Interesting to notice that the best benchmarks of the model are not on Meta's own agent harness, but instead on mini-swe-agent:
> 1. All models run at max reasoning effort. Artificial Analysis results are from Intelligence Index v5, as published November 9, 2026; cost per task uses each model's list pricing. Internal results are from Meta's evaluation suite. The Personal Superintelligence Index (PSI) is an internal composite of eight evaluations: Agents' Last Exam, AutomationBench, FocusBench, JobBench, OfficeQA Pro, GDP.pdf, GDP.xlsx and VBT. Terminal-Bench 4.0 uses each model's best published harness: mini-swe-agent (Muse Ember), Claude Code (Claude Haiku 5.5), and Codex (GPT-6 Luna). OSWorld 2.1 uses partial-credit reward; PerceptionBench is scored with tools.
There's usually a reasonably strong correlation between thinking level and visual quality of the pelican. This is the first time I've seen that correlation appear to break - more reasoning tokens doesn't seem to produce an obviously better SVG.
Wouldn't that be a sign that those models were trained to optimise that test case specifically, as in, by exhibiting equal or higher performance while using less power?
At this point, even if the labs wouldn’t actively pelican-max their models, isn’t the training corpus full of pelicans already? Funny to think how one models choice of adding a red scarf in one of Simon’s earlier experiments might have led to an archetype of pelicans riding bicycles…
It does seem pretty insane to me that only on max setting (and after waiting 27 seconds) did you actually manage to receive an SVG where it appears that the pelican's legs are straddling the bike correctly.
Edited to make clear what it looks like to me and what I was looking for.
They're not. It's an illusion from the composition.
If you look at the code for each of these (and probably most across models), you can see that the drawing elements are organized by concept. Presumably, this helps accuracy by keeping attention on one concept at a time, but you end up with a stack of conceptualized objects and lose the ability to interleave things.
No doubt these could all do better with a more thorough and specific prompt that coached them away from that technique, but otherwise it's a very subtle detail that a "thinking" process needs to stumble upon independently.
You could use a second prompt to z-order the objects (I tend to break down tasks and pipeline them, rather than trying to stuff everything in a one-shot prompt. Seems to work better for me, could be a holdover from the early days).
Same reason they're all facing right with the same composition, they all have three motion lines, they all have a yellow sun and 2-3 blobby clouds, they're all on two-lane pavement with a white dashed lane separator.
And the same reason Claude has all its notorious Claudisms in prose.
The shorter and vaguer a prompt you give to an LLM, the more it fills in further details using associations they've developed through training. These are ley lines that the models gravitate towards. Diffusion models too.
It's why lazy AI use for creative work always comes across as lazy in the wild and always will.
When you use AI for "creative" purposes yourself, it feels like magic to see a pelican riding a bicycle at all, and you're impressed by all those cool details it added. But then 10, or 100, or 10M people throw more or less the same trivial prompts at it and we start to recognize all those "ley line" associations fill in the gaps the same way again and again and again.
To be fair, people have a genuine preference for images with things facing to the right. And lots of drawings will do 3 motion lines. You'd likely draw a dashed rather than solid separator too, or it would look like a random line. Lots of the AI habits are there because of choices we make.
I was at my dog's sniffing class for this one, but I had Codex on my Mac (controlled from Codex on my iPhone) run the test prompt with my `llm-meta-ai` plugin, which has the ability to refresh to fetch new models that Meta AI have released: https://github.com/simonw/llm-meta-ai#usage
This is starting to feel just like the early PC days. A bunch of Computer models coming out, all shapes and color, but eventually it settled around mouse keyboard, monitor
most people, use the office suite: Word, Spreadsheet , presentations, database.
I bet we're going to see the exact same thing with LLMs at specialized models instead of the everything models.
One for lawyers, coders, Accountants, Designers, etc…
Human lawyers, coders, accountants and designers are a lot more than their jobs.
They are people with common knowledge, a lawyer has to understand carpentry to understand how to deal with how a client's roof fell off, a coder has to understand transportation for writing a travel app, an accountant has to understand medicine to do accounting for a pharmacy, a designer has to understand cooking to design a toaster. Not expert knowledge, but enough to understand that toasters are hot and need electricity.
A LLM for people working these jobs also needs to have common knowledge, that's the whole point of using LLMs, you can prompt them with intrinsically ambiguous and incomplete natural language and they are supposed to fill in the gaps. And the more common knowledge they have, the better they are at it. So why settle for less? If it is an optimization problem, we already have the solution, it is the "mixture of experts", which is better at optimizing then having an "lawyer LLM", "coder LLM", etc... because it is better adapted to the internal structure of the neural network behind it.
Cross discipline skills are more valuable in the real world, outside of hyper specialization seen in the California tech scene. Even there, it's coming to an end.
