Meta has released Muse Ember, the smallest and most efficient model in its Muse family, and it’s pitching it as the cheapest way to run the repetitive work that AI agents do all day. According to Meta, Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 and GPT-6 Luna at every reasoning setting, matching Haiku’s best score on the Artificial Analysis Intelligence Index at about a third of the cost.
Announcing the model, Meta AI chief Alexandr Wang said Ember offers more intelligence per dollar than Haiku, Luna and every other “flash” model at any reasoning setting, at a third of the cost or less. Muse Ember is available now on Muse Code and the Meta Model API.
What Is Muse Ember? #
Muse Ember is built for the high-volume steps that make up most of an agent’s workload: tool calls, coding subtasks, routing, extraction and long document parsing. It’s also the first Meta model trained using Muse Wonder Full, the larger model Meta released in October as part of its Muse Wonder series.
The model is natively multimodal, taking in text, images, audio and video, and has a 1M-token context window with up to 512K tokens of output. Meta says that’s enough to ingest an entire codebase or document set in a single call.
Ember rounds out a three-tier lineup: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Ember for repeatable agentic tasks at scale.
Muse Ember Benchmarks: Same Score As Claude Haiku 5.5, A Third Of The Cost #
On the Artificial Analysis Intelligence Index, Muse Ember at max reasoning scores 36, effectively tied with Claude Haiku 5.5 at max (36.2 against Ember’s 35.8) and well ahead of GPT-6 Luna at max, which scores 28.
The difference is what it costs to get there. Artificial Analysis puts Ember’s cost per task at $0.136 at max, against $0.411 for Haiku 5.5 at max. That’s roughly a third of the price for the same score. Even Ember’s lower xhigh setting (33.9 at $0.116 per task) beats Luna’s best result (27.8 at $0.152).
The gaps hold across the rest of the curve:
- Ember max vs Haiku xhigh: 35.8 vs 34.3, at $0.136 vs $0.248 per task
- Ember xhigh vs Haiku high: 33.9 vs 29.6, at $0.116 vs $0.159 per task
- Ember max vs Luna max: 35.8 vs 27.8, at $0.136 vs $0.152 per task
Meta also says Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard, meaning no model scores higher for less per task.
Muse Ember vs Claude Haiku 5.5 vs GPT-6 Luna: Benchmark Table #
All three models were run at max reasoning effort. Muse Ember leads or ties on nine of the ten evaluations Meta published. The one exception is FrontierCode 2.0-AA, where Haiku 5.5 is well ahead.
The biggest margin is on AutomationBench-AA, which tests end-to-end business workflows. Ember scores 37.8, nearly triple Haiku 5.5’s 13.9 and almost double Luna’s 20.2. It also leads clearly on the Coding Agent Index, 46 against 37 for Haiku and 31 for Luna, and on computer use, where it edges Haiku on OSWorld 2.1 and is far ahead of Luna.
Muse Ember Pricing: Same Rate Card, Fewer Tokens #
On the Meta Model API, Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. The model ID is muse-ember-1.0, and it supports the Responses API.
Those per-token prices are identical to GPT-6 Luna’s $0.10 and $0.50, and match Haiku 5.5’s standard rate for prompts up to 100K tokens. So Ember’s advantage in cost per task comes from how many tokens it burns to reach an answer, not from a lower sticker price. For agents that make thousands of calls a day, that’s the number that ends up on the invoice.
Against Meta’s own models the gap is wider. Compared with Muse Spark 1.3, which scores 37 on the same index, Ember scores 36 but costs $0.14 per task against $2.39, and its token prices are $0.10/$0.50 versus $1.25/$4.25 per million. That’s about 17 times cheaper per task for a one-point drop.
Pairing Muse Ember With Muse Wonder #
Meta’s suggested setup is to use both ends of the lineup. Muse Wonder, which is priced at $4 per million input tokens and $20 per million output tokens, plans the work, and Ember handles the volume. It’s the same planner-and-worker pattern Anthropic pitches for Haiku 5.5 alongside Opus 5.5 and Sonnet 5.5, and it’s aimed at teams trying to keep agent bills down as usage scales.
Meta is also lining this up with its consumer agent push. Muse, its AI agent, has been signing up partners like Shopify and Expedia, and a model cheap enough to run in every product is the kind of thing that makes that economics work.
Caveats On The Muse Ember Benchmarks #
A few things are worth keeping in mind. Several of the evaluations, including Agents’ Last Exam, the Personal Superintelligence Index, OSWorld 2.1 and PerceptionBench, are Meta’s own internal runs, and the Personal Superintelligence Index is an internal composite. On Terminal-Bench 4.0, each model was run in its best published harness (mini-swe-agent for Ember, Claude Code for Haiku, Codex for Luna), so the comparison isn’t like for like.
The Artificial Analysis numbers are also from the new Intelligence Index v5, which reset the leaderboard with majority-private datasets. That’s why Haiku 5.5 scores 36 here after scoring 43 on the earlier version, and why the two sets of scores shouldn’t be compared directly. Ember’s costs are also calculated at Meta’s list pricing.
On safety, Meta says Ember was evaluated under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and stayed below the framework’s high-risk thresholds. Meta also tested how it behaves when acting on a user’s behalf, including whether it resists instructions hidden in web pages, documents or tool outputs.
How To Access Muse Ember #
Muse Ember is available today on Muse Code and the Meta Model API as muse-ember-1.0. The release continues the pace Meta Superintelligence Labs has kept up since Muse Spark debuted in April, under Alexandr Wang’s leadership. Meta says it has even more efficient models in training.