Public transcript

Sam Altman: "Codex is Another ChatGPT Moment".

By: ExpovistaTV
Length: 18 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In highlights from the Cisco AI summit, OpenAI CEO Sam Altman discusses the launch of Codex, calling it the first time since ChatGPT that he has felt another 'ChatGPT moment' and a clear glimpse of the future of knowledge work. He covers agent computer use, the security and software redesign problems it raises, the surprisingly slow absorption of AI by enterprises, and OpenAI's outlook on compute demand, open source, business models and future model improvements. He predicts models will subjectively feel about 10x better by the end of the year and warns that companies slow to adopt AI co-workers will be at a huge disadvantage.

Key points

Sam Altman says the Codex launch feels like the first time since ChatGPT that he has felt another 'ChatGPT moment'.
He gave Codex full control of his computer within about two hours because it was so useful, and now runs two laptops to manage the risk.
He believes code plus generalized computer use is here to stay, and that OpenClaw did an incredible job making agent computer use feel usable and real.
He sees a capability overhang bigger than any time before, even as diffusion and absorption of AI by enterprises feels surprisingly slow.
Companies not set up to quickly adopt AI co-workers will be at a huge disadvantage.
He worries the US lacks an open source lead and says people increasingly want locally running private models.
He expects models to subjectively feel about 10x better by the end of the year.
He imagines future business models like AI cloud subscriptions and OpenAI acting as an investor in AI-driven scientific discoveries such as curing a disease.

Questions it answers

00:00Codex as a ChatGPT moment

Why does Sam Altman call Codex another ChatGPT moment?

He says the models have hit a threshold and the interface and harness have caught up, making Codex feel like the first real ChatGPT moment since ChatGPT itself.

Codex is a software development agent that writes, understands, debugs and refactors code from natural language.
A coding model that can build full software and interact with the real world to build a company around it is a very big deal.

01:13Computer use and OpenClaw

What does Altman think about giving AI agents control of your computer?

He gave Codex full control of his computer within about two hours because it was so useful, then was persuaded to stop and now uses two laptops. He thinks code plus generalized computer use is here to stay and credits OpenClaw with making it feel usable and real.

Agents with full computer and browser access lead to incredible stuff.
He imagines new social networks where people's agents interact and collaborate on their behalf.

04:05Constraints and security

What constraints does Altman see for AI?

Obvious ones are energy, manufacturing and hardware. Non-obvious ones include balancing security and data access versus utility, and rewriting software to be equally usable by humans and AIs.

A new security or data access paradigm needs to be invented.
His agent using Slack marked threads as read and broke his workflows.
Always-on computing is not supported by current hardware, permissioning or legal systems.

06:03Capability overhang and slow diffusion

How big is the gap between AI capability and adoption?

He says the capability overhang feels bigger than ever, even bigger than right before ChatGPT launched, yet diffusion and absorption by enterprises is surprisingly slow.

AI is already doing research and large amounts of economic work.
Companies not set up to quickly adopt AI co-workers will be at a huge disadvantage.

09:13Codex as a teammate

What makes the Codex app feel different from ChatGPT?

He says it is the first thing that truly felt like interacting with a teammate rather than a tool, showing how much value there is in packaging and interface even with amazing underlying models.

ChatGPT always felt like a tool; Codex feels like a collaborator.
That feels like the shape of things to come.

10:14Compute demand and infrastructure

Will there be enough compute and infrastructure for AI?

He expects demand for tokens to grow for decades even as efficiency improves, with possibly temporary supply gluts, and compares AI demand to electricity demand at different price levels.

Around $5 trillion may be spent; if spent quickly that could be enough.
Capable and cheap AI will create a ton of demand at some price.

12:54Open source and local models

Does Altman worry about the US open source gap?

Yes, he worries the US lacks an open source lead and says OpenAI is limited by focus and time, though he thinks leading on frontier models matters most.

People want control of their own models and local inference.
Always-on devices that see your whole life should run on inference you control.

14:04Business models

How will OpenAI make money from AI?

Many tens of millions of consumers pay subscriptions, which surprised him, and businesses want AI cloud subscriptions combining agents, enterprise licenses and API access; he also imagines OpenAI investing in capital-intensive discoveries like curing a disease for a royalty.

Advertising may work for mass-scale consumer businesses but must be done carefully.
He imagines billions of humanoid robots building data centers and power plants, and unprecedented economic growth.

Notes

Codex as another ChatGPT moment

Sam Altman says the models have hit a threshold and the interface and harness have now caught up.
He calls Codex the first thing since ChatGPT that feels like another ChatGPT moment: a clear glimpse of the future of knowledge work.
A coding model that can build full complex software and interact with the real world to build a company around it is a very big deal.

Agents with full computer use

Altman gave Codex full control of his computer within about two hours despite intending not to; it was so useful.
He was persuaded to stop running it that way and now uses two laptops until he figures out how it should work.
He credits OpenClaw with bringing many ideas together to make agent computer use feel usable and real; Moldbook may be a fad, but agent-driven social interaction points to a real future social network.

Constraints and open problems

Obvious constraints: energy, manufacturing, hardware.
Non-obvious ones: balancing security and data access versus utility — a new security paradigm needs to be invented.
Software must be rewritten to be equally usable by humans and AIs; his agent using Slack broke his workflows by marking things as read.
Always-on computing (an AI watching meetings and your screen) is not supported by current hardware, permissioning systems, or the legal system.

Capability overhang and adoption

The capability overhang feels bigger than ever, even bigger than right before ChatGPT launched.
AI is already doing research and large amounts of economic work, but diffusion and absorption by enterprises is surprisingly slow.
Companies not set up to quickly adopt AI co-workers will be at a huge disadvantage.

Compute, business models and open source

He expects demand for tokens to grow for decades even as efficiency improves; possibly temporary supply gluts but massive infrastructure buildout ahead.
Consumers surprised him by paying subscriptions — many tens of millions do.
He worries the US lacks an open source lead; people want local, private models, especially for always-on devices that see your whole life.
He imagines AI cloud subscriptions for businesses and OpenAI acting as an investor in capital-intensive scientific discoveries, like curing a disease for a royalty.
He expects models to subjectively feel about 10x better by the end of the year.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.