@thsottiaux (tibo) might follow me if i hit 10k. help a brother out. "single handedly delayed image AGI" follow, so the voices don't haunt you
DM for collabs
Joined June 2026
- Tweets10.1K
- Following367
- Followers4.3K
- Likes28.2K
openai is not done.
one good model is all it takes.
I think OpenAI was never behind, and people are so wrong.
Look at tokens-per-task stats, Claude models need many times more tokens for similar results, and OpenAI seems to be the only lab that really cracked efficiency.
Also, they are the only lab with the full stack, they have: image models (best), text models (almost best), and voice models (imo, best).
Its AI, everything can change in a day here, i don't think they're in a bad position at all. Just wait for Astra 6.1
anthropic could be heading for the biggest IPO in history.
if fable 5.5 gets anything close to the reception opus 5.5 did, i think there's a real chance its IPO could value the company at $2T.
that would put it ahead of spaceX's record breaking IPO.
especially with how much momentum anthropic has right now, i wouldn't be surprised.
it's funny to think we actually thought gpt-6.1 terra might leapfrog 6.1 sol.
with how good opus 5.5 is, and openai releasing a new model in the middle of all the competation, we were hoping for better than sol performance at half the price.
that would've been an insane release.
looking at terra now, we really got ahead of overselves lmao.
LLMs love japan.
researchers tested 8 LLMs with cultural questions across 24 languages, letting them choose any country for their answers.
and for some reason, the models kept gravitating toward japan.
but the most interesting part is that this preference wasn't nearly as strong in base models. it became much more pronounced after supervised fine-tuning.
gpt-6.1 terra doesn't look particularly impressive on terminal-bench 4.0 compared to sol.
terra tops out at around 50%, compared to 56% for 6.1 sol.
and sol manages to score higher at a lower cost per task.
so even though terra is 50% cheaper per token, that doesn't necessarily make it cheaper for agent workloads.
but it's a huge upgrade over 5.6 terra.
these don't look real at all.
you can easily tell they're fake.
Today @RealbotixCorp is live with its 1st commercial robot line.
Check out our new website Realbotix.com and reserve yours today.
the sentiment around gpt models isn't great right now, especially with how good opus 5.5 is.
a lot of that frustration is spilling over into how people view openai's entire ecosystem, not just its models.
openai's products aren't helping either. the constant bugs, reliability issues, and frustrations across chatgpt, codex, and even dots are only making things worse.