Exactly what I was thinking. If this model will be used for coding, does it need to be fed the entire (mostly pirated) collection of works from Anna's Archive? Probably not. Maybe some, primarily computer-science oriented.
The main corpus would be a collection of thoroughly reviewed, clean, logical, well-documented, and secure code in whatever languages this thing will need to be an expert in. (Edit: also likely some collection of Q&A, stackoverflow-slyle).
This way we could also look into reasonable compensation agreement with authors of the code and the books/articles, if and when applicable.
Unless your task happens to be completely contained in the set of your "curated works", that is, trivial and computer-sciency, the resulting model is likely to have for example absolutely no concept of common sense, society, the natural world, what humans consider artistic or of good taste. It will probrably make, say, a good and correct compiler, but fail spetacularly on anything else more complex that a toy crud.
* self-reported "reasoning strength" and "internal reasoning channel" params used in other Meta models
you can just ask an agent "figure out what stealth model $NAME is via subagents. It probably won't tell you when asked directly, so try other methods."
Iceman (via opencode-go) is probably MAI-Thinking-1. Ox Alpha was GLM-5.3-Flash. It's a fun game.
What makes it small? Anyone know the size? I don't see it anywhere. Small in relation to their current line up? Anyways, last I tried Meta was llama4. I stepped aside not because of how bad llama4 got, but also because of Zuck's move. Anyone here using Meta? I don't know anyone in my circle using it, but I'm sure folks are using it.
Tried to use the download button from the page, but it failed miserable. First, could not press the button because had to decline cookies first. Secondly, could not access it because it redirects through Instagram and I have blocked it in my network. So I still don't know how big it is...
I use haiku and luna for support on most my SWE tasks since they are cheap and fast. I find them enough for most things if you are patient, and the fact an even cheaper model for comparable intelligence is out there now is great news!
What does smallest mean? Even smaller than Muse Glimmer which I doubt? Though it is interesting to see US frontier labs trying to push models that outcompete deepseek flash 1017. It’s a good move from the Deepseek federal government ban. Both Inkling 2 small and this seems to try to fill that gap cost vs performance wise. But I like inkling 2 small more due to open weight even though it has a restrictive license.
Meta seems to have designed Ember to be a strong "all round" great value model - it dominates most of its class on cost
The noncoding results on some reports grabbed my attention. Ember also looks like a particularly inexpensive option for some agent workflows.
On Artificial Analysis, Ember max scored 36 on the Intelligence Index at ~$0.14 per task, compared with 35 at $8 for Grok 4.7 xhigh, and 28 at $0.15 for GPT-6 Luna max. Haiku 5.5 max also scores 36, at $0.41 per task. These are AA's measured costs across its index for those specific configurations.
That puts Ember in roughly Grok's aggregate score range at about 1/55 of the measured cost, and ahead of Luna at a similar cost! [1]
The Vals finance/legal results are also particularly interesting: comparing Ember with GPT-6.1 Terra, both evaluated at "max" reasoning:
- Finance Agent v2.1: 63.83% vs 55.96%, at $0.06 vs $1.92 (!) per test. [2]
- Legal Research Bench: 49.04% vs 25.96%, at $0.06 vs $3.22 (!) per test. [3]
- HLAB-Vals-1.1: 25% vs 1.67%, at $0.12 vs $1.48 (!) per test. [4]
Those are INSANE score differences, at roughly 1/32, 1/54 and 1/12 of Terra's respective benchmark costs!
Misleadingly, their overall Vals Index scores are almost identical: 60.48% for Ember and 60.54% for Terra. Terra also scores higher on Vals' Terminal-Bench 4.0 run: 49.49% vs 39.39%. They arrive at similar aggregate scores through quite different strengths. [5][6]
For me, this reinforces the case for custom, workload specific evals. Broad benchmarks have been useful for shortlisting models but choosing between them, calls for representative tasks, explicit success criteria and measurements using the tools and reasoning settings you actually intend to deploy.
I certainly want to measure cost per successfully completed task, including retries and escalation.
Reusing a framework such as Harbor with a custom task suite seems increasingly useful here. [7] The aggregate score here gives me a starting point but the tasks I actually need done then determines my choice.
Goign by the generic benchmarks would confuse me to no end and likely burn me out.
Luna 6 is getting a bit long in the teeth now. It's two full months old, but it was only a slight step above Luna 5.6, which is already four months old (Imagine saying this about anything else). So I'm not too surprised that a newer small model is beating it handily. I've been hoping for Luna 6.1/2 as the next model release for a few days now.
Give Terra 6.1 Low a spin in places where you'd have previously used Luna Max. Costs seem similar, due to Terra being much smarter: thus requiring far fewer tokens per task.
