely to pass the threshold. Ember still undercuts Haiku 5.5 there, since Haiku charges $0.50 in and $2.50 out above 100,000 prompt tokens, according to MarkTechPost. But “cheap” turns into “cheaper than the alternative”, which is a different promise.

What the benchmarks show, and who ran them

Meta published ten comparison rows against Haiku 5.5 and GPT-6 Luna. Ember’s strongest result is AutomationBench, Artificial Analysis’s run of 657 Zapier business workflows, where it completes 37.8% of tasks against 13.9% for Haiku and 20.2% for Luna. Its one clear loss is FrontierCode 2.0, which grades real code changes on whether a maintainer would merge them: Haiku scores 34.9, Ember 27.2.

Fello AI counted who produced each number. Five rows come from Artificial Analysis, four are Meta’s own runs, and the Terminal-Bench row mixes setups, an Artificial Analysis run for Ember against self-reported figures for the rivals. Fello also notes that the Artificial Analysis model page lists Ember’s inputs as text and image only, so audio and video support rests on Meta’s word for now.

Who Muse Ember is for

Meta’s launch post says Ember’s cost is “low enough to run in every product, for everyone.” For now it runs in none of Meta’s consumer products: Fello AI reports it is absent from the Meta AI app, WhatsApp and the Muse agent. It is a part for other companies’ agents, and the partner quotes Meta chose say so. Cline founder Saoud Rizwan pitched it to “teams cutting costs by running it as the subagent in their harness.” Cloudflare CTO Dane Knecht said it “ran up to 30% faster at the tail than the comparable model we tested.”

That is the bet behind Ember. Meta is not trying to win the chat window with it. It wants to be the call that runs a thousand times an hour inside someone else’s agent while a bigger model does the thinking, and in that job the bill is decided by a number no vendor publishes: how many tokens the model needs to finish.

Tags: meta, Mark Zuckerberg, AI models, Claude Haiku, Muse Ember