Meta Muse Ember: $0.10 in, ties Haiku 5.5 for agent work

Meta released Muse Ember today, its smallest Muse model, at $0.10 per million input tokens and $0.50 per million output. Meta says it scores 36 on the Artificial Analysis Intelligence Index, the same as Claude Haiku 5.5. The pitch is a division of labor: a big model plans, Muse Ember does the repetitive agent work. For a small business running agents on real volume, that split is where the money is.

What actually happened

From Meta's Muse Ember model page and launch post:

• The price. $0.10 input, $0.50 output and $0.004 cached input per million tokens for prompts up to 200K tokens. Above 200K, it triples: $0.30 in, $1.50 out.
• The specs. Model ID muse-ember-1.0. 1M-token context. Text, image, audio and video in.
• Where it runs. The Meta Model API, which Meta says is OpenAI-SDK compatible, and every Muse Code subscription plan. Meta's pages do not mention open weights. Treat it as API-only.
• The benchmarks. Meta's table puts Ember at 40.9 on Terminal-Bench 4.0 against 39.2 for Haiku 5.5, and 37.8 on AutomationBench-AA against 13.9. Haiku 5.5 wins on FrontierCode 2.0-AA, 34.9 to 27.2. These are vendor-reported numbers.
• The intended setup. Meta says Muse Wonder handles the hardest problems and Ember handles "everyday agent workloads." In its Python-to-Rust example, Ember read 158M input tokens against Wonder's 4.4M, and the whole run cost about $3.96.

TechCrunch frames it as Meta's cheapest, most cost-efficient model yet.

Why Muse Ember matters for your business

Look at that Rust example again. The planner used 4.4 million tokens. The executor used 158 million. In a real agent workload, almost all the tokens are the boring part: reading files, calling tools, extracting fields, retrying. Pay frontier rates for that and your bill scales with volume. Pay $0.10 and it barely moves.

The pattern works with any vendor, not just Meta. Split your agent into a planner and an executor. Write the rules once with a strong model. Run them across ten thousand support tickets, product listings or invoices with a cheap one. Then check the cheap model's output on a sample against your own answers before you trust it.

Two cautions. First, the 200K price step: if your agent stuffs whole document sets into each call, you pay triple, so keep prompts tight. Second, a new API is a new vendor. Wire it through an abstraction layer so Ember is one config line, not a rewrite. If it wins your eval, keep it. If the next cheap model wins, swap it.

Sources: Meta: Muse Ember model page, Meta: Start building with Muse Ember, TechCrunch.