1. Home
2. News
3. Meta

Meta's Muse Ember matches Claude Haiku 5.5 at about a third of the cost per task

Artificial Analysis found it matches Claude Haiku 5.5 for about a third of the cost per task, mostly by declining to guess.

By Think Facility · November 9, 2026 · Meta AI models Save

Part of a running story: Meta's Muse agent and models (7 parts)

On November 9, 2026, Meta released Muse Ember, the smallest model in its Muse family. Meta built it for the repetitive part of agent work: tool calls, coding subtasks, routing, pulling data out of documents and reading long ones.

It costs $0.10 per million input tokens and $0.50 per million output, the same list price as Claude Haiku 5.5 and GPT-6 Luna. Artificial Analysis, which tests models independently, found it does the same work for much less.

Released

November 9, 2026

Model ID

muse-ember-1.0

Price

$0.10 input, $0.004 cached input, $0.50 output per 1M tokens

Context window

1M tokens

Intelligence Index

36, matching Claude Haiku 5.5 (max)

Where

Meta Model API and Muse Code

Why it's cheaper than its price

Meta announced it on X a little after 1 pm Eastern.

The list price is a tie, so the gap comes from elsewhere. In Artificial Analysis's agent tests, cached input is most of the bill, and Ember's is 60% cheaper than its rivals'. It also writes roughly a third as many output tokens as Haiku 5.5. Put together, Ember at max effort scored 36 on the group's Intelligence Index for $0.14 a task. Haiku 5.5 got the same 36 for $0.41.

It isn't better at everything. Ember beat Haiku 5.5 on Terminal-Bench 4.0 (40.9% to 32.8%) but trailed it on FrontierCode 2.0, 27.2% to 34.9%. Running inside Meta's own Muse Code, it scored 46 on the Coding Agent Index, ahead of Haiku 5.5 in Claude Code at 37 and GPT-6 Luna in Codex at 31.

It says "I don't know" a lot

Knowledge is the weak spot. Ember got 21.1% on AA-Omniscience, the lowest of the similar models Artificial Analysis compared, against 34.7% for Haiku 5.5. Most of that's because it declined to answer 64% of the questions.

That trade has an upside. Its hallucination rate came out at 17.4%, half of Haiku 5.5's. For a model meant to pull fields out of contracts, I'd take a blank over a made-up answer.

One thing to check before you rely on the chart on Meta's page. Meta lists Ember at 37.8 on AutomationBench-AA, while Artificial Analysis reports 68.1% for the same test name, and the two give different Terminal-Bench scores for Haiku 5.5 (39.2 on Meta's page, 32.8% in the article). Neither page explains the difference.

What it means for people's work

Ember is aimed at office work done in bulk, and the customer quotes on Meta's page say what kind. Box says it can score a contract against a company's negotiation playbook clause by clause, and screen job candidates against a brief while flagging borderline ones as alternates.

That second use reaches people who never touch an API. If a model priced at a few cents a task is sorting applications, the first reader of a CV may be a subagent. Meta's page doesn't say how those screening results were checked, and Box's quote gives no error rate.

For individual coders, Muse Code subscriptions start at $5 a month, and Meta's page lists plans up to $200. Ember sits alongside the larger Muse Wonder, which Artificial Analysis scored 44, eight points higher, at about 50 times the cost per task.