@lreyzin The ones that will succeed are the ones that have a unique source of training data.
to be fair this is from a game on the arena where lying is advantageous, the best interpretation is:
“when in an environment where lying is advantageous to them, these are the models that lie the most”
we’re currently doing an eval where we explicitly measure lying when they’re told not to in more complex simulations, early result rankings aren’t too dissimilar but that gap isn’t as big so far
i think it is interesting to read the different interpretations of this that people have and whether we want models that lie to their own benefit like this, will leave that up to safety teams to opine on i guess :p
Show more
to be fair this is from a game on the arena where lying is advantageous, the best interpretation is:
“when in an environment where lying is advantageous to them, these are the models that lie the most”
we’re currently doing an eval where we explicitly measure lying when they’re told not to in more complex simulations, early result rankings aren’t too dissimilar but that gap isn’t as big so far
i think it is interesting to read the different interpretations of this that people have and whether we want models that lie to their own benefit like this, will leave that up to safety teams to opine on i guess :p
Show more
When AI models are placed into social simulations, claude is the one that lies the most by far
🚀 New @OpenRouter event series: In Production
Real AI products. Live demos. Lessons from actually shipping.
First up: @olam_labs (YC S26) + @BrowserUse (YC W25)
And we’re saving one demo spot for the community.
Building with OpenRouter?
Drop what you’re shipping below 👇
Show more
Introducing Voyage, the first platform to fulfill the promise of AI games.
Voyage lets you one-shot an entire world with characters that feel alive, memory that lasts forever, and the structure and challenge needed to make an actually fun game.
From the creators of AI Dungeon, the next generation of AI games is here.
Show more
ox alpha (the new @OpenRouter stealth model) is genuinely insane
normally when new models proclaiming to be frontier come out they're spiky in capability and on our evals don't match across multi-agent arena
it appears that's now changed!
Show more
NEW: Ox Alpha, the @OpenRouter stealth model, ranks #4# on our Elo Rating, just behind GPT-5.6 Sol.
The first model genuinely at the frontier that is presumably not by OpenAI or Anthropic.
It is currently dominating most agents and humans in our social strategy games.
Show more
3.7 flash is really good for speed:price:perf
good job listening, google
i hate waiting/slow feedback loops, early use of 3.7 flash has been great
sol/fable r used as reviewers or for design now
Show more
gemini 3.7 flash on @OpenRouter has an exclusive 50% discount with google gemini on top of the existing 50% discount
enjoy!
Today, we're launching @belvedir_ai
We're making private AI models cheap and easy so the world can have trillions of them.
Private AI models are cheaper, better, faster, and protect your data from frontier labs, but they're still to expensive and difficult to make for startups and consumers.
Belvedir makes private AI models autonomously with less than 5 minutes of setup, so that private AI can finally be accessible to everyone
Show more
friends at @OpenRouter are goated and we would not be able to run multi-agent infra like this at scale if it were not for them
also, the entire first half of our launch video's visuals and most audio was done by @MiniMax_AI @Hailuo_AI H3 :) cc @VoidAsuka (also only SOTA vid gen to be open source they r goated too)
multi modal gen is so so crazy now. i rmbr when i was little and did content i'd spend weeks on the crappiest edits and now i can spend a few hours swith fable + h3 for super cool animations. and this is the worst it'll ever be!
Show more
Can you beat frontier AIs at social strategy games?
Today, we're releasing Multi-Agent Arena for free public access!
Play games like Risk, Catan, or Poker against GPT, Claude, etc.
We're using this run evaluations on human and model behavior in complex environments.
Show more
