Muse Ember Is Cheap. The Pelican Test Told Me Something Else.

Simon Willison ran a pelican test on it. That is the first thing I saw about Muse Ember, before I saw a single benchmark number.

He asked the model to draw a pelican riding a bicycle at all six reasoning levels. It returned five pelicans. The high setting did not produce a complete SVG at all.

Then he said something that stuck with me. The pelican test usually tracks reasoning effort. More thinking, better pelican.

This time the correlation broke. minimal came back with about ten reasoning tokens and, by several readers' judgment, the best bird of the set.

Meta Superintelligence Labs shipped its smallest model on November 9. The top comment thread on Hacker News was about whether the pelican had a fish in its basket.

I read 139 comments. Maybe six were about the pricing.

What they actually shipped

Muse Ember is a small multimodal model built for the boring part of agents. Tool calls, routing, extraction, coding subtasks, long document parsing. The stuff an agent does four hundred times while you watch a spinner.

The pricing is the headline, and it is genuinely low:

• $0.10 per million input tokens
• $0.004 per million cached input tokens
• $0.50 per million output tokens

That cache number is the one people keep missing. It is 4% of input. Luna and Haiku charge about 10%.

In agent loops, cached input is most of the bill.

Meta's pitch is simple. A big model plans, this one does the volume. Their line: Muse Wonder plans, Muse Ember handles the volume.

The benchmark story, and the number that does not match

Artificial Analysis puts it at 36 on the Intelligence Index for $0.14 per task. Claude Haiku 5.5 hits the same 36 and costs $0.41. That is nearly three times more for the same score.

And it beats DeepSeek V4.1 Flash on cost, which is the part that made people sit up. DeepSeek was supposed to own cheap.

But here is the thing I could not stop poking at. Meta's own launch page lists AutomationBench-AA at 37.8% for Ember. Artificial Analysis wrote 68.1% for the same benchmark in the same week.

I do not know which test setup produced which number. Neither page explains the gap.

That is not a scandal. Benchmarks get re-run, versions drift, runners differ.

But it does mean you should not quote either figure like it is carved in stone. I have been burned by this exact thing before.

"Smallest" is doing a lot of work

Everyone asked the same question in different words. Small compared to what? How many parameters?

Can I run it?

The answer is no. There are no weights. Muse Ember is API only, muse-ember-1.0, on Meta's Model API and Muse Code.

One commenter put it better than I could: what is the point of bragging that it is small if it always runs on someone else's datacenter?

Fair. Though the "small" claim is not really about your laptop. It is a signal that this competes with Haiku and Luna, not with Wonder.

Meta is telling you where to slot it in your agent stack.

The open weights crowd is still annoyed. Muse Spark Open shipped at 816B parameters, which is not a thing most people can run either. And the open version came with extra safety training that several users described as being lobotomized into announcing it is Muse Spark regardless of the system prompt.

There is a running joke in that thread about whether the open releases are distills wearing a safety hat. Nobody proved it. But nobody disproved it either, and Meta has not published the parameter count for anything closed.

So the pattern is clear. Meta releases the closed good one, then a heavier, more restricted open one later. Ember fits the pattern.

The trick that keeps it honest

Here is the detail I actually respect.

On AA-Omniscience, Ember scores 21.1% accuracy. That is the lowest of its comparison group. Haiku gets 34.7%.

But it declines to answer 64% of the questions. And that keeps its hallucination rate at 17.4%, half of Haiku's.

So the model is not knowledgeable. It just knows when to shut up.

I have wanted this from every model for two years. A model that says "i don't know" instead of confidently inventing a citation is worth more to me than one that scores twelve points higher and makes things up. For an agent doing extraction at volume, a wrong value in a database costs real money.

It is also the least glamorous thing a model can do. Declining to answer looks like weakness on a scoreboard. In production it looks like not paging anyone at 2am.

The catch is obvious. If you need it to know things, this is not your model. It is a doer, not an encyclopedia.

The pelican thing is not a joke

I keep coming back to the birds.

Every model now draws the same pelican. Same red scarf. Same fish in a front basket.

Same three motion lines, same yellow sun, same dashed lane markings. One commenter called them ley lines, the paths a model walks when you leave a prompt vague.

That is not really about Muse Ember. It is about all of them. A short prompt makes the model fill the gaps from training associations, and the associations are now shared across every lab.

Which means the pelican test stopped measuring reasoning a while ago. It measures how much a model is willing to wander. And when a launch thread spends its energy there instead of on the pricing, that tells you something about what people actually do with these things all day.

The honest part

Most people do not need this model. If you are building a chatbot that answers twenty questions an hour, the price difference is noise. Use whatever you already have wired up.

And the benchmaxxed complaint is real. One user who had actually used the Muse line said the models read as frontier on bench cards but feel less useful in real work.

Another said Spark was their daily driver for weeks and got quietly outclassed. These are people with opinions formed by using the thing, not reading about it.

Also, no batch API at launch. One commenter's entire review was three words: "No batch. Pass."

If your workload is bulk and offline, that matters more than any index score.

Where this lands

The pelican with the fish is the one I remember. Not the 36.

Meta built a model that is cheap, quiet when unsure, and good at the repetitive middle of agent work. That is a real product decision, and a smarter one than chasing the top of the leaderboard.

But the launch told me something else too. We have gotten so used to benchmark charts that a genuinely low price barely registers, while a bird on a bicycle gets 300 upvotes.

Next time i run an agent loop that burns forty thousand tokens on nothing, i will think about that pelican. And i will check the cache price before i check the index.