Meta's Muse Ember Lists at Haiku's and Luna's Price; Its Edge Is Spending Fewer Tokens per Task

Muse Ember does not undercut its rivals on the rate card. Meta's new small model lists at exactly the starting price of Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna. Where it claims an advantage is in how many tokens it burns to finish a task, which is the number that decides what an agent fleet actually costs.

What Meta released#

Meta released Muse Ember on November 9, calling it its "smallest and most efficient model," built for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." It is available as muse-ember-1.0 on the Meta Model API and in Muse Code, with a 1M-token context window and up to 512K output tokens, and accepts text, images, audio and video. Meta said Ember, Muse Spark and Muse Wonder are also available on Google Cloud and Oracle Cloud.

For prompts up to 200K tokens, Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Above 200K, the rates rise to $0.30, $0.012 and $1.50, and VentureBeat notes that the higher rate then applies to the whole request. Meta says Ember is the first model it trained using Muse Wonder Full, the top tier of the Muse Wonder release in October.

Claims vs checks#

Meta's headline claim rests partly on an outside evaluator. In Artificial Analysis's launch evaluation, Ember scored 36 on the Intelligence Index at maximum reasoning effort, the same rounded score as Haiku 5.5, at about $0.14 per benchmark task against $0.41 for Haiku, VentureBeat reported. Meta's chart puts the unrounded scores at 35.8 for Ember and 36.2 for Haiku. GPT-6 Luna scored 27.8 at its highest setting. Artificial Analysis attributes part of Ember's cost edge to cheaper cache reads and says Ember used roughly a third as many output tokens as Haiku at maximum effort, according to VentureBeat.

The results are not one-sided. On FrontierCode 2.0, which grades whether code changes would be merged, Haiku scored 34.9% to Ember's 27.2%. On Meta's own OSWorld 2.1 desktop test, Ember's 69.5% partial-credit score falls to 33.8% when tasks are graded all-or-nothing. Meta's strongest lead is on AutomationBench-AA business workflows, where Ember fully completed 37.8% of tasks against 20.2% for Luna and 13.9% for Haiku.

Several of Meta's other figures, including its Personal Superintelligence Index and Agents' Last Exam scores, come from Meta's internal evaluations, and its Terminal-Bench 4.0 comparison runs each model in a different harness. Those numbers are not directly comparable.

The list-price tie also cuts the other way for some workloads. VentureBeat notes that OpenAI's long-context rates for Luna are lower than Ember's above 200K tokens, so a pipeline with very large prompts could favour Luna on token rates even if Ember is cheaper on the benchmark suite.

Safety and what stays closed#

Meta says it evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks, and found it below the framework's high-risk thresholds. It has not published Ember's weights or parameter count, unlike the Muse Spark Open release in October.

Limits#

• Artificial Analysis's figures come from its launch evaluation as reported by VentureBeat and charted by Meta; we did not independently query its leaderboard.
• Cost per task depends on list prices, cache use and output length; real agent bills will vary with prompt size and retries.
• Meta's safety findings are self-reported, and no outside safety evaluation was published with the launch.

Sources#

• Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember (November 9, 2026)
• VentureBeat: Meta's proprietary Muse Ember nearly matches Claude Haiku 5.5 at 1/3 the estimated per-task cost (November 9, 2026)
• AI at Meta on X: Muse Ember launch post (November 9, 2026)