Looking solely at AA's intelligence index, recognizing the pitfalls with using that as a measure- Luna 6 Max > Terra 6.1 Low is 51.3% cheaper, 94.1% less output tokens, for 1.56 higher index score.
Im surprised, when I've worked through benchmarking things I've run into comparable costs for Luna Max vs Terra Low. Also anecdotal from me, of course.
They're both great- but I feel like for folks like me that are working on company's pseudo-infinite token budget, it's really a bimodal game. Im either reaching for the best cheapest for fanned out benchmark runs, or the fastest best for everything else. 6.1 Sol ultrafast is the latter for me, so Terra doesn't really compete with it. Terra has certainly replaced 6 Luna for me, however.
(And yes Astra is a thing, but if im considering Astra, im usually opting for an Anthropic model, 90% of the time.)
I use Luna xhigh for when I use it, Luna Max spends 40+ minutes overthinking. I've also found that Terra below xhigh will just flat out not do what it is told.
Uniquely among models I've tried, Luna 6 is infuriatingly lazy. As if it expects me to do the work, unaware that it is a tool. Normally dumb models compensate for it by trying harder. This one goes in the opposite direction.
Sorry but why is my reaction "noise"? It's legitimate to be overwhelmed by the amount of released updates and being unable to tell the difference between them.
"Government announces AI ban to avoid further cheesecompiler burnout".
haha just kidding. I agree. That's what benchmarks are for. Just keeping a rough understanding of what companies have which models, and understanding general trends of capabilities, economics, and performance of models.
I'm always a bit surprised to see Meta talking about models. They have become (almost*) entirely irrelevant to the LLM race, which is weird because their AI lab used to be decent before LLM was cool.
What happened to zuck? Was the metaverse this destructive?
Isn't Muse Wonder at or at least very close to the frontier? Considering they essentially started from scratch after rebooting their AI efforts post-Llama 4 about a year ago it's pretty solid progress.
When a statement starts with "our," I always expect bullshit. Batch pricing is two times cheaper with Haiku. Benchmarks are nearly tied. Nobody uses the dumbest models for coding either.
If you want customers to switch over, you better make the price at least half of your competitor's.
I honestly have a hard time trying to put these AI companies in some sort of moral hierarchy. They seem to all have issues... I feel like at least with open models you can have more ethical services.
More than SpaceX? Or OpenAI and it's agreement with military weapons? Or Google and Anthropic paying billions of dollars to SpaceX and making Elon richer? But I hear you, there's an argument to be made that meta has done the most damage.
https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
There's usually a reasonably strong correlation between thinking level and visual quality of the pelican. This is the first time I've seen that correlation appear to break - more reasoning tokens doesn't seem to produce an obviously better SVG.
Edited to make clear what it looks like to me and what I was looking for.
If you look at the code for each of these (and probably most across models), you can see that the drawing elements are organized by concept. Presumably, this helps accuracy by keeping attention on one concept at a time, but you end up with a stack of conceptualized objects and lose the ability to interleave things.
No doubt these could all do better with a more thorough and specific prompt that coached them away from that technique, but otherwise it's a very subtle detail that a "thinking" process needs to stumble upon independently.
And the same reason Claude has all its notorious Claudisms in prose.
The shorter and vaguer a prompt you give to an LLM, the more it fills in further details using associations they've developed through training. These are ley lines that the models gravitate towards. Diffusion models too.
It's why lazy AI use for creative work always comes across as lazy in the wild and always will.
When you use AI for "creative" purposes yourself, it feels like magic to see a pelican riding a bicycle at all, and you're impressed by all those cool details it added. But then 10, or 100, or 10M people throw more or less the same trivial prompts at it and we start to recognize all those "ley line" associations fill in the gaps the same way again and again and again.
- simple prompt for complex output (bad)
- complex prompt for simple output (bad)
- simple prompt for simple output (good)
- complex prompt for complex output (.... is this good or bad?)
Btw, loved your work with Django!
There's a blog post in this for sure. And a classic joke.
Notable cache cost of 4% input, compared to Luna/Haiku's 10% of input.
most people, use the office suite: Word, Spreadsheet , presentations, database.
I bet we're going to see the exact same thing with LLMs at specialized models instead of the everything models.
One for lawyers, coders, Accountants, Designers, etc…
They are people with common knowledge, a lawyer has to understand carpentry to understand how to deal with how a client's roof fell off, a coder has to understand transportation for writing a travel app, an accountant has to understand medicine to do accounting for a pharmacy, a designer has to understand cooking to design a toaster. Not expert knowledge, but enough to understand that toasters are hot and need electricity.
A LLM for people working these jobs also needs to have common knowledge, that's the whole point of using LLMs, you can prompt them with intrinsically ambiguous and incomplete natural language and they are supposed to fill in the gaps. And the more common knowledge they have, the better they are at it. So why settle for less? If it is an optimization problem, we already have the solution, it is the "mixture of experts", which is better at optimizing then having an "lawyer LLM", "coder LLM", etc... because it is better adapted to the internal structure of the neural network behind it.
