Meta's Muse Ember matches Claude Haiku 5.5 at a third of the cost, at $0.10/$0.50 per million tokens
- Meta released Muse Ember, its smallest model, on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens, aimed at tool calls, routing, extraction, coding subtasks and long document parsing.
- On the Artificial Analysis Intelligence Index versus cost per task, Muse Ember scores 33.9 at $0.116 per task, near Claude Haiku 5.5's best of 34.3 at $0.248, which Meta describes as matching Haiku at about a third of the cost.
- It beats GPT-6 Luna at every reasoning setting; Luna's top point is 27.8 at $0.152 per task.
- Headline benchmark gaps: AutomationBench 37.8 against 13.9 for Haiku 5.5 and 20.2 for Luna, Agents' Last Exam 61.6, Terminal-Bench 4.0 at 40.9, but it loses FrontierCode 2.0 to Haiku, 27.2 against 34.9.
- The model has six reasoning effort levels from minimal to max; Box, Cline, Oracle and Cloudflare tested it early, and Cloudflare reported it ran up to 30% faster.
Hacker News opinions
Six effort levels, so someone ran the pelican SVG test on all of them. The weird part is the correlation broke: more reasoning tokens didn't produce a better pelican, and minimal at like 10 tokens looked as good as max.
Only on max, after 27 seconds of waiting, did the pelican's legs actually straddle the bike correctly. That's a long wait for a bird on a bicycle.
A single SVG is not a benchmark. It's fun to look at but it tells you nothing about what the model can do.
Pretty sure this is the crouton alpha stealth model from OpenRouter. I used it a lot: starts strong, gets rough around turn 10, unusable by turn 20. Fine for short tasks, bad for long ones.
Luna 6 is two months old and a newer small model is already beating it handily. It was barely a step above 5.6 anyway, so I've been waiting on a 6.1 for days now.
Luna 6 is also infuriatingly lazy, like it expects me to do the work. Dumb models usually compensate by trying harder; this one goes the other direction.
Try Terra 6.1 Low where you'd have used Luna Max. Luna 6 Max is 51.3% cheaper, uses 94.1% fewer output tokens, and scores 1.56 higher on the index.
Meta talking about models still surprises me. They went basically irrelevant after Llama 4, hired new people and reorganized, and now Muse Wonder is at least close to the frontier.
Still frustrating that I can't give Muse Wonder Full a spin.
Refusals should score negative, not zero. Treating a refusal as a 0 erases the difference between lacking the capability and refusing to do it.
No weights, so no local use. Shame, I'd love to run this myself while they're still catching up. They did put out Muse Spark Open for exactly that.
Feels like the early PC days. A pile of models in every shape and color, then it settles on a few specialized ones for lawyers, coders, accountants, designers.
Disagree on specialization. A coder writing a travel app needs to understand transportation, a lawyer needs to understand how a roof fell off. Mixture of experts already handles that better than a lawyer-only LLM.
I use Haiku and Luna for SWE support because they're cheap and fast, and they're enough if you're patient. Something cheaper at comparable intelligence is good news.
Why is there no comparison with DeepSeek?