Meta’s Ember Model Targets Cheaper, Smarter AI Agents

Meta has released Muse Ember, a compact model built for the repetitive work that makes AI agents useful. It handles tool calls, coding subtasks, routing, extraction, lookups, document reading, and long document parsing, aiming to bring stronger reasoning to every step an agent takes.

The launch lands on Monday with a clear target: make agent execution cost less without turning every task over to a larger model. Ember can run tests, make small code fixes, read multiple media types, and complete the routine operations that agents repeat all day.

A Smaller Model With a Big Agent Workload

Meta describes Muse Ember as its smallest and most efficient model for complex reasoning execution. The company positions it as a partner to Muse Wonder, with Alexandr Wang writing, “Muse Wonder can plan, ember can handle the volume.” Meta’s launch message puts the idea in even clearer terms: “Muse Wonder plans, Muse Ember handles the volume.”

That division matters because an agent may need to perform many modest actions before producing one answer. Ember is built to manage those steps, including lookups, document reads, tool calls, test runs, and small coding changes. It also accepts text, images, audio, video, and PDFs, then returns text.

Meta says agents can use Ember to read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model. The model supports six reasoning levels: minimal, low, medium, high, xhigh, and max, with max set as the default.

Mark Zuckerberg described Ember as “our smallest model that delivers higher intelligence per dollar than others in its class.” The model is available through Muse Code, giving coding workflows a direct route to its reasoning and execution capabilities.

Pricing Built Around Long Context and Repeated Tasks

Muse Ember supports a context window of 1,048,576 tokens and a maximum output of 524,288 tokens. That puts it alongside GPT-6 Luna and Claude Haiku 5.5, which also support up to a million input tokens, although all three models raise prices after a threshold.

For up to 200,000 input tokens, Meta charges $0.10 per million uncached input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Above 200,000 tokens, the rates rise to $0.30, $0.012, and $1.50, respectively.

For shorter prompts, Ember matches GPT-6 Luna and Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens. Claude Fable 5.2, another major model release, is listed at $8 per million input tokens and $40 per million output tokens.

Meta offers Ember as muse-ember-1.0 through its Responses, Chat Completions, and Messages endpoints. Oracle Cloud is another route, while Google Cloud’s preview documents text and image inputs.

Artificial Analysis attributes part of Ember’s benchmark cost advantage to cheaper cache reads. The firm also found that Ember used roughly a third as many output tokens as Haiku at maximum effort, a difference that matters when an agent repeats a task across thousands of operations.

Ember’s Scores Put Pressure on Larger Models

In Artificial Analysis’s launch evaluation, Ember scored 36 on the Intelligence Index at maximum reasoning effort, the same rounded score as Claude Haiku 5.5. Ember cost about 14 cents per benchmark task, compared with 41 cents for Haiku. Meta’s underlying chart lists unrounded scores of 35.8 for Ember and 36.2 for Haiku.

GPT-6 Luna scored 27.8 at about 15 cents per task at its highest setting, while a lower setting brought Luna’s cost to about 7 cents with a score of 21.4. Meta’s comparison therefore focuses on the balance between capability and the number of tasks a model can complete before costs become a concern.

On AutomationBench-AA, Ember fully completed 37.8% of tasks on its first attempt, compared with 20.2% for Luna and 13.9% for Haiku in Meta’s launch results. Artificial Analysis’s headline measure puts Ember at 68.1%, but the two numbers answer different questions.

In the Coding Agent Index, Ember running in Muse Code scored 46, ahead of Haiku in Claude Code at 37 and Luna in Codex at 31. Haiku led Ember on FrontierCode 2.0, which measures whether changes are ready to merge, scoring 34.9% to Ember’s 27.2%.

Ember’s results also show why benchmark labels matter. In Meta’s desktop-app evaluation, OSWorld 2.1, Ember earned a 69.5% partial-credit score, but that result fell to 33.8% when tasks received an all-or-nothing grade.

Meta’s Expanding Model Lineup

Meta has kept Ember’s weights and parameter count private. The scores above use Artificial Analysis’s current v5 methodology, after the firm revised its index in October.

DeepSeek-V4.1-Flash-1017 is a close alternative for inexpensive hosted agent work, with a million-token window and image input. Its API charges $0.15 input and $0.60 output per million tokens off-peak, or $0.30 and $1.20 at peak, while cache hits cost $0.003 or $0.006. In Artificial Analysis’s Ember comparison, DeepSeek Flash scored 33 at about $0.60 per benchmark task.

Meta also offers Muse Spark Open for teams that want to run a million-token multimodal Muse model on infrastructure they control. It lists 816 billion total parameters and 42 billion active per token, but a four-bit checkpoint weighs about 537 GB before serving overhead. For a smaller local option, the 30-billion-parameter Muse Glimmer is available under Apache 2.0.

Ember is the fourth model under the Muse brand, following Spark, released in April; Glimmer, released in August; and Wonder, released in October. With agents moving toward always-available work, the pressure now falls on models that can reason well while handling large volumes of small tasks. Ember is Meta’s answer to that pressure, and its pricing and benchmark results make the next agent battle about efficiency as much as intelligence.

Stay connected via Google News

Follow us for the latest AI News