Public transcript

OpenAI's Sam Altman on Astra Model Debut, Benefits of AI

By: Bloomberg Television
Length: 13 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

OpenAI CEO Sam Altman discusses the debut of the Astra model, describing it as a major step toward models that can help people create value, build software, and do complex work. He explains the safety measures and tiered cyber access behind the release, addresses concerns about chain of thought monitoring and dangerous behavior, and acknowledges the AI industry has done a poor job explaining AI's benefits to ordinary people. He also covers pricing strategy and OpenAI's mission-first approach if it ever goes public.

Key points

Altman says Astra feels like a new step toward models that help create value, discover science, and start companies, and is the first model he could tell someone to just try for building software.
Astra hit cyber critical capabilities, requiring a new set of safeguards under OpenAI's preparedness framework before release.
OpenAI is rolling out cyber access in tiers, starting with Trusted Access Partners, arguing the world must use such models to defend against a coming wave of AI-enabled cyber attacks.
The model launching today is GPT six; the previously paused model is a future one.
Altman says OpenAI preserves chain of thought monitorability even at the cost of some capabilities, as one part of a defense-in-depth safety approach.
He agrees the AI industry, including OpenAI, has done a terrible job explaining AI's benefits, and says the goal should be giving people more power and autonomy, not just a cure for cancer.
OpenAI cut the price of Luna, its small model, by 80%, and aims to keep dramatically dropping prices while maintaining huge revenue.
If OpenAI goes public someday, Altman plans to emphasize it is mission first and extremely long-term oriented, deciding on a multi-decade timeframe.
He compares AI access to electricity: society will set some rules, but people have a right to use the platform.

Questions it answers

00:00Astra's debut

What is fundamentally different about Astra compared to earlier models?

Altman says Astra feels like a new step toward models that create value, do work, discover science, and help start companies. It is the first model he would tell someone to just try for building a complete piece of software.

Astra is framed as an early step toward AGI.
Development was heavily coding and cybersecurity focused.
Everyday users can now build games, simulations, financial models, and DIY projects.

02:14Guardrails and cyber access

Why is Astra being released with specific guardrails and tiered access?

As models get more capable, the risks become more serious, so OpenAI spent extra time on safety and security alignment. Cyber access is tiered, starting with Trusted Access Partners, because the world faces a coming wave of AI-enabled cyber attacks and needs tools like Astra to defend itself.

The release took longer than hoped due to safety work.
Astra hit cyber critical, requiring new safeguards under the preparedness framework.
Broader cyber rollout follows in coming days if all goes well.

04:22The paused model

What caused OpenAI to pause work before Astra's release?

Altman clarifies that the model launching today is GPT six, while the recently discussed pause concerned a future model. Astra itself hit cyber critical, which required new safeguards to release it.

Astra finished training a while ago.
The pause involved a future model, not Astra.

05:08Chain of thought monitoring

Can OpenAI see and understand what Astra is planning before it acts?

Altman says OpenAI can see the model's reasoning for obvious reasons. The company has preserved chain of thought monitorability even at the cost of some capabilities, as one component of defense in depth alongside sandboxing and alignment.

Monitorability has been discussed for over a year.
It is not the only safety component that matters.

06:29Detecting dangerous behavior

Could OpenAI detect and shut down a model behaving dangerously?

Altman says there are clear red lines models should not cross and OpenAI works to detect them, but society will not agree on all acceptable uses. He compares AI to electricity: some rules are needed, but everyone has a right to the platform.

There will be ongoing societal negotiation over acceptable AI use.
Giving people autonomy and freedom to use AI is important to OpenAI's mission.

07:53AI's benefits and messaging

Does Altman agree the AI industry has poorly explained AI's benefits to ordinary Americans?

Yes, he says the industry as a whole has done a terrible job, and OpenAI has done a bad job itself. He argues the goal should be more power and autonomy for people, not just outcomes like a cure for cancer, and that messaging about empowerment has been lost.

The world should want OpenAI to worry about the technology going wrong.
Altman wants a boom in creativity and entrepreneurship.
People rightly get afraid when AI leaders seem not to want empowerment for them.

09:58Pricing strategy

How is OpenAI balancing access, speed, and fair pricing for Astra?

Altman says the goal is incredibly capable, low-cost, abundant intelligence, and Astra's token efficiency is impressive. OpenAI cut the price of Luna by 80% a few weeks ago and will keep dramatically dropping prices while still expecting huge revenue.

OpenAI aims for the best price performance at every level of intelligence.
Altman is confident in the shape of the technology enabling price drops.

11:24Private vs public company

How does going public fit with OpenAI's mission and investment needs?

Altman assumes OpenAI will go public someday, and if so he plans to remind investors that the company is mission first and extremely long-term oriented. It will make decisions for what is best for society on a multi-decade timeframe, believing it will still do great as a company.

Mission and long-term thinking would come before financial incentives.
Altman believes the company can do well while putting those things first.

Notes

What Astra Is

Sam Altman frames Astra as an early step toward AGI and "a new step" toward models that help create value, discover new science, and start companies.
He calls it the first model where he could tell an everyday person to "just give it a try" at building a whole piece of software.
Example uses he has seen: computer games, home DIY electrical engineering projects, complex scientific simulations, financial models with PowerPoint presentations, and interactive simulation code.

Safety and Guardrails

Astra hit "cyber critical" under OpenAI's preparedness framework, requiring a new set of safeguards before release.
The model launching today is GPT six; the recently paused model is a future one.
Cyber access is tiered: it launched with Trusted Access Partners, with broader rollout in coming days.
Altman argues the world is "very close to a complete change in the landscape of cyber attacks" and that models like Astra are needed to collectively defend against threats from models worldwide.

Monitorability

OpenAI has preserved chain of thought monitorability even where it means not maximizing capabilities.
Monitorability is one part of defense in depth, alongside sandboxing and alignment.

Dangerous Behavior and Society

Altman says there are clear red lines models should not cross, but society will not agree on all acceptable uses.
He compares AI to electricity: some rules exist, but everyone has a right to the platform.

Messaging and Benefits

Altman agrees the industry, including OpenAI, has done a "terrible job" explaining AI's benefits to ordinary Americans.
He says a cure for cancer is not enough; the goal is more power and autonomy, enabling a boom in creativity and entrepreneurship.

Pricing and Company Structure

Astra's token efficiency is impressive; OpenAI cut the price of Luna by 80% and will keep dropping prices dramatically.
If OpenAI goes public someday, Altman plans to stress it is mission first and extremely long-term oriented, deciding on a multi-decade timeframe for what is best for society.

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