[Live Blog] OpenAI Releases GPT-5

Welcome to the live blog of OpenAI’s much-anticipated release of GPT-5.

This might be the biggest tech launch in years. GPT-5 will be the next iteration of OpenAI’s AI models following a slightly disappointing GPT-4.5. OpenAI says it will combine its reasoning and GPT series of models into a unified structure with GPT-5.

Along with GPT-5, it’s widely expected that OpenAI will launch GPT-5-mini, GPT-5-nano and GPT-5-chat. These model names were given away in a Github leak yesterday, and the model images had also been found in the API.

Post livestream

GPT-5 pricing

Sam Altman has revealed the pricing for GPT-5. GPT-5 has gone aggressive on pricing its APIs. For GPT-5, its most powerful model, it charges $1.25 per million input tokens, and $10 for output tokens. The other models are even cheaper, with GPT-5-nano — the smallest model, charging just $0.05 per million input tokens, and $0.40 per million output tokens.

[11:35 am PST]

Cursor co-founder comes up. Says they’ve been using GPT-5 for a while. Says GPT-5 figured out an oddity in its codebase, which would’ve taken humans might longer to figure out. Says GPT-5 is incredibly smart and easy to use. It’s fast and interactive. Good at back-tracking too.

[11:20 am PST]

Some benchmarks are being saturated, says Greg Brockman.

GPT-5 can work as a pair programmer. GPT-5 can modify its own prompts through meta prompting.

[11:11 am PST]

OpenAI President Greg Brockman talks about how GPT-5 will change software engineering. GPT-5 is the best model for agentic coding tasks, he says. The model can maintain context for long period. It’s great at front-end development and instruction following. It’s extremely fast as well.

OpenAI shipping 3 reasoning models in the API — GPT-5, GPT-mini and GPT-5 nano. There’s a “minimal reasoning” option for low reasoning tasks.

API has custom tools. These are plain text, which can be used for outputs instead of json. The model’s outputs can be constrained to a phrase or format. It also has tool call preambles, and the model can give additional information. It also has a “verbosity” parameter, to determine how terse or expansive it is with outputs.

These are the results for SWE-bench verified and Aider Polyglot benchmarks.

Results of GPT-5 on T2 benchmark, which tool calling.

GPT-5 has a context of 400k in the API.

[11:06 am PST]: Health

Sam Altman returns. Says health queries are a top use-case of ChatGPT. Says GPT-5 is the best model in the world for health. It has a top score on Health Bench, a new benchmark that OpenAI has created.

A user says she was detected with three different types of cancers. She put in her lab reports in ChatGPT. ChatGPT explained the lab reports in plain language. The user says this helped in a moment of panic. By the time she was ready to talk to a doctor, she had a baseline understanding of what to wrong. User says she’s been using it to make critical decisions and advocate for treatments. When faced with a decision to use radiation, the doctors themselves didn’t agree on the course of action. The doctors asked her what she wanted to do. She approached ChatGPT, which explained both sides of the decision and weigh pros and cons, and helped her make a decision. Her husband says that ChatGPT filled the knowledge gap and become an active participant in her care journey.

User says she’s been testing GPT-5. Says it’s extremely fast, and very thorough. Says it feels like a thought partner. GPT-5 explained the biopsy results even better than GPT-4o, which they’d used earlier.

[11:04 am PST]: Synthetic data

Previous generations of models are helping create data to train new models. This can lead to recursive self-learning, says the presenter.

[11:00 am PST]: GPT-5 safety

GPT-5 is less deceptive than o3 and o4-mini, says the presenter. GPT-5 goes beyond just the prompt to understand intent. GPT-5 has “safe completions”, through which rather than judging the user’s prompt, it tries to understand the intent. Based on this, it can refuse to answer, or give a partial answer. GPT-5 will also tell the user why it’s unable to answer a question.

[10:30 am PST]: Memory features

ChatGPT gets access to Gmail and Google Calendar. The data flows seamlessly into ChatGPT, and can create plans, and suggest times for meetings. It can also surface emails that haven’t been read.

[10:25 am PST] Voice features

Free users to get voice mode “for hours”.

GPT-5 can help users practice a new language, such as Korean. GPT-5 can speak Korean which will allow users to catch the pronunciation. It’ll also explain what each word means. “The voice sounds a lot more natural than the voice we’d demoed a year ago,” says the presenter.

Paid subscribers can change colour of chats. GPT-5 can also take on different personalities, sort of like Grok’s new avatars..

[10: 18 am PST]

GPT-5 is trying its hand at vibe coding. GPT-5 was tasked with building a web app to learn French. GPT-5 seems to be basically creating a simple Duolingo in minutes. The app has flashcards, a quiz, and an animated mouse and cheese game. The app also has a progress bar. There’s a snake-like game, in which losing teaches the learner a new French word.

[10:15 am PST]

All previous models will be deprecated with GPT-5.

GPT-5 wrote a eulogy for the older models. OpenAI compared the outputs of GPT-4 and GPT-5. “GPT-5’s text has better rhythm,” says the presenter. It seems to create more emotionally more resonant outputs. “Responses feel less like AI, and more like chatting with your high EQ and IQ friend,” she says.

[10:10 am PST]

GPT-5 can create a moving demo with an SVG to explain the Bernoulli effect. GPT-5 wrote more than 400 lines of code over 2 minutes to create the animation. The demo is interactive — users an pick the variables to see how they’ll impact flow.

[10:00 am PST]

Altman says that it’s been 32 weeks since ChatGPT was released. Now 700 million people use it every week.

GPT-5 is a major update, and a significant step to AGI.

GPT-3 was like a high-school student. GPT-4 was a college student. GPT-5 is a PhD level expert.

Mark Chen, Chief Research Officer at OpenAI says OpenAI pioneered reasoning. GPT-5 will automatically decide whether to reason or not, instead of users having to pick a model themselves. He indicates GPT-5 will be the best coding model on the market today.

GPT-5 benchmarks

GPT-5 will be available to the free tier: Free users will get some free GPT-5 queries.

[9:53 am PST]: OpenAI is billing the GPT-5 launch as its “Summer Update”.

[9:51 am PST]: After several hints (including saying “Live5tream” in the announcement) It’s become pretty clear that OpenAI will indeed launch GPT-5 today. The thumbnail for the livestream boldly says GPT-5.

gpt-5 launch live blog
Posted in AI