Release of GPT-5 - OpenAI LIVE5TREAM: 7th August 2025

We all know what we are waiting for. But what else will be revealed?
Expect a one hour live stream with many different updates.

The YouTube livestream starts at:
2025-08-07T17:00:00Z

Join Sam Altman, Greg Brockman, Sebastien Bubeck, Mark Chen, Yann Dubois, Brian Fioca, Adi Ganesh, Oliver Godement, Saachi Jain, Christina Kaplan, Tina Kim, Elaine Ya Le, Felipe Millon, Michelle Pokrass, Jakub Pachocki, Max Schwarzer, Rennie Song, Ruochen Wang as they introduce and demo GPT-5.

Usually the comments are turned off. Share your thoughts and impressions here instead.

Let’s have a great time!

14 Likes

We need Bingo Cards stat! Hehe…

Wish I had more time!

ETA: What did we use to make them last time?

There! May we all be winners! :wink:

2 Likes

Fill in the boxes, if something announced:

  • now on ChatGPT, hoping to rollout to API in the coming weeks
  • requires organization verification to use
  • requires organization verification to stream
  • requires tier level to use
  • mini and nano, for you unqualified users
  • fixes structured outputs writing international characters as /u1234 or /x000xx that cannot work
  • fixes calling functions over and over without stop
  • fixes repeated structured outputs without stop
  • fixes functions and structured outputs with endless /n and /t
  • fixes reliably writing markdown, code fences
  • doesn’t spew emoji and unicode as formatting
  • doesn’t color the output with em-dashes, and ā€œit’s not about x, it’s about yā€
  • has no more system message, only ā€œdeveloperā€ that is degraded
  • has even more held-out context for OpenAI system messages injected
  • has well documented vision offering patches and high resolution
  • these new voices…
  • doesn’t pester the user, offering to do other things it can’t do
  • doesn’t say it is ChatGPT when asked unprompted
  • here’s benchmarks for a model we’ll never release and results of $1000 per question
  • no more lazy with response length cutoffs
  • we’re deprecating these models you use…
  • where’s GDB; there he is!
  • our partners have already developed…
  • ā€œagentā€ ā€œagentā€ ā€œagentā€
6 Likes

7 Likes
4 Likes

I don’t understand why anyone cares about GPT5 when GPT4.5 is already $75/input $150/output. What app are you planning to build on top of a model that’s even more expensive?

FWIW I’ve found little difference between 4.1 and 4.5 in my Plus account usage.

Honestly, the API is starting to mature. I’d like to see speed improvements in Agents (Agents and Response APIs are both still half the speed of Completions given the same params). It would also be nice if GPT itself understood the OpenAI APIs well enough to provide useful code and guidance. Even if you give it web search and point it to the documentation, it hallucinates or falls back to Completion API code quite readily.

I’d like to see stronger, cost-focused, task-focused models. The more reliable and less costly working with the API is, the more services we can build on top of it.

2 Likes

GPT 4.5 is a bit special in this regard and it’s more likely to expect pricing comparable to the other top models available to developers.

But who knows? Maybe there will be an offer to run the model for hours on end?

1 Like

Sam Altman posted that today’s event will take about one hour. There will be a long line-up of announcements, most likely covering both ChatGPT and the API.

2 Likes

I appreciate your take, you keep it real.

1 Like

GPT-4.5 is shut-off on the API (and nearly shut-off on ChatGPT, just a bunch of failed attempts). You’d have to offer OpenAI a lot more money than that to talk them into your own instance. And: was worth it, in a ā€œknows more than anythingā€ and a ā€œcomprehend everything that was inputā€ manner.

I suspect that anything new won’t be a similar ā€˜largest ever’ model, but an incremental model that it was time to put a new name on. Then, with your choice of ā€œstrongā€ vs ā€œspeedā€. There are hints of its underpinnings I need not detail.

A lot of fast tokens is the new paradigm: inference-time reasoning elevating models instead of a quality token #0. ā€œAgentā€ and ā€œreasoningā€ is just doing that where you can’t see it.

2 Likes

Sam must have known it’s my birthday this month :grin:

So excited. It’s been nearly two and a half years since gpt-4 was announced.

I want to see Simple-Bench demolished among other things.

3 Likes

Hey, stranger. Good to see ya!

These big announcements usually bring a lot of people back around haha.

2 Likes

I’m delighted to see that you’re all still here! Apologies for dropping off the face of the earth, failed solo startup attempt yada yada.

Definitely can’t miss this though!

2 Likes

Woohoo! One hour long? Exciting

Wow, time flies. I’m remembering the days of Davinci being barely capable of writing a proper JSON object. Now we’re watching AI churn complete modules effortlessly. Can’t wait to see what OpenAI has cooked up for us today

3 Likes

heh

OpenAI’s new GPT-5 models announced early by GitHub | The Verge

3 Likes

spoiler alert! haha - I think I’ll still watch though

1 Like

5 Likes

A little update from OpenAI while we wait:

3 Likes

Okay but seriously, can someone make the bingo cards?

:up_arrow: note the spelling of Live5tream (sic) :thinking: :up_arrow: