ChatGPT's Next Act: From Chatbot to Super Assistant

ChatGPT was supposed to be a demo. One month, then shut it down. Three and a half years later, it has 900 million weekly active users and is being rebuilt into something fundamentally different: a super assistant that doesn’t wait for your questions but works on your behalf.

The Accidental Product

Nick Turley, OpenAI’s VP of Product, joined the company through what might be the most effective recruiting pipeline in tech history: he asked to skip the DALL-E 2 waitlist, and got an interview instead. Three and a half years later, he’s overseeing one of the fastest-growing consumer products ever built.

The origin story of ChatGPT as a business is surprisingly scrappy. The product launched as a free demo with a planned one-month lifespan. When it went viral, the team realized they had an actual product on their hands, but one that kept crashing at capacity. Subscriptions weren’t a monetization strategy; they were a demand-shaping tool, a way to gracefully turn away users when servers couldn’t handle the load.

“ChatGPT originally was entirely free and the reason for that was that it was intended to be a demo and we were going to wind it down after a month.”

The same pattern repeated with GPT-4: too expensive to serve to everyone, so it went behind the Plus paywall. OpenAI stumbled into its subscription model by solving infrastructure problems, not by running a pricing strategy.

The North Star: Long-Term Retention

When asked to allocate 100 points across all possible metrics, Turley puts everything on one: long-term retention.

“I care a lot about long-term retention and I would put all my points there. Because ultimately the sign of durable value is whether or not people are coming back in three months.”

ChatGPT’s retention curves are “smiling,” a rare phenomenon where users who initially churned come back months later. Turley attributes this to the nature of AI delegation: it takes people multiple months to discover all the ways ChatGPT can fit into their lives. Search and personalization were the two biggest product investments that drove this, transforming ChatGPT from a “worky” weekday tool into a mobile-first personal assistant with steady weekend usage.

The growth has been roughly one-third, one-third, one-third:

• Classic friction removal: removing the login wall, improving onboarding (Sam Altman apparently pushed for no-auth from day one)
• Core product investments: search, personalization, writing blocks, all built jointly between research and product teams
• Pure model improvements: step changes like GPT-3.5 → GPT-4 → o-series, plus the less splashy iterative updates (5.3, 5.4) that systematically address user feedback

Why the Next Billion Users Need a Different Product

ChatGPT has reached about 10% of the world. The remaining 90% won’t come through the same path. Turley identifies two fundamental barriers:

The delegation problem. ChatGPT is a “raw appliance,” a power tool that doesn’t tell you what it’s for. Users have to discover use cases on their own. For the next wave of users who are busier and less tech-curious, the product needs more affordance, more software-like structure rather than a blank terminal.

The proactivity gap. Most people are too busy to figure out what to delegate. The solution isn’t to make prompting easier; it’s to flip the model: instead of the user prompting the AI, the AI prompts the user. Pulse was the first step, but it’s limited because it can only produce information, not take action or connect to your life.

The vision is that these compound: actions + proactivity = super assistant. Imagine an AI that notices your metrics dropped and proactively runs an analysis, or detects you’ve landed at your destination and calls a cab. Not because you asked, but because it understands your goals.

The Coding Agent as Proving Ground

Domain-specific agents are already working. In coding, they’ve achieved what Turley calls “escape velocity”: many OpenAI engineers don’t open their IDE anymore, ever.

“We’ve got so many engineers who don’t open their IDE like ever.”

The key insight is why coding agents came first: code is testable (you know if it worked), RL-friendly, and the failure modes are legible. This creates the hill-climbing loop that makes products great: people try it, get partial credit, which generates real task data, which the team uses to improve.

General-purpose agents are the next frontier. The previous ChatGPT Agent attempt was “slightly too early,” models weren’t good enough for escape velocity, so users never learned to trust it. The remaining use cases were too niche (migrating file servers to the cloud) to drive broad adoption. Turley believes they’re approaching the threshold where general agents work well enough for partial credit on meaningful real-life tasks, and then the improvement flywheel begins.

Pricing in the Age of Infinite Intelligence

The current subscription model is an artifact of historical infrastructure constraints, not a deliberate pricing philosophy. As test-time compute allows intelligence to scale up nearly without limit, the metaphor shifts:

“It’s possible that in the current era having an unlimited plan is like having an unlimited electricity plan. It just doesn’t make sense.”

Power users are extracting enormous value, token consumption per user is climbing steeply (Turley describes the internal charts as “mindboggling”), and GPU consumption correlates with value delivered. The Uber/Lyft 2015 comparison is apt: subsidized usage during the growth phase, with inevitable pricing evolution as the product matures.

For casual users and emerging markets where credit cards aren’t common, ads become an access tool rather than a monetization play. The most common support inquiry about their ads pilot isn’t “how do I disable ads” but “how do I run an ad,” the ecosystem wants in.

The GPU Zero-Sum Game

GPUs are the hardest trade-off at OpenAI. Unlike traditional software, they face a genuinely zero-sum resource constraint: every GPU serving ChatGPT is one not training models or running Codex.

The naive approach would be to optimize for incremental revenue per GPU. But that would kill zero-to-one bets. Deep Research, for example, couldn’t have been justified on a per-GPU revenue basis, but not shipping it meant never discovering whether consumers wanted a research product (they did).

Turley’s heuristic: start planning from the most constrained resource (GPUs), prioritize existing user experience first (fast, reliable), then balance proven use cases against exploratory bets. He has no line of sight to a world where this trade-off disappears; demand keeps rising even as per-unit costs fall.

Code Red as a Management Tool

When Google’s Gemini gained momentum and Marc Benioff publicly switched to it, OpenAI declared a “Code Red,” pausing work on ads, health agents, and shopping to focus entirely on making ChatGPT better.

Code Red isn’t a crisis signal; it’s a focusing mechanism. It gives the company permission to drop other projects and converge on what matters: latency, reliability, model quality, personalization. The output was GPT-5.3 (great for everyday users) and 5.4 (workhorse for knowledge work).

The lasting artifact isn’t any specific product improvement; it’s focus itself. In a company with as many opportunities as OpenAI approaching AGI, the risk of death is distraction.

“If you were to premortem why a company like OpenAI does not achieve its mission, it’s probably focus.”

Peter Steinberger and the OpenClaw Inspiration

Turley is visibly excited about Peter Steinberger (of OpenClaw fame) joining OpenAI. OpenClaw demonstrated a vision the ChatGPT team had been pursuing in various forms: a fully embodied AI that exists across different UIs, maintains state, takes actions, and interacts in short, natural message bursts rather than formal prompts.

The codebase name for ChatGPT’s server? “SA Server,” short for Super Assistant Server. Proof, Turley says, that this was always the vision.

Closing Notes

A few observations worth sitting with:

• The best metric is invisible. Turley jokes about “incremental hours of sleep” as the real north star, but he’s not entirely joking. The product team genuinely discusses how to measure self-actualization. When your metric is whether someone’s life got better, you end up making different decisions than when you’re optimizing for engagement.

• The Mac OS mental model. Turley aspires to progressive disclosure: magical simplicity for casual users, terminal-level power for advanced ones. Both extremes teach the team different things. Casual users force you to nail the interface; power users do your product discovery for you.

• Curiosity as the permanent skill. When asked what students should focus on, Turley doesn’t say coding or math. He says curiosity, because “if the machine can answer all your questions, you better have good questions.” And writing, because expressing what you want to a machine requires precise thinking.

• Timing is the hardest prediction. Turley can see what form factors are coming but can’t predict when. His planning horizon: “anything between three months and eventually is difficult.”