Meta's Muse Ember Lists at the Same Price as Claude Haiku 5.5. The Savings Are in the Tokens It Doesn't Use, and the Accuracy Gap Is Real.
Meta released Muse Ember on November 9, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K. Claude Haiku 5.5 lists at the same $0.10 and $0.50 up to 100K. So the headline that Ember costs a third as much is not about the price list. It is about how many tokens each model burns to finish a task, and that is a claim worth testing before you rewire anything.
What the numbers say
Artificial Analysis scores both models at 36 on its Intelligence Index at maximum effort. Ember costs about $0.14 per task on their test and Haiku 5.5 costs about $0.41, roughly a 3x gap. Artificial Analysis says Ember uses about a third as many output tokens, which explains it. Same price per token, far fewer tokens.
Meta's own blog adds agent-focused results. On its Coding Agent Index, Ember in Muse Code scores 46 against 37 for Haiku 5.5 in Claude Code and 31 for GPT-6 Luna in Codex, all at max effort. Meta also reports 37.8% on AutomationBench-AA against 13.9% for Haiku 5.5 and 20.2% for Luna. Artificial Analysis' own write-up gives Ember a different AutomationBench figure of 68.1%, and I could not tell whether that is a different scoring mode or a mismatch. When two sources disagree on a number, I treat both as marketing until I run my own.
Specs from Meta's blog: a 1M-token context window, up to 512K tokens of output, and input across text, image, audio and video. Prompts over 200K cost $0.30 input and $1.50 output. These come only from Meta. I found no second source.
Where Ember is weaker
This is the part the cost headline skips. On Artificial Analysis' Omniscience test, Ember's accuracy is 21.1% against 34.7% for Haiku 5.5. It declines to answer about 64% of questions, and its hallucination rate is 17.4%, about half of Haiku's. So it is more cautious, not more knowledgeable. On Meta's FrontierCode 2.0 table, Ember scores 27.2 against Haiku's 34.9. Meta's table and Artificial Analysis agree that Ember trails there.
For an indie builder, the profile reads like this: good at tool calls, routing and extraction loops where the answer is in the context, worse at answering from memory.
Things I could not confirm
Meta's pages I read do not state a license, and nothing says Ember is open-weight. It is available through the Meta Model API and Muse Code, with a trial at dev.meta.ai. One chart caption in Meta's blog says costs are estimates at "proposed pricing" while a table says list pricing, so I would check the live price before budgeting on it. I read both Meta's blog and the Artificial Analysis article through a summarizing tool, so I would verify any figure you plan to build on against the pages themselves.
What I'd actually do
If I run an agent that makes dozens of small tool-call decisions per job, I would route those subtasks to Ember and keep a stronger model for the steps that need recall or judgment. The test is cheap. Take 30 real subtask logs, run Ember and Haiku 5.5, and measure pass rate and total tokens. If Ember passes at least as often and uses a third of the tokens, you have a real saving. If its pass rate drops on tasks that need facts not in the prompt, you have found the boundary.
I would not move a customer-facing answer step onto it. A model that declines two-thirds of factual questions will produce visible refusals, and your users will not read that as caution.
Where this could be wrong
Per-task cost depends on the test harness, and Artificial Analysis says Haiku 5.5's score and cost are measured at max effort. If you run Haiku at a lower effort, the gap may shrink. And since the pricing language in Meta's blog is inconsistent, the saving could change when pricing is finalized.
Author
Lukas
@lukcombinator