Public transcript

FULL INTERVIEW: Sam Altman Responds to Anthropic’s Attack Ads, Live on TBPN

By: TBPN
Length: 37 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Sam Altman joins TBPN live to discuss the launch of 5.3 Codex, which he calls the best coding model in the world, along with the Frontier enterprise platform and the Codex desktop app. He shares views on agent orchestration, the capability overhang, SaaS disruption, Sora and the Disney partnership, and compute bottlenecks. He also responds directly to Anthropic's Super Bowl attack ads, saying OpenAI will never put ads into the LLM stream, and describes internal OpenAI as calm despite intense media scrutiny. The conversation closes with Altman prioritizing coding and science models as the most exciting frontier.

Key points

OpenAI launched 5.3 Codex, which Altman calls the best coding model in the world, with mid-turn interaction, better personality, and strong computer use.
Altman says there is a capability overhang: models are so capable that better tools and orchestration matter more than intelligence for a while.
On Anthropic's attack ads, Altman says OpenAI will never put ads into the LLM stream and calls the ads deceptive, while admitting they were funny and well played.
He says chips are the current bottleneck for compute, and society should fund more wafer capacity and fab supply chains.
Codex desktop has been a surprise hit, and Altman wants a unified AI backend across surfaces, including mobile task kick-off.
He is excited about the Disney partnership for generating characters, and says memes on group chats are the killer use case for Sora.
He describes internal OpenAI as calm and busy, while media coverage is largely divorced from reality.
He would rather prioritize coding and science models over a writing model, and expects a great voice mode by the end of the year.

Questions it answers

00:005.3 Codex launch

What did OpenAI just launch and why does it matter?

OpenAI launched 5.3 Codex, which Altman calls the best coding model in the world. It is faster, smarter at programming, supports mid-turn interaction, and is good at computer use.

Mid-turn steering unlocks more than one-shot execution for long tasks.
Expert users noticed the change mid-deploy.

02:48Agents and orchestration

How does Altman see agent orchestration evolving?

He sees people managing teams of agents at higher levels of abstraction. Because there is a capability overhang, tools that make orchestration easy will matter more than intelligence for a while.

Forward deployed engineers connect non-AI-native companies to platforms like Frontier.
Security of data is the most frequent enterprise concern.

09:48Benchmarks and GDP

What metrics will matter as agents scale?

No AI chart lasts more than a few years; a joke at OpenAI is that the chart that matters will be GDP impact. Agents can subdivide work rather than needing super long context.

GDP could fall while quality of life rises.
Agents orchestrating sub-agents mirrors how people work.

12:24Research labs and SaaS

What does Altman think of the new research lab boom?

He thinks it is fantastic that industry is doing research again, something OpenAI hoped to help restart. The best efforts will look like a mixture of product and research.

Compute power is the statement closest to true for him.
Some SaaS companies are durable; some feel like a thinner layer now.

16:00Codex desktop and OpenClaw

How is Codex desktop going and what about OpenClaw?

Codex desktop has been a surprise hit; he wants mobile task kick-off and a unified AI backend across surfaces. He loves the OpenClaw homebrew spirit that a big company could never ship.

He built a magic auto-completing to-do list with the app.
A general knowledge-work agent version is coming.

20:00Social, Sora and Disney

Where are social and video generation heading?

Bots on existing platforms are making everyone crazy; he favors assertion of humanity over AI detection and doubts platforms want to solve it. Sora's killer use case is memes on group chats, and the Disney character rollout will be important.

He would open the floodgates on Disney generation.
The Super Bowl ad celebrated the revolution for researchers, not average users.

26:00Anthropic's ads

What does Altman think of Anthropic's attack ads?

He says the ads are wrong because OpenAI will never put ads into the LLM stream, and using deceptive ads to criticize deceptive ads does not sit right with him, though he admits the ad was funny and well played.

OpenAI has not even started ad tests yet.
He calls the feud a sideshow compared to product momentum.

30:00Compute, IQ and media

What are the compute bottlenecks and how is OpenAI handling scrutiny?

Chips are the current bottleneck, and society should fund wafer capacity and the fab supply chain. Internally OpenAI feels calm and busy while media coverage is largely divorced from reality.

He expects a great voice mode by the end of the year.
No known upper bound on model IQ.

Notes

5.3 Codex Launch

OpenAI launched 5.3 Codex, which Sam Altman calls the best coding model in the world, incorporating feedback from 5.2.
It is much smarter at programming, faster, supports mid-turn interaction, has a better personality, and is good at computer use.
Expert users noticed the difference mid-deploy on launch morning.

Agents and Orchestration

People now run multi-hour tasks; steering agents mid-task unlocks much more than one-shot execution.
Altman sees a workflow where people manage a team of agents at ever-higher levels of abstraction.
There is a capability overhang: tools and orchestration will matter more than intelligence for a while.
Forward deployed engineers connect non-AI-native companies to platforms like Frontier, addressing security and data concerns.

Benchmarks and Economics

No AI chart lasts more than a few years; a joke at OpenAI is that the chart that matters will be GDP impact.
Agents can subdivide and orchestrate work without needing super long context windows.
GDP could go down while quality of life goes up, which we have little practice handling.

Industry Views

New research labs across Silicon Valley are welcome; the best efforts will mix product and research.
Compute power is the new oil; data alone was wasted a decade ago.
Software is not dead but will transform; every company is an API company now, and agents will use and pay for other services.
Some SaaS companies will be durable, some are a thinner layer now.

Codex Desktop and OpenClaw

Codex desktop has been a surprise in how much people love it; 10% polish goes a long way given the overhang.
Altman built a magic auto-completing to-do list with the app and wants mobile task kick-off soon.
He loves the spirit of OpenClaw and the homebrew computer club energy of open source.

Social, Sora, Disney

Bots on existing social platforms are making everyone crazy; assertion of humanity beats AI detection.
People love generating Sora videos; memes on group chats are the killer use case.
Generating characters with Disney will be important; Altman would open the floodgates.

Ads and Anthropic

OpenAI's Super Bowl ad celebrated the AI revolution for researchers, not average users.
Anthropic's ads are wrong: OpenAI will never put ads in the LLM stream; using deceptive ads to criticize deceptive ads doesn't sit right, though the ad was funny and well played.
OpenAI hasn't started ad tests yet.

Compute and Media

Chips are the current bottleneck; society should fund wafer capacity and the fab supply chain.
No upper bound on model IQ is known; Altman can't reason about 2,000 IQ.
Internally OpenAI feels calm and busy; media coverage is 99% divorced from reality.

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