Meta's Muse Ember matches Claude Haiku at a third of the cost

What you could do with it

• A software development team could deploy an AI agent to handle repetitive coding subtasks, tool calls, and request routing around the clock at roughly a third of the cost compared to rivals ¹.
• A business running customer-support agents could route requests, pull data from files, and sort inquiries without paying premium prices per call ¹.
• A developer could pair Ember with Meta's larger Muse Wonder model so Ember processes the heavy daily volume of text, images, audio, and video while Wonder plans harder problems ².

Cheaper for the same smarts

For every million input tokens, Ember costs $0.10, while output tokens run $0.50 per million. Its two closest rivals, Claude Haiku 5.5 and GPT-6 Luna, carry the same list prices.

But Meta says you get more done for each dollar. The claim is backed by the Intelligence Index from Artificial Analysis, an independent benchmarking firm. On its highest effort setting, Ember scored 36 — the same score as Haiku 5.5. Luna scored 28.

Running that benchmark test cost about $0.14 per task with Ember, $0.41 with Haiku, and $0.15 with Luna. That makes Ember roughly level with Haiku at about a third of the cost per equal result ¹².

"Our teeny model muse ember is now on the efficient frontier, matching haiku at 3x cheaper cost per task," Meta AI chief Alexandr Wang wrote on X ¹.

How it scores elsewhere

On AutomationBench, which tests full business workflows end to end, Ember scored 37.8%. Luna scored 20.2% and Haiku only 13.9% ².

It also won Meta's own Internal Terminal-Bench coding test and the Agent Index from Artificial Analysis ². But it does not win everything. On FrontierCode 2.0 — a benchmark that checks whether real engineers would accept a model's code changes — Haiku earned 34.9% and Ember scored 27.2% . Several other comparisons, including Meta's Personal Superintelligence Index, come from Meta's own internal tests ¹.

Built to work with a bigger brain

Ember handles text, images, audio and video. It can read up to one million tokens at once — enough for a whole codebase. Meta plans to pair it with Muse Wonder, its bigger frontier model. "For the hardest problems, pair it with Muse Wonder: Muse Wonder plans, Muse Ember handles the volume," Meta wrote ².

Ember is also the first Meta model trained using the larger Muse Wonder Full system ². Weights stay private, as with previous Muse releases ¹. Meta put Ember through its own safety framework, which includes tests for chemical and biological weapons, cybersecurity threats, and loss of control risks. The company says Ember stayed below high-risk thresholds in every category .

Ember is available now on the Meta Model API and in Muse Code, Meta's coding tool ¹². More efficient models are still in training, Meta said .

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

Key facts

¹ - Muse Ember charges $0.10 for every million input tokens and $0.50 for every million output tokens.

• Ember matches Claude Haiku 5.5's score of 36 at about a third of the cost per task ($0.14 vs $0.41)
• Ember scores 37.8% on AutomationBench, beating GPT-6 Luna (20.2%) and Claude Haiku 5.5 (13.9%)
• Haiku beats Ember on FrontierCode 2.0 (34.9% vs 27.2%); some results from Meta's internal tests
• Available now via Meta Model API and Muse Code

Sources

1. Meta’s Muse Ember matches Claude Haiku at a third of the cost — thenextweb.com
2. Introducing Our Smallest, Most Efficient Model: Muse Ember