Sam Altman on GPT-5.3 Codex, the Capability Overhang, and Why Anthropic's Attack Ads Don't Matter

The most interesting thing Sam Altman says in this interview isn’t about GPT-5.3 Codex or Anthropic’s attack ads. It’s a quiet admission: intelligence is no longer the bottleneck. There’s such a capability overhang in current models that better tools matter more than smarter models, at least for a while.

The Interview

Sam Altman joins TBPN hosts John Coogan and Jordi Hays on launch day for GPT-5.3 Codex, OpenAI’s latest coding model. The conversation ranges from the new model’s capabilities to SaaS disruption, Anthropic’s Super Bowl campaign, and where compute bottlenecks stand. Altman is relaxed, candid about competitors, and clearly most excited about the Codex desktop app.

The Capability Overhang

Altman frames the current moment around a single concept: capability overhang. The models are already far more capable than most users realize or utilize. The real constraint is tooling and user education.

“I think tools to make that easy to do will matter more than intelligence for a little while because there’s such an intelligence overhang already.”

This shapes his view on almost everything. Forward-deployed engineers exist because companies don’t know how to use what’s already available. The Codex desktop app matters because 10% more polish unlocks dramatically more value. Even OpenAI’s advertising strategy is about reducing the overhang: teaching people what they can do, not what the models can do.

The trajectory he describes is people managing increasingly capable agent teams, working at higher and higher levels of abstraction, always at the edge of their cognitive bandwidth. Not smarter models, but better orchestration of already-smart models.

GPT-5.3 Codex and Mid-Turn Interaction

The new model introduces mid-turn interaction for long-running tasks. Instead of one-shot prompting for multi-hour coding sessions, users can steer the model while it works, correcting course without starting over.

Altman draws an analogy to training a new coworker: you wouldn’t let them work for hours without feedback and then judge the final result. Current models demand exactly that. Mid-turn interaction breaks that pattern.

“It is incredible what these models can do without any feedback… but they can do much more amazing things if you steer them along the way.”

Expert users noticed the model improvement during deployment, before any announcement, calling it “really different” from GPT-5.2 Codex.

Benchmarks Are Dying, GDP Impact Is Next

When asked about long-task benchmarks, Altman makes two observations. First, no benchmark chart in AI lasts more than a few years. Second, agent orchestration through sub-agents and parallelization is already proving more effective than brute-force long-context approaches, echoing how human teams subdivide work.

The internal joke at OpenAI: the only chart that will matter soon is GDP impact. But Altman adds a twist. GDP itself may become the wrong metric as AI drives massive deflation. Quality of life could go way up while measured GDP goes down, and we don’t have frameworks for that.

Is Software Dead?

Altman’s answer: different, not dead.

The SaaS market will remain volatile. Big sell-offs will continue as models improve, but so will big booms. Some SaaS companies with strong systems of record will survive the transition. Altman reports that SaaS companies he talks to aren’t demoralized; they recognize that AI lets them rebuild faster too.

The framing that stuck with him most recently:

“Every company is an API company now whether they want to be or not.”

Agents will interact with services directly, which threatens ad-supported business models (as Uber’s CEO acknowledged on the same show) but also creates new distribution channels. Business models and revenue sharing will evolve.

No public SaaS company has approached OpenAI for an acquisition or “soft landing” that Altman is aware of.

Codex Desktop: The Surprise Hit

The Codex desktop app surprised Altman with how much people loved it, including himself. He calls it a “profound shift” in his own workflow, extending beyond coding to general knowledge work.

His personal use case: an auto-completing to-do list. You add tasks, and the AI tries to complete them. If it can, it does. If it needs clarification, it asks. If you need to handle something manually, you still can. A single interface where you describe what you want done and the system tries to do it.

The vision is a unified AI backend across devices: desktop for deep work, mobile for adding tasks and quick interactions, all sharing context, memory, and data. Still too technical for non-technical users, but that will change.

OpenClaw and the Open Source Spirit

Altman is openly enthusiastic about OpenClaw, the open-source AI coding agent built by Peter Steinberger. He acknowledges what makes it possible: a solo open-source developer can move fast without worrying about lawsuits, data privacy, or hyperscaler partnerships that would paralyze a large company.

“Letting the builders build the equivalent of the homebrew computer club spirit go here is so important.”

He frames it as the natural innovation cycle: something starts in open source, proves clearly amazing, then someone finds a way to make a mass-market version.

Anthropic’s Attack Ads

Anthropic ran multiple Super Bowl ads criticizing the prospect of ads in AI chatbots. Altman’s response is measured but pointed.

OpenAI’s first principle on ads: nothing goes into the LLM stream. Putting ads in model responses would feel “crazy dystopic” and would rightfully drive users away. The deceptive part, Altman argues, is using a deceptive ad to criticize deceptive ads.

He concedes the ad was “well played” and that it accurately captured what most annoys him about ChatGPT’s personality, which he says will be fixed soon. But he also notes that Anthropic’s blog post included language reserving the right to later revise their no-ads position.

His bottom line: it’s a sideshow. The Codex launch and the groundswell of excitement around coding agents matter far more.

Chips, Energy, and the Scaling Curve

The current bottleneck is chips, though it alternates with energy. Altman advocates for a societal decision to increase global wafer capacity, though he acknowledges normal capitalism might solve it.

On whether models have an IQ upper bound: “It seems certain” they can get much smarter. But what 2,000 IQ or 10,000 human-years of thinking actually means is something no one can reason about yet. His intuition: it won’t feel as strange as it sounds, because humans are so focused on their own lives that super-intelligent AI running companies and inventing science will just be “very weird” rather than “impossibly weird.”

On the broader AI trajectory, the scaling curve has been “incredibly smooth” for six to eight years. Every prediction that it would top out has been wrong. The magic relationship: more compute, data, and new ideas go in, and roughly the log of it comes out as improvement.

“Compute power is the new oil is the statement that feels closest to true to me.”

Space data centers providing meaningful compute for OpenAI? “No.” Not in two, three, or five years.

A Few Observations

This is a launch-day interview, and Altman is in product-promotion mode. But several threads cut deeper than PR:

• The capability overhang thesis is the most strategically significant claim. If true, it means the competitive moat shifts from model intelligence to product craft and distribution, at least temporarily. This explains OpenAI’s heavy investment in desktop apps, enterprise platforms, and forward-deployed engineers.
• The “every company is an API company” framing has real teeth. If agents can directly call any service, the entire advertising and discovery layer of the internet becomes contestable. Companies that resist agent access will get routed around.
• Altman’s reaction to Anthropic’s ads reveals an interesting tension. He’s dismissive (“sideshow”) but engaged enough to have asked Claude for the definition of “playing dirty.” The fact that he found the personality critique accurate and promised a fix suggests Anthropic landed at least one punch.
• The GDP-as-wrong-metric observation deserves more attention. If AI drives deflation so severe that GDP declines while quality of life rises, we’ll need fundamentally new economic measurement frameworks, and policy built on GDP growth targets could become actively counterproductive.
• Altman is preparing a blog post on predictions for the next five years (ten years felt too far). Given his notably accurate 2016 predictions about chatbot attachment, this is worth watching for.