We wrapped up this year's competition circuit with a full score on the ICPC, after achieving 6th in the IOI, a gold medal at the IMO, and 2nd in the AtCoder Heuristic contest!

When Mythos came out, my immediate thought was "if our models can prove 80-year-old theorems, surely they can find cyber vulnerabilities too." And they did. I imagine the researchers there are thinking the same thought in reverse.

The big story here is that GPT 5.5 (high/xhigh) outperforms claude-opus-4.8 (max/xhigh) by 20.7% succeeding on 12 additional tasks! More impressive: GPT is roughly half the cost and twice as fast. OpenAI is back in the game. Overall, this competition is healthy for the industry. I'd love to see a third player rise to the top of the leaderboard!

nothing like switching to claude for a few days to try out a new model and going back to codex xhigh to remind you how much better 5.5 is right now it's really not close

Very proud that an OpenAI model disproved Erdős’s longstanding unit distance conjecture, with an elegant and intricate proof that brings sophisticated ideas from algebraic number theory to bear on geometry. For whatever reason, mathematics has been the field most amenable to research breakthroughs with AI. I consider it lucky that it was mathematics after all - a field where experts have been willing to engage deeply with us, and with proofs generated by our models. I'm grateful for that, and don't take it for granted. Math is an artistic endeavor, and perhaps for artists, it is precisely their appreciation for art that saves them from the possibly grotesque feeling of a machine producing it. Our goal is not to replace humans. We aim to chart a path forward where humans continue to have a significant role to play, even as we build exceptionally powerful AI. I am excited to use math as a domain to explore these paths, and @SebastienBubeck, @merettm, and I are excited to engage with the broader mathematical community to chart them together. Please reach out if you are interested! I'm optimistic this will help us navigate how AI impacts society in domains like coding and general co-working.

This is just one eval, but it's an important one - UK AISI’s cyber range tests long-horizon, agentic capability. 5.5 performs similarly to Mythos. The risks for frontier models are real. But we do our best to deploy AI people can actually use - through hard work on mitigations.

When GPT-5.5 misses on a Frontier Math question

We tried a new thing with NVIDIA to roll out Codex across a whole company and it was awesome to see it work. Let us know if you'd like to do it at your company!

And he’s doing a phenomenal job!

why isn't chatgpt the perfect personal AGI? what is most disappointing about it? what feature, model improvement, or bugfix would do the most to make it more useful in your daily life? what is most frustrates you that chatty can't do, or can't do well enough?

Not true. Science is more important than ever. The future can’t be about dumping results on the community en masse. We need to work with scientists to use AI to accelerate discovery without stripping away the artistry. Excited for @SebastienBubeck and @ahelkky (who are both amazing scientists) to take on this mandate!

Copy of FreeBSD from Jan 1 2026 (68d6abd9714384a41028dc0d5086b4930366bbea), then prompted GPT-5.4 with a similar prompting strategy to the Mythos red team harness from their whitepaper, via OpenCode. This reproduces. Going to attempting reproduction for the other bugs they disclosed. Concerning; maybe the only insight from the Mythos whitepaper is that they were willing to spend millions on compute to do this for a bunch of open source. But they could have saved millions by just using Opus; Mythos had little to do with it.

At @OpenAI, Chief Scientist @merettm helps lead the research roadmap to AGI including a research intern-level AI system by September 2026 and a fully automated AI researcher by March 2028. I sat down with Jakub to check on those timelines and ask him all of my top-of-mind AI questions including: ▪️ How OpenAI thinks about extending RL beyond code and math ▪️ The current state of alignment research as more powerful models loom ▪️ The future of continual learning ▪️ How startups should think about building their own models/harnesses And he also shared some great stories around OpenAI’s pioneering work on math. YouTube: youtu.be/vK1qEF3a3WM Spotify: bit.ly/4sjUyrN Apple: bit.ly/41jAdrN 0:00 Intro 1:53 Research Intern Capability Timelines 4:59 Math Breakthroughs 7:59 RL Beyond Verifiable Tasks 12:32 RL vs In-Context 19:01 Allocating Compute Internally 28:18 AI for Science 31:40 Pattern Matching 33:23 Solving the Hardest Math Problems 37:40 Chain of Thought Monitoring 44:33 Generalization and Value Alignment in Models 47:57 Inside OpenAI 51:55 Quickfire

We’re excited to launch the OpenAI Safety Fellowship - supporting rigorous, independent research on AI safety and alignment, including areas like evaluation, robustness, and scalable mitigations. Applications are open through May 4, 2026!

Introducing the OpenAI Safety Fellowship, a new program supporting independent research on AI safety and alignment—and the next generation of talent. openai.com/index/introducing…

Really proud of how our auto compaction turned out. I hope you notice a clear difference in how long Codex stays coherent!

we shipped a new version of 5.3 instant to chatgpt yesterday. 5.3 was unintentionally pretty annoyingly clickbait-y. it's better in yesterday's model and we're going to keep stamping that behavior out. keep the feedback coming! help.openai.com/en/articles/…

Insane how leaky OpenAI is smh

Just use Codex. That might have been a single prompt and worked within your $20 sub

If you give GPT-5.4 a raw dump of the GPT-2 weights and ask for a <5000 byte C program to inference it, GPT-5.4 succeeds in under 15 minutes! I remember working on a similar exercise to compare results against a proprietary model in a previous paper - it took days!