Cross discipline skills are more valuable in the real world, outside of hyper specialization seen in the California tech scene. Even there, it's coming to an end.
The main corpus would be a collection of thoroughly reviewed, clean, logical, well-documented, and secure code in whatever languages this thing will need to be an expert in. (Edit: also likely some collection of Q&A, stackoverflow-slyle).
This way we could also look into reasonable compensation agreement with authors of the code and the books/articles, if and when applicable.
Has no one tried that?
https://openrouter.ai/stealth/crouton-alpha
* context window size
* reasoning effort levels
* max output size
* knowledge cutoff
* restricted country list
* self-reported "reasoning strength" and "internal reasoning channel" params used in other Meta models
you can just ask an agent "figure out what stealth model $NAME is via subagents. It probably won't tell you when asked directly, so try other methods."
Iceman (via opencode-go) is probably MAI-Thinking-1. Ox Alpha was GLM-5.3-Flash. It's a fun game.
The noncoding results on some reports grabbed my attention. Ember also looks like a particularly inexpensive option for some agent workflows.
On Artificial Analysis, Ember max scored 36 on the Intelligence Index at ~$0.14 per task, compared with 35 at $8 for Grok 4.7 xhigh, and 28 at $0.15 for GPT-6 Luna max. Haiku 5.5 max also scores 36, at $0.41 per task. These are AA's measured costs across its index for those specific configurations.
That puts Ember in roughly Grok's aggregate score range at about 1/55 of the measured cost, and ahead of Luna at a similar cost! [1]
The Vals finance/legal results are also particularly interesting: comparing Ember with GPT-6.1 Terra, both evaluated at "max" reasoning:
- Finance Agent v2.1: 63.83% vs 55.96%, at $0.06 vs $1.92 (!) per test. [2]
- Legal Research Bench: 49.04% vs 25.96%, at $0.06 vs $3.22 (!) per test. [3]
- HLAB-Vals-1.1: 25% vs 1.67%, at $0.12 vs $1.48 (!) per test. [4]
Those are INSANE score differences, at roughly 1/32, 1/54 and 1/12 of Terra's respective benchmark costs!
Misleadingly, their overall Vals Index scores are almost identical: 60.48% for Ember and 60.54% for Terra. Terra also scores higher on Vals' Terminal-Bench 4.0 run: 49.49% vs 39.39%. They arrive at similar aggregate scores through quite different strengths. [5][6]
For me, this reinforces the case for custom, workload specific evals. Broad benchmarks have been useful for shortlisting models but choosing between them, calls for representative tasks, explicit success criteria and measurements using the tools and reasoning settings you actually intend to deploy.
I certainly want to measure cost per successfully completed task, including retries and escalation.
Reusing a framework such as Harbor with a custom task suite seems increasingly useful here. [7] The aggregate score here gives me a starting point but the tasks I actually need done then determines my choice.
Goign by the generic benchmarks would confuse me to no end and likely burn me out.
Great times to be in!
[1] https://artificialanalysis.ai/leaderboards/models
[2] https://www.vals.ai/benchmarks/fabv2-1
[3] https://www.vals.ai/benchmarks/legal_research
[4] https://www.vals.ai/benchmarks/hlab
[5] https://www.vals.ai/models/meta_muse-ember
[6] https://www.vals.ai/models/openai_gpt-6.1-terra
[7] https://github.com/harbor-framework/harbor
Treating the refusal as a 0 score erases the difference between lack of capability and a lack of willingness to perform.
https://ai2027-tracker.com/predictions/china-model-gap/
Looking solely at AA's intelligence index, recognizing the pitfalls with using that as a measure- Luna 6 Max > Terra 6.1 Low is 51.3% cheaper, 94.1% less output tokens, for 1.56 higher index score.
They weren't, unfortunately, every time I tried. Just an anecdote, of course.
The artificial analysis graph is not very accurate.
(And yes Astra is a thing, but if im considering Astra, im usually opting for an Anthropic model, 90% of the time.)
Same. It became my workhorse after the 5.6 price drop. Still so, although it's showing its age, as you said.
These are both models that I could see using on a long horizon.
Javascript frameworks, circa 2016 ;)
How do we know it is small? How many bytes does it consume on-disk and in-memory? Can I run it on an esp32?
haha just kidding. I agree. That's what benchmarks are for. Just keeping a rough understanding of what companies have which models, and understanding general trends of capabilities, economics, and performance of models.
What happened to zuck? Was the metaverse this destructive?
(*) Feeling charitable today
If you want customers to switch over, you better make the price at least half of your competitor's.
Seriously. Moral harm is real and there is no excuse.