Two stealth models just appeared on the leaderboard, both ahead of Fable 5.2
contralabs.comI don't understand, where are those models available in stealth mode? Usually people do this on openrouter but I couldn't find any model called "gestalt". Does anyone know or is this just black hat marketing for this contralabs platform, and the model does not even exist?
Openrouter will have these from time to time
Sheesh if I was wrong you could just tell me! I'm not trying to shill openrouter, I've just seen stealth models there before.
To clarify, it is not two models. Its a single model (Gestalt). The other bar on the benchmark is a human created dataset. Suffice to say we didn't expect anyone to beat it, but here we are. And no, it is not Contra model.
Imagine the Contra team discovering this while compiling the rating data, they might be so shocked that their eyes nearly pop out!
It is possible that anthropic/open ai may no longer release their best models due to their massive size and cost of running, instead releasing distilled version for public use.
Google/Gemini just went to a Flash/Standard/Pro model where Standard is $20 a month and Pro is $100 a month. I would surmise there's no way their top model can be given out for that price.
its more possible the IPO signals the death of any significant training runs and the rest of the operations will be goto:
1. price discrimination.
2. inference discrimination
3. national security extortion racket.
Glm 5.5 I bet, but still too early to tell
I’d also guess that, considering their be massive approach to sourcing training data, it would make sense that it’s much less sloppified than anything before it
I believe the next version is going to be glm-6.0 but yeah that is my guess as well.
This Thanksgiving period will be a cambrian explosion of capable models.
Deepseek V5 trained on Huawei Ascend?
Yes, Ascend 950pr or dt.
Very curious to find out which model that is :O
The name is probably a clue. I wonder if someone just bolted Fable to Astra in a loop? I do that when I'm stuck, and get great results that exceed what either can do alone... Though it is wickedly expensive.
How do you have them interacting in a loop? Plan-execute-repeat? Execute-review-repeat?
Usually I do it when one of them is stumped on their own, or seems to be thrashing. I guess it's more like... adjacent to the main loop.
I'll spin up both TUI's in herdr, tell them both to run "hc --skill" and then talk to each other until they figure out a solution. Also, for really complex tasks I will give one of them the title "Advisor" and just leave it running in the background while the main agent runs. The main agent gets instructions to ask the Advisor when stuck. So far it's basically always worked.
Before herdr I had them use tmux and files to talk to each other. That works too, but with a little more fumbling.
My money is on GLM 6
The name gestalt indicates it’s probably a router model like paretto by unbiased
interesting! any model performing better than human created outputs is very unexpected. i wonder how fast these models are gonna get better, thinking about the progress we made in a year on coding, design would be soon too
This is a clever bit of content marketing
That web site is insufferable.
I normally wouldn't even bother commenting on an insufferable site (just close the tab and move on), but this was so bad that I came to the comments looking for validation of my feelings.
Like, above the fold is a bunch of moving links to AI-generated pages that have apparently nothing to do with the page title. Perhaps it's relevant, but I find it non-obvious. The first one I clicked gave a bad gateway.
Then we get the leaderboard, then just more AI-generated projects with opaquely dense descriptions like "Real client briefs, each with its golden: the design set beside the models’ outputs as the bar.".
Just reading this makes me feel like I'm having a stroke.
</rant>
Thanks for confirming the site is insufferable.
I closed it before it could finish loading because it was so slow. Which does speak to its insufferable nature.
"Random AI lab makes up a new benchmark that puts at least one of its models ahead of Fable 5.2"
Yeah this is 100% an ad
Saying this when none of the gallery examples show the new model's output is the ultimate "trust me, bro" move.
Perhaps soon we will know. I think on of the reasons for such stealth benchmarks is testing the model before uncovering details
the benefit of the doubt some of you give to everyone else but just swallow anything anthropic/oai has to say is insufferable
who is --John Galt-- Gestalt?
Why would this not be GPT 6.2 Astra?
"on the leaderboard [that we created so we could lead it]"
It is interesting that the Kimi model gets a very high “slop score” but “professional critiques” gave it a pretty high score. Do critiques have no taste or the “slop” is not slop anymore?
Producing a work that looks and feels great but with a mess under the covers is a thing.
If slop score is mostly “looks like AI”, Kimi scoring high there while human designers rate it well isn’t that weird. Generic can still nail the brief.
some obscure design benchmark