ChatGPT

@deepfates would you like at that? it's time to run the numbers

giving ChatGPT increasingly insane situations and then saying "let's run the numbers"

@deredleritt3r nice

Added to prinzbench: GPT-5.4 I love this model so much that I benchmarked it *THREE* times: - "Heavy" thinking (in ChatGPT app) - "Extended" thinking (in ChatGPT app) - xhigh (in Codex) "Heavy" thinking and "xhigh" yielded essentially the same results, and these results are *absolutely superb* (see graph below). GPT-5.4 simply obliterates all other models I've benchmarked to date. "xhigh" achieved the highest overall score (69/99), the highest Legal Research score (50/99) and the highest Search score (19/24) to date on prinzbench. In its GPT-5.4 release blog post, OpenAI seemingly called out my benchmark by stating that GPT-5.4 is great "particularly for 'needle-in-the-haystack' questions". I can confirm that this is fully accurate and then some. GPT-5.4 (xhigh) scored 19/24 in Search, whereas the highest Search score a non-OpenAI model has ever managed on my benchmark is 9/24 (Gemini 3.1 Pro/Grok 4.20). Congrats to OpenAI! This model is just fantastic. GPT-5.4 Pro will be benchmarked soon(TM).

140 million people use ChatGPT to help them understand math and science concepts every week.

New Model New Me

ChatGPT's Code Blocks can support many kinds of interactive visualizations!

🆕 Fresh coat of paint to Code Blocks in @ChatGPTapp ! I esp like using them for in-chat data visualizations - Code Blocks natively support mermaid diagrams, vega, HTML, React, with more 🔜 This is an early launch as we make this a more interactive and useful part of ChatGPT

Lots shipped @ChatGPTapp this week! Super focused on getting the small things right, we - Doubled file upload limit (you can now attach up to 20 files in a single message to chat) - Improved select-all (cmd+a) behavior to copy just the conversation vs. any excess on screen - Fixed share button color contrast (now matches your accent color too!) - Now support more file types (text/code files) - (android) New quick-tools menu from composer - (android) Simplified composer into a unibar - (ios) Many performance fixes to make the app faster - Many reliability improvements to cut long “Thinking” chat failure rates by more than half Lots more to come in the coming weeks both in terms of polish to make the app quicker and shipping long-requested features Let us know what you'd like to see!

@viemccoy banger.

I think the "how many R's are in the word strawberry" is a sort of optical-illusion that affects minds of the transformer type. You wouldn't dismiss a human genius because they are susceptible to misperceiving the color of a dress! Total red herring, and not the only example.

research grade intelligence at your fingertips

@SteveMoraco these are API metrics sam was reporting. moreover he was counting output tokens too. the number of directly input words from real humans on ChatGPT is a more special metric

@ChatGPTapp Wait so usage is down quite a bit from two years ago?

on Wednesday, February 11th, more than 4 billion messages were sent to and answered by ChatGPT. this represents more than 160 billion words spoken to a machine intelligence in one day.

Tomorrow at 10am PT legacy models (GPT-5, GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini) will be deprecated in ChatGPT. openai.com/index/retiring…

You can just build things.

@MrMekzy_ i feel seen

@TheJennyAI wrong x.com/patrickc/statu…

@ChatGPTapp Stalin said this about soldiers OpenAI said this about parameters History rhymes

Quantity has a quality all its own

@ichbinilya @tszzl wrong x.com/patrickc/statu…

@ChatGPTapp @tszzl ah yes, Josef Stalin

@MillionInt thank you JT

@ChatGPTapp Banger