not gpt-5, not a search engine, but we’ve been hard at work on some new stuff we think people will love! feels like magic to me. monday 10am PT.

GPT-4o is our new state-of-the-art frontier model. We’ve been testing a version on the LMSys arena as im-also-a-good-gpt2-chatbot 🙂. Here’s how it’s been doing.

I’m blown away by GPT-4o. Realtime + multimodal + desktop app. You’ll have an AI teammate on your device that’s able to help you with anything you’re working on - and it runs 2x faster and costs 50% less than before. OpenAI doesn’t make AI models. They make magic.

i think people are misunderstanding gpt-4o. it isn't a text model with a voice or image attachment. it's a natively multimodal token in, multimodal token out model. you want it to talk fast? just prompt it to. need to translate into whale noises? just use few shot examples.

AIs getting smarter isn't just about hitting benchmarks. It's also about affordability and speed. Meet GPT-4o: for devs it's twice as fast & half the cost of GPT-4 Turbo, effectively delivering double the inference within the same latency and budget! openai.com/index/hello-gp…

GPT-4o harmonizing with itself 🤯

psst ... rollout complete! 4o rate limits are now 10M TPM for any established dev

POV when you're in the midst of exponential change

hidden gem: GPT-4o mini supports 16K max_tokens (up from 4K on GPT-4T+GPT-4o) openai.com/index/gpt-4o-m…

With Amazon S3, the cost per GB of storage dropped 86% in its first decade (2006 - 2016), or ~97% if you include Glacier. The cost of gpt-3.5 level intelligence has dropped 99% in only 2 years.

4o mini is almost beating out 4o as the top user preferred model! (it's 95% cheaper) And we've just released FREE fine-tuning through Sept. There's never been a better time to build x.com/OpenAIDevs/sta…

impossible to overstate the dedication and care Michelle Pokrass (and many others!) put into Structured Outputs developers can now reliably depend on our models to deliver the exact structure they need

welcome o1 - our first reasoning model 🍓 o1 dynamically chooses to think more depending on the difficulty of the question openai.com/index/introduc…

Scaling LLMs was one of the most fascinating talks at DevDay yesterday Colin Jarvis and Jeff Harris shared their take on how to * Increase accuracy * Decrease latency & cost This tweet blew up so I put it into a blog post instead leverage.to/learn/dev/post…

The default version of the GPT-4o model is now `gpt-4o-2024-08-06`. - 50% cheaper input token costs - 33% cheaper output token costs - 100% adherence to JSON schema with Structured Outputs

The Realtime API is fully rolled out! You can build rich + fast voice experiences anywhere. Check out what a couple partners have done already Tinker, use it, please send us your feedback platform.openai.com/playground/rea…

This is one of the paradigms I'm excited about with the realtime api, I think of it as "interface control". If you expose tools to the model it's *really* good at interpreting what you want and takes action quickly. Lots of cases where you could replace point-and-click with "say

pretty soon talking to computers will feel as natural as talking to anyone else

the first chat model that can hear and speak!