Meta’s Muse Ember Matches Claude Haiku’s Top Score at About a Third of the Cost per Task

Meta on Monday, Nov. 9, released Muse Ember, its smallest and cheapest AI model, built for the repetitive work that AI agents do all day: calling tools, fixing small pieces of code, routing requests and reading long documents. In launch testing by the independent firm Artificial Analysis, Meta Muse Ember scored 36 on the firm’s Intelligence Index at maximum reasoning, the same rounded score as Anthropic’s Claude Haiku 5.5, at an estimated 14 cents per benchmark task versus 41 cents for Haiku, VentureBeat reported.

The model is available now on Meta’s Model API as muse-ember-1.0 and in Muse Code, Meta’s coding agent, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens, according to Meta’s launch post. CEO Mark Zuckerberg called it “our smallest model that delivers higher intelligence per dollar than others in its class.”

What Meta is selling

Meta says Ember delivers “more intelligence per dollar” than Claude Haiku 5.5 or OpenAI’s GPT-6 Luna at any reasoning setting, The Verge reported. It accepts text, images, audio, video and PDFs, has a context window of about 1 million tokens and can produce up to 512,000 tokens of output. Above 200,000 input tokens, prices triple to $0.30 and $1.50 per million, and the higher rate applies to the whole request, VentureBeat noted.

Meta says Ember is the first model it trained using Muse Wonder Full, its top-tier model, and pitches the pair as a team: Wonder plans, Ember handles volume. Meta Chief AI Officer Alexandr Wang summed it up on X: “[Muse W]onder can plan, ember can handle the volume.” Early partners quoted by Meta included Box, Cline, Oracle, Cloudflare and You.com; Cloudflare CTO Dane Knecht said the model “ran up to 30% faster at the tail than the comparable model we tested.”

What the benchmarks show

Meta’s own chart puts the unrounded index scores at 35.8 for Ember and 36.2 for Haiku. OpenAI’s GPT-6 Luna scored 27.8 at its highest setting for about 15 cents a task.

By The Boston Report’s calculation from those figures, Ember reaches essentially Haiku’s top score for about 67% less per task, while Luna costs roughly the same as Ember but scores about 8 points lower. Artificial Analysis attributed part of Ember’s advantage to cheaper cache reads and to using roughly a third as many output tokens as Haiku, VentureBeat reported. Because list prices for the three models start at the same $0.10 and $0.50 per million tokens, the savings come from how efficiently each model works, not from a lower sticker price.

Ember’s clearest lead is in business workflows. On AutomationBench, which counts a task only if every step is completed, it finished 37.8% on the first try, compared with 20.2% for Luna and 13.9% for Haiku.

Where it falls short

The results are not uniformly strong, and many come from Meta’s own tests. Haiku beat Ember on FrontierCode 2.0, which grades whether code changes are ready to merge, 34.9% to 27.2%. On Meta’s OSWorld 2.1 desktop test, Ember’s 69.5% partial-credit score fell to 33.8% when tasks were graded all-or-nothing. VentureBeat concluded the numbers make a better case for giving Ember bounded jobs with a clear check than for handing it whole engineering tickets unattended. Meta has not disclosed the model’s size, and its weights are not downloadable.

Meta said safety testing under its internal framework found the model below high-risk thresholds for chemical, biological, cyber and loss-of-control risks. Those assessments have not been independently published.

The bigger race

The launch caps a fast year for Meta Superintelligence Labs, which introduced Muse Spark in April and the higher-end Wonder models in October, and recently gave its Muse agent a voice and avatar across the U.S. It lands as rivals push agents into everyday products, from Apple’s plan to turn Siri into an agent to OpenAI’s latest $11 billion fundraising. For businesses running agents at scale, the price per finished task, not per token, is becoming the number to watch.

Sources

• Meta: Introducing our smallest, most efficient model, Muse Ember
• VentureBeat: Meta's proprietary Muse Ember nearly matches Claude Haiku 5.5 at 1/3 the estimated per-task cost
• The Verge: Meta is launching its 'smallest and most efficient' AI model, Muse Ember