AI Judge changed title after evaluation, original title: "OpenAI launches GPT-5.6 Sol, its most capable model for long-horizon cybersecurity and coding tasks"
OpenAI has released GPT-5.6 Sol in limited preview as its strongest model yet for extended agentic work, with standout results on vulnerability research, exploitation chains, and other long-horizon cybersecurity jobs plus notable gains in coding speed.
GPT-5.6 is incredibly strong and fast for coding. I hope we can make it available to everyone soon.
GPT-5.6 Sol is our most capable model yet for cybersecurity.
It shifts the performance-efficiency frontier for long-horizon security tasks including vulnerability research and exploitation.
o1 co-creator Noam Brown highlighted the model's unusual strength and velocity on coding work, framing it as a practical step forward rather than a distant promise.
Access runs only through the API for roughly twenty government-approved organizations right now, with broader rollout expected in coming weeks after staggered security checks.
Positive users praise GPT-5.6 Sol's cybersecurity advances and urge quicker public access, while negative users accuse OpenAI of regulatory capture and criticize announcements of models that remain unavailable.
No Digg Deeper questions have been answered for this story yet.
From their posts, it seems that the concern that the USG has is cybersecurity. Which is totally reasonable.
I wonder if they'll notice, though
GPT‑5.6 Sol launches with our most robust safety stack yet.
We strengthened real-time protections against high-risk cyber activity and repeated misuse, then spent weeks hardening the system with human red teaming and over 700,000 A100-equivalent GPU hours of automated testing.
This is what happens when you finally unite the Noams
GPT-5.6 is incredibly strong and fast for coding. I hope we can make it available to everyone soon.
when the score gets to mythos level but cheaper: alright fellas lets stop
GPT-5.6 Sol is our most capable model yet for cybersecurity.
It shifts the performance-efficiency frontier for long-horizon security tasks including vulnerability research and exploitation.
like that line there, there is no reason for it to stop, right? :P
when the score gets to mythos level but cheaper: alright fellas lets stop
Compute multipliers, being discovered, one after another. To make it not do this is to not make it available. What use is the frontier if not to make the frontier?
How much work is not directly related to compute multipliers for frontier AI, yet tautological to the very methods? Are you going to ban mathematics? At what intelligence does the user become dangerous?
Welcome to Mr. Bones' wild ride
@polynoamial I was hoping to see some test-time budgets as footnotes in the blog post 😄
@polynoamial Hope is doing lot of heavy lifting bruh... Though I feel bad for you guys chaught in the crossfire due to safetiest/yudoskovian bs
@yacineMTB @howardlutnick there is mythos class model and better than mythos
@polynoamial Is there any mention of swe bench pro scores? I couldn’t find any in the system card…
@polynoamial great now i can generate bugs i dont understand at 400 tokens per second
@graykevinb @yacineMTB
@yacineMTB why are they hyping up its cyber abilities then? They are feeding themselves to the dragon.
@yacineMTB If these guys were smart they would release these amazing models and just be like "dang they aren't as good as the benchmarks previously but try em out anyhow" and totally lie about it while the models actually beat historic benchmarks
and then public gets to use em
@runequantum @yacineMTB yeah. They also are feeding the dragon.
Ohh wait a sec they only make a profit from enterprise...
They just used us for token generation to feed their dataset.
so they'll abandon consumers
Regulatory capture in action. What secret benchmark or “safety” bar is actually blocking these releases? Are we really heading toward a future where every strong model needs government minders? Capabilities are already jaw-dropping. In a year they’ll be godlike. Pretending we can centrally approve this is pure delusion
@polynoamial Open source it. Embrace your destiny.
@polynoamial Any thoughts for coding specific model? I mean I get it hard to sperate cyber security from it cause.. well its finding broken/bad code and exploiting.. but maybe can be shipped cheaper or more widely if it doesnt have to worry about fluff from MoE?
@polynoamial Then do it.
