Meta's Muse Ember Matches Claude Haiku 5.5 on Intelligence at About a Third of the Cost per Task

Meta has a new answer to the most boring and most expensive part of running AI agents: the thousands of small calls they make every hour. On Monday the company released Muse Ember, its smallest model, built for tool calls, coding subtasks, routing, data extraction and long document parsing.

The pitch is cost per job, not price per token. Ember lists at $0.10 per million input tokens and $0.50 per million output tokens, exactly what Anthropic charges for Claude Haiku 5.5 and OpenAI for GPT-6 Luna. Meta says it still does more work for each dollar than either rival at every reasoning setting.

Same score, smaller bill

On the Intelligence Index run by the independent benchmarking firm Artificial Analysis, Ember scores 36 at its highest reasoning setting, level with Haiku 5.5 at its highest. Luna tops out at 28. Running one index task cost about $0.14 with Ember against $0.41 with Haiku, which makes Ember the cheapest model on the leaderboard to reach that score.

The savings come from two places. Cached input, which dominates the bill in agent workloads that resend the same context over and over, costs $0.004 per million tokens, about 60% less than on Haiku and Luna. Ember also writes far less: around 84,000 output tokens per index task, roughly a third of what Haiku uses at max effort.

Meta's own tables favor Ember on agent work. It scored 37.8% on AutomationBench, a test of end to end business workflows, against 13.9% for Haiku and 20.2% for Luna, and 40.9% on Terminal-Bench 4.0. Several of the other numbers, including Meta's Personal Superintelligence Index, come from internal tests nobody outside the company can rerun.

Where it falls short

Ember does not win everything. On FrontierCode 2.0, which judges whether maintainers would actually merge a model's code, Haiku scored 34.9% and Ember 27.2%. Its weakest spot is general knowledge, where Artificial Analysis measured 21.1% accuracy against 34.7% for Haiku. The flip side is a hallucination rate about half of Haiku's, so the model more often says it does not know.

Meta is not pretending Ember can run a whole project. Its developer guide tells builders to use it for repeated, bounded tasks and hand the planning to Muse Wonder or Muse Spark. Wonder scores eight points higher on the index for roughly fifty times the cost per task.

Ember is live on the Meta Model API and in Muse Code as muse-ember-1.0, with a one million token context window and up to 512,000 output tokens. Prompts above 200,000 input tokens cost three times more. It slots in under the Muse family Meta launched in September, as Meta's bid for the cheap, high volume calls that keep agents running.

Some offers on this page may be paid placements or contain affiliate links.