OpenAI just split its model into three, right as the people who ran it left

OpenAI shipped a lot this week. Three new models. A work agent. A voice upgrade. A push toward families. And at the same time, the executive who ran product and business quit. If you build on the OpenAI stack, you need to hold both of those facts at once. Let me catch you up.

One model became three, and free users got left behind

GPT-5.6 is not one thing anymore. It ships as three named tiers: Sol, Terra, and Luna. Sol is the smart one for coding and deep research. Terra is the everyday workhorse, roughly last flagship's level. Luna is cheap and fast for low-stakes tasks. The prices split too: Sol runs $5 in and $30 out per million tokens, Terra half that, Luna a fifth.

Here is the part that stings for your growth funnel. Free users get none of it. They stay on GPT-5.5. Plus and Business users only get Sol at medium effort. The top gear, Sol Pro, is locked to Pro and Enterprise. If you route users to OpenAI, you now have to think about which tier each task actually needs, because the smart one hits usage limits fast.

The bench got thin at exactly the wrong time

Fidji Simo, OpenAI's No. 2, is stepping down from her full-time role. She ran the whole business and product side, with the COO, CFO, and CPO all reporting to her. Her medical leave got longer and harder than expected, so she is going part-time advisory.

She is not the only one gone. CPO Kevin Weil left. CMO Kate Rouch left. Lightcap moved to "special projects." The person who was supposed to steer consumer product through a possible IPO is now Altman's open seat to fill. That matters because ChatGPT's growth cooled late last year and missed internal revenue targets, which is why the company is leaning harder into coding, where it still trails Anthropic.

The product is walking into your living room

OpenAI is hiring a product manager to build for families, caregivers, and older adults. The audience is shifting under them. Users 35 and up climbed to 31% of the global base, while the 18-to-24 group slipped to 29%. Nearly one in four U.S. parents used ChatGPT last quarter.

That shift brings safety weight. Parents underestimate the problem: 27% said their kid used generative AI last week, but 38% of kids said they did. OpenAI is already facing lawsuits from parents, including cases involving suicide, and has bolted on parental controls, distress routing, and a "Trusted Contact" alert. Stephen Balkam of the Family Online Safety Institute called it "safety by redesign", which is a polite way of saying it was not built for kids the first time.

The super app and the sunsetting you missed

OpenAI also launched ChatGPT Work, its long-awaited "super app" that blends the chatbot, the coding tool, and the new models into one agent that acts on your files and works in the background. It runs on Codex and GPT-5.6. On desktop, it is free for everyone.

But read the fine print. The browser inside ChatGPT Work means OpenAI is sunsetting Atlas, its standalone web browser. If your team built anything around Atlas, that is a dead end now. This is the pattern to watch: OpenAI ships fast, folds features into the flagship, and quietly retires the pieces that lost.

The deep cut

The real risk is not the price hike or the org chart. It is that the three moves point in different directions at the same time the people who set direction are leaving. Coding push, family push, super app, hardware fight with Apple, all live at once, with a thin product bench above them. Atlas dying tells you how fast a bet can get cut.

So do not marry a single OpenAI product surface. Pick the model tier per task and price it out. Keep one real fallback provider wired in, not just on a slide. And put a line in your next vendor review that names continuity risk out loud, because right now the roadmap owner is a job posting.

Three questions for your team

• Which of our features actually needs Sol, and which can run on Terra or Luna without users noticing? Put a number on the token cost before renewal.
• If OpenAI sunsets a surface we depend on, like they just did with Atlas, how many days does it take us to switch? If the answer is more than a sprint, we have a problem.
• If we serve families or minors through ChatGPT, whose job is it to check the safety controls, and have they read the actual parental control settings this month?