Meta's Muse Ember Ties Haiku at 1/3 the Cost

Meta on November 9 released Muse Ember, which it calls its "smallest, most efficient model," at $0.10 per million input tokens and $0.50 per million output tokens. Independent benchmarking firm Artificial Analysis says the model ties Anthropic's Claude Haiku 5.5 on its Intelligence Index while costing roughly a third as much per task.

Key takeaways

• Muse Ember is available as muse-ember-1.0 on the Meta Model API and Muse Code, with a 1M-token context window and up to 512K output tokens.
• Artificial Analysis scores it 36 on its Intelligence Index, the same as Claude Haiku 5.5 (max), at $0.14 per task versus $0.41.
• Its weak spot is knowledge: it declined to answer 64% of AA-Omniscience questions, though that kept its hallucination rate at 17.4%.

Muse Ember price and availability

According to Meta's launch post, Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It supports the Responses API and natively handles text, images, audio and video.

Meta says Ember is the first model trained using Muse Wonder Full, its most capable model. The company also said Muse Ember, Muse Spark and Muse Wonder are now available on Google Cloud and Oracle Cloud.

Meta pitches the model for the "repeatable coding and agentic tasks" that agents run at volume, such as reading files, running tests and calling tools in sequence. It positions Muse Wonder for the hardest problems and Muse Spark for coding and agentic workflows.

Muse Ember vs. Claude Haiku 5.5 and GPT-6 Luna

Artificial Analysis, which published its own evaluation the same day, notes that Ember carries the same input and output token prices as Claude Haiku 5.5 and OpenAI's GPT-6 Luna. The cost gap comes from elsewhere: cached input, most of the bill in agentic tests, is 60% cheaper, and Ember used roughly a third as many output tokens as Haiku 5.5.

The results are mixed in coding. Ember leads Haiku 5.5 on Terminal-Bench 4.0 but trails it on FrontierCode 2.0. Running inside 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.

Why the benchmark numbers don't always match

Readers comparing sources will find different figures for the same test. Meta's chart lists Ember at 37.8% on AutomationBench-AA, labeled "strict pass@1," against 20.2% for GPT-6 Luna and 13.9% for Haiku 5.5. Artificial Analysis's write-up puts Ember at 68.1% on the same benchmark, level with Muse Wonder. Neither page explains the gap.

Terminal-Bench 4.0 shows a similar split. Meta's chart, which uses "each model in its best published harness," gives Haiku 5.5 39.2%, while Artificial Analysis reports 32.8%. Ember's 40.9% is the same in both.

Context: a cheap tier for Meta's agent push

Ember fills the low-cost slot under Muse Wonder, which Artificial Analysis scores 44 on its index at $6.95 per task. That makes Ember eight points lower for about one-fiftieth of the cost. Meta also compares it with Muse Spark 1.3, which scores 37 but costs $2.39 per task at $1.25 input and $4.25 output per million tokens.

In a developer post, Meta showed the intended pattern: Muse Wonder planned a migration of 16 Pygments language lexers from Python to Rust, and 16 Ember agents each ported one lexer in parallel. Meta says the 16 porters finished within about six minutes of each other. The release follows Meta's rollout of Muse, its personal agent, in September.

What's next

Meta says it has "even more efficient models in training." On safety, the company says Ember stayed below the high-risk thresholds of its Meta Superintelligence Scaling Framework across chemical and biological, cybersecurity and loss-of-control evaluations. Several of Meta's own charts, including OSWorld 2.1 and PerceptionBench, come from internal testing that has not been independently reproduced.

FAQ

How much does Muse Ember cost?
$0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, according to Meta.

Is Muse Ember better than Claude Haiku 5.5?
It depends on the task. Artificial Analysis rates them equal overall at 36, with Ember cheaper per task and stronger on Terminal-Bench 4.0, but weaker on FrontierCode 2.0 and knowledge recall.

Where can developers use it?
Through the Meta Model API and Muse Code as muse-ember-1.0. Meta says it is also available on Google Cloud and Oracle Cloud.

References

• Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta AI Research (Nov. 9, 2026)
• Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index — Artificial Analysis (Nov. 9, 2026)
• Start building with Muse Ember, our smallest, most efficient model — Meta Model API