Suvansh Sanjeev

Research @OpenAI. ex-🤖 @berkeley_ai, @CMU_Robotics.
Teamwork makes the dream work, and Dreamworks makes Shrek.

[Suvansh Sanjeev @suvansh]

GPT-5 is what you’ve been waiting for – it defines and extends the cost-intelligence frontier across model sizes today.
it’s been a long journey, and we’ve landed pivotal improvements across many axes in the whole GPT-5 family.

and hey no more model picker (by default)!

[Mike Lefko @MikeLefko]

Wins at Levi’s Stadium in 2026:

-Seahawks: 2
-USMNT: 1
-49ers: 0

[Suvansh Sanjeev @suvansh]

Boris is a thoughtful and positive-sum leader. I couldn’t have asked for a better first manager, and I can’t imagine anyone better-suited to the job of careful last-mile AI deployment in this critical period.
Congrats!! 💥

[Boris Power @BorisMPower]

1/ I’m excited to share that I’ll be taking on an additional role as Head of Research for Thrive Holdings, alongside my work at OpenAI.

[Suvansh Sanjeev @suvansh]

@distributionat careful there folks are gonna community note you with a job listing https://t.co/typ698oQU4

[Suvansh Sanjeev @suvansh]

damn dario walking his walk — word is anthro stopped hiring below L6

[Suvansh Sanjeev @suvansh]

@andrebalg gotcha! does seem there’s only more smoke since i first posted that this is true in some real, if not absolute, sense

[Suvansh Sanjeev @suvansh]

damn dario walking his walk — word is anthro stopped hiring below L6

[Suvansh Sanjeev @suvansh]

saw old colleagues now at cursor and TBD grabbing lunch the other day and i can’t decide who’s poaching whom

[Suvansh Sanjeev @suvansh]

@hersh_godse @trajectorylabs congrats hersh!! 🚀

[Sebastien Bubeck @SebastienBubeck]

We now know that with an appropriate harness both Mythos and GPT-5.5 can reproduce what our internal model did in one-shot for the unit distance problem. Clearly there is an insane overhang of capabilities with this generation of models, and no ceiling in sight for what scientific advances they can bring. You can go and try to discover new things with 5.5 right now!

[Suvansh Sanjeev @suvansh]

“sure the models are good but they aren’t creative” is the most recent iteration of human exceptionalism

creativity is high-powered search. we deem something “creative” when we can’t see the path that led to it

this proof marks a big step towards demonstrating AI “creativity”

[OpenAI @OpenAI]

Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946.

For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids.

An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better.

This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.

[Suvansh Sanjeev @suvansh]

@andrebalg ah fair enough word seems mixed then, could it be scoped to some orgs

[Suvansh Sanjeev @suvansh]

@andersonbcdefg we been knowing

https://t.co/xwMFGrKWj0

[Suvansh Sanjeev @suvansh]

there is no job AI researchers are more eager to automate than our own

[Suvansh Sanjeev @suvansh]

@philcuvin https://t.co/Uo6Ows64QW

[Suvansh Sanjeev @suvansh]

going with ‘thropic from here on out

[Suvansh Sanjeev @suvansh]

@theiyobosa trainium

[Suvansh Sanjeev @suvansh]

@benchatenu as long as AI remains an amplifier for technical employees, you’d increase the strength of your preference for hiring the most productive folks, right up until AI replaces them/us altogether

[Suvansh Sanjeev @suvansh]

if dario believes (and indeed he can’t seem to stop saying it) that humans are only useful for technical work for a short order of months this seems like a reasonable policy

why spend half the new hires’ remaining tenure training them when you can hire staff+ to race to RSI

[Suvansh Sanjeev @suvansh]

damn dario walking his walk — word is anthro stopped hiring below L6

[Suvansh Sanjeev @suvansh]

going with ‘thropic from here on out

[Suvansh Sanjeev @suvansh]

learning anthro is quite a controversial abbreviation

[Suvansh Sanjeev @suvansh]

@SebastienBubeck @tszzl damn seb, certified guyknowballogist

[Suvansh Sanjeev @suvansh]

might need a “research mode” where the models throw caution to the wind on gnarly problems and send it on a crazy approach with the understanding that it’s ok to come up short on this try

possibly equivalent to getting them to utilize more TTC effectively (better search)

[Suvansh Sanjeev @suvansh]

sorely needed long-horizon swe benchmark from @ProximalHQ

i hear mythos is too dangerous to run it on tho

[Justus Mattern @MatternJustus]

Introducing FrontierSWE, an ultra-long horizon coding benchmark.

We test agents on some of the hardest technical tasks like optimizing a video rendering library or training a model to predict the quantum properties of molecules.

Despite having 20 hours, they rarely succeed

[Suvansh Sanjeev @suvansh]

interesting finding suggesting codex is more collapsed, higher floor/lower ceiling

seems aesthetically similar to when models have to be coaxed to solve problems that are easily found to be “open”