💻 OpenAI showed a new generation of reasoning models: o3
💻 OpenAI showed a new generation of reasoning models: o3
OpenAI CEO Sam Altman announced the new reasoning model o3 and its smaller version o3-mini.
The main advantages of o3 are:
➡️ o3 writes code at nearly the level of OpenAI developers. The primary model outperforms o1 in programming by a factor of about 1.5.
➡️ Solves PhD-level tasks 10% better than o1.
➡️ In the ARC AGI reasoning benchmark, o3 scores 87.5%—more than humans. The o3-mini result is 76%. Last generation o1 scored just over 30%, and regular neural networks scored around 5%.
More on the graphs ⤴️
💡 The light o3-mini will be available in three variants, differing in depth of reasoning. The o3-mini-low is roughly between the o1 and o1-mini, and the o3-mini-high is ahead of the o1.
"o3 is a very, very smart model, and o3-mini is an incredibly smart model but still really good performance and cost," according to Altman.
The models are not yet available to users. OpenAI is now accepting applications for security testing.
This was the last of 12 "shipmas" gifts from OpenAI.
Most interesting:
➡️ Full version of OpenAI o1 and Pro subscription for $200 per month.
➡️ Sora Video Generator
➡️ Video Mode for ChatGPT
➡️ ChatGPT Search Web Search
➡️ ChatGPT on WhatsApp
What do you think of the marathon?
❤️ — it's great, lots of new stuff!
🎃 — wait, where's the GPT-4.5?!?!
