Welcome to the AI Brief from the YPO Technology Network. I'm Stephen Forte. So, a week ago today, something happened that almost nobody at the C-level was told about, and I think it matters. OpenAI swapped the default model inside ChatGPT, not announced from a stage, not rolled out behind a feature flag, not previewed at a conference. They just changed it. On Tuesday, May 5th, GPT 5.5 Instant became the default. GPT 5.3 Instant got moved to the back of the cabinet with a 90-day timer on it for paid users. Most of your companies use ChatGPT in some form. Some of you have it sanctioned. Some of you have it in shadow. Either way, last Tuesday afternoon, your sensitive workflows ran on a different model than they did Tuesday morning. The hallucination profile on legal output, on medical references, on financial summaries all changed. And nobody asked your permission. Nobody briefed your CIO. Nobody put a memo on your desk. If that bothers you, good. If it does not bother you, I would gently suggest. It probably should. So today, we are going to talk about it. By the end of this episode, you will have a clean read on what actually changed, why most enterprises did not notice, and a one-page audit you can run this week that will tell you exactly which side of the AI economic divide your company is on. Spoiler, the divide is bigger than you think, and the test is simpler than you would expect. Let's set the table. On May 5th, OpenAI released GPT 5.5 Instant. It replaced GPT 5.3 Instant as the default chat GPT model. Paid users still have three months of access to 5.3 if they go hunting for it in a settings menu. The free user got the swap Tuesday morning without a pop-up. OpenAI's headline claim GPT 5.5 instant reduces hallucination in sensitive domains, specifically law, medical, financial outputs. The benchmarks they cite are real benchmarks. The improvement is real. Honest credit where it is due. Reducing hallucination on legal and medical content is the single most important thing a frontier lab can ship right now. It is also the least glamorous thing they can ship. There is no demo video for the model is wrong less often. OpenAI did the boring work. That matters. Two days after the default swap, they shipped something else worth noting. GPT 5.5 Cyber, a specialized variant for vetted cybersecurity teams limited rollout gated access application required We are going to come back to that one because it is the inverse pattern But the headline event of the week is not the cybersecurity model It is the silent default rotation because that is the one that affected millions of enterprise users And almost nobody at the C-level was told. Here is something you can do in the next five minutes. Walk over to whoever runs your AI strategy or pick up the phone or send the Slack. Ask three questions. The first, which model is our team actually using in chat GPT right now? The second, when did that model last change? And what was it before? The third, did anyone evaluate the new model against our existing workflows before the swap? If they can answer all three quickly and crisply, you are in good shape. Stop listening. Go do something more interesting with your afternoon. If they cannot answer any of them, and statistically, four out of five of you cannot, we have a structural problem to talk about. And I do not mean a technology problem. I mean a posture problem. Here's the lens that makes the rest of this make sense. There are two ways a company can relate to artificial intelligence right now. You can consume it as a feature or you can run it as infrastructure. Consuming AI as a feature looks like this. Somebody signs an enterprise contract with OpenAI or Anthropic or whoever. The IT team turns it on. People log in. They use the chat box. Whatever model is behind the box on any given day is the model they use. The vendor changes something on Tuesday. The features change on Tuesday. Nobody owns the chain between the vendor made a change and our outputs changed because there is no chain. There is a chat box. Running AI as infrastructure looks like this. You know which model your team is on. You know which version. You have an evaluation harness, three or four representative tasks from your business run automatically with a benchmark of what good output looks like. When the vendor pushes a new model, your harness runs the new model against your tasks before your team does. You make a decision, promote it, hold it, or roll back to the prior version. You treat the model the way you treat any other piece of production infrastructure, with version control, with change management, with a roadmap. The companies in the first camp last Tuesday were running on a different model by lunchtime and did not know it. The companies in the second camp ran their harness Tuesday afternoon, confirmed the new model, performed equal or better on their sensitive workflows and promoted it Tuesday evening or they flagged a regression and held the prior version. Either way, they made a decision. That is not a difference in tools. That is a difference in posture. Last month PWC published an AI performance study based on 200 executives across 39 countries The headline finding 74 of AI economic value is being captured by 20 of companies Read that one more time Three quarters of the economic gains one fifth of the companies. That is not a normal distribution. That is a power law. And power laws and business outcomes do not come from differences in adoption. Every company in the survey is adopting AI. They are not separated by whether they use it. They are separated by how they use it. PwC's finding is that the leaders are using AI for growth. New products, new revenue surfaces, new customer experiences. The followers are using it for productivity. Same outputs, fewer hours, both produce value. But growth compounds. Productivity decays the moment your competitor does the same thing. Layer that finding on top of the GPT 5.5 swap and a picture emerges. The companies running AI as infrastructure can pursue growth use cases because they have the discipline to put a model into production with confidence. The companies consuming AI as a feature are stuck on productivity use cases because anything bigger requires a level of trust the chat box cannot give them. They are not behind by accident. They are behind by posture. The 74% gap is not measuring adoption. It is measuring the difference between companies that watch their model stack and companies that get watched by it. Three concrete moves. None of them are exotic. All of them are unsexy, which is why most companies have not done them. The first, they version control their model stack the way they version control their code. When OpenAI changes the default, somebody on their team knows about it within 24 hours. Not because they have a fancy monitoring tool, because they actually look. There is a person whose job includes know what model we are on. The second, they run an evaluation harness on their sensitive workflows. Five to ten representative tasks run weekly or whenever a model changes. Output goes to a dashboard a human actually reads. If you cannot afford an in-house version, vendors will sell you one. If you can afford an in-house version, build it. The investment is two engineers for one quarter. The return is knowing whether your AI got better or worse this week. The third, they pick growth use cases on purpose, not by accident, not let us see what AI can do for the support team. Instead, we're going to use AI to launch a new product line we could not have staffed for two years ago. That is the move PWC's leaders are making. It looks like product strategy because it is product strategy AI is the unlock not the goal One more thing worth noting because it is the canary Two days after the default swap OpenAI launched GPT 5 Cyber specialized variant vetted teams only, application required, limited preview, notice the pattern. The general purpose model rotates silently underneath you. The specialty model is gated, version locked, and rolled out with explicit access controls. That is the same pattern the Pentagon got. Contract level scoping, named restrictions, control distribution. We talked about that yesterday. What that tells you is the mature side of AI procurement is starting to look like every other mature procurement category. Generic equals commodity equals silent updates. Specialty equals contract equals version control. The vendors already know this. They are just waiting for their customers to catch up. If your enterprise AI is all generic chat box access today, you are buying commodity, which is fine. Commodities are useful, but you should know that is what you are buying. And you should not be surprised when the vendor treats it like one. The PwC 7420 split is not a snapshot of where companies are today. It is a forecast of where companies are going to be next year and the year after. And the year after that, power laws widen. They do not converge. The companies running AI as infrastructure get to compound. Better posture leads to better evaluation, leads to better use cases, leads to better outcomes, leads to better posture. The companies consuming AI as a feature are running on whatever the vendor decided to ship last Tuesday, and they're going to keep being surprised by it. The fix is unglamorous. There is no platform to buy. That turns you into the 20%. There is no McKinsey deck that transforms your posture. There is one engineer, one evaluation harness, and one person whose job description includes, tell me when the model changed. That is it. That is the gap. OpenAI changed the model last Tuesday. Your company probably did not notice. That is not a technology problem. That is the whole problem. The 74% of economic value being captured by 20% of companies is the macro version of the same story. The companies that noticed are running infrastructure. are the companies that did not are running a chat box and hoping. