Loading page…
The model handles text, images, audio and video. Meta’s launch benchmarks show gains on business workflows, but not a clean sweep on coding.
Listen to this story
Meta’s Muse Ember is designed to take routine, high-volume subjobs off larger agents, with Muse Wonder handling harder planning. That makes it a delegation-oriented release rather than an all-purpose coding replacement. Available through Muse Code and Meta Model API, Ember also supports multimodal inputs and a one-million-token context window. Its benchmark results are strong on business workflows but mixed on coding, while Meta’s safety review says it remained below specified high-risk thresholds in the categories assessed.
Meta lists Muse Ember at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens.
On AutomationBench-AA, Ember scored 37.8%, ahead of GPT-6 Luna at 20.2% and Claude Haiku 5.5 at 13.9%.
Coding results varied: Ember scored 40.9% on Terminal-Bench 4.0 but 27.2% on FrontierCode 2.0-AA, where Haiku led at 34.9%.
Meta released Muse Ember on November 9, 2026, giving developers a smaller model for the repeated steps inside AI agents: calling tools, reading documents and handling coding subtasks. Available through Muse Code and the Meta Model API, it starts at $0.10 per million input tokens and $0.50 per million output tokens.
Meta’s starting point is the volume of small, structured jobs an agent performs. The company describes Ember as its smallest and most efficient model, designed for routing requests, extracting information and parsing long documents. For coding, its intended assignments include reading files, running tests and making targeted fixes.
The proposed division of labor leaves harder planning to Muse Wonder. Meta recommends pairing the two: Wonder plans, while Ember handles the volume. That positions the new release as a model for delegated work, rather than a replacement for every part of a demanding assignment.
Ember also natively understands text, images, audio and video, Meta says. Its examples include screenshots, charts, scanned documents, voice memos and meeting recordings. Those inputs can be handled without a separate model, making the release relevant to document-heavy workflows as well as software development.
Meta’s listed rate for input tokens.
Meta lists a separate, lower rate for cached input tokens.
Meta’s listed rate for generated output tokens.
Meta’s launch table shows a substantial lead on AutomationBench-AA, a test of end-to-end business workflows. At maximum reasoning, Ember scored 37.8%, against 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5. Meta attributes these results to Artificial Analysis and specifies a strict, single-attempt scoring method.
The coding picture is less uniform. On Terminal-Bench 4.0, Ember scored 40.9%, compared with Haiku’s 39.2% and Luna’s 16.4%. Each model used its best published agent setup. But on FrontierCode 2.0-AA, which judges real engineering tasks on whether changes can be merged, Haiku led at 34.9%; Ember scored 27.2% and Luna 25.2%.
Meta also makes a cost-efficiency case using Artificial Analysis results. Its chart puts one Ember setting at a 35.8 intelligence score and $0.136 per task, versus Haiku’s highest displayed score of 36.2 at $0.411. Ember’s task costs are estimates using its listed pricing, not a guarantee of what a developer’s workload will cost.
Early partners offer examples of the intended work. In a testimonial published by Meta, Box AI products chief Yashodha Bhavnani described scoring contracts against negotiation rules and screening candidates against detailed briefs. Cline founder Saoud Rizwan highlighted using Ember as a subagent—a model assigned part of a larger job—for coding and document analysis.
Before release, Meta says it evaluated chemical and biological risks, cybersecurity risks and loss-of-control risks under its Meta Superintelligence Scaling Framework. Ember remained below the framework’s high-risk thresholds in the categories assessed. That is Meta’s safety assessment, not a claim that agent use carries no risk.
Meta also tested behavior closer to everyday deployment: respecting user intent, handling sensitive information, resisting instructions hidden in documents or tool outputs, and seeking human input before high-impact actions. These are the boundaries the company evaluated for a model meant to act on users’ behalf; the launch describes those checks without publishing numerical results for them.
Story updates
Meta’s new Muse Ember has reached the same rounded benchmark score as Claude Haiku 5.5 while costing about a third as much per task. Artificial Analysis’s November 9 evaluation puts Ember at 36 on its Intelligence Index at maximum reasoning effort—the cheapest model reaching that mark, at an estimated $0.14 per task.
Ember’s advantage is not a lower starting price for ordinary input and output. Its listed rates match those of Haiku 5.5 and GPT-6 Luna. Instead, Artificial Analysis identifies two contributors to the savings: cheaper cached input and less generated text. Cached input accounted for most of the cost in its agent evaluations, making that discount especially important to the benchmark bill.
Ember’s cached-input rate is 60% lower than those competitors’ rates. It also used roughly a third as many output tokens as Haiku at maximum effort. Tokens are the units of content a model processes or generates; fewer billed units can produce a cheaper completed task even when the headline rates look identical.
Uncached input: $0.10 per million tokens.
Cached input: $0.004 per million tokens.
Output: $0.50 per million tokens.
Long requests change the calculation. Above 200,000 input tokens, those rates triple to $0.30, $0.012 and $1.50, respectively. The higher prices apply to the entire request, including output—not just the portion above the cutoff. Cached input and carried-forward conversation history also count toward that threshold.
Artificial Analysis measured a rounded Intelligence Index score of 36 at maximum reasoning effort.
Haiku also scored a rounded 36 at maximum reasoning effort, with a higher estimated cost.
Ember’s clearest strength is AutomationBench-AA, which tests work across simulated business applications. Artificial Analysis measured a 68.1% score, level with Meta’s larger Muse Wonder at maximum effort. That measure awards credit for individual objectives completed without breaking the test’s rules; it is not the share of entire workflows finished successfully.
Meta’s launch results supply that stricter view: Ember fully completed 37.8% of AutomationBench-AA tasks on its first attempt, versus 20.2% for Luna and 13.9% for Haiku. Ember still leads that comparison, but the distinction matters for work that only counts as successful when every required step is done.
Coding produces a split verdict. At maximum effort, Ember beats Haiku on Terminal-Bench 4.0, scoring 40.9% against 32.8%. Haiku wins on FrontierCode 2.0, which grades whether code changes are ready to merge: 34.9% against Ember’s 27.2%. The aggregate intelligence score therefore conceals a meaningful difference between coding assignments.
The comparison shifts again when models run inside their respective coding agents. Ember in Muse Code scores 46 on Artificial Analysis’s Coding Agent Index, ahead of Haiku in Claude Code at 37 and Luna in Codex at 31. Those results compare model-and-agent combinations, rather than three models operating through an identical setup.
Knowledge is Ember’s main weakness in the evaluation. On AA-Omniscience, its knowledge test, Artificial Analysis measured accuracy of 21.1%, versus Haiku’s 34.7%. Ember declined to answer 64% of questions; its measured hallucination rate was 17.4%, about half Haiku’s. The evaluator attributes Ember’s low accuracy mostly to those refusals. Its lower hallucination rate comes alongside fewer answers, not stronger knowledge-test accuracy.
Meta released the proprietary model on November 9, 2026, through its API and Muse Code. VentureBeat describes support for text, images, audio, video and PDFs, with text output and a 1,048,576-token context window—the material it can hold in a request. Meta has kept the model’s weights and parameter count private, so this evaluation offers a hosted-service comparison, not a route to running Ember independently.
venturebeat.comtoday we released our smallest model yet: Muse Ember 🔥 Muse Ember offers nearly Muse Spark 1.3 (max) intelligence at a 17x cheaper price point (per AA) Muse Ember is the first model we’ve released trained using Muse Wonder Full excited for y’all to try on API and muse code!
Loading discussion...
Join the conversation
Explain which tasks you would keep with the stronger model.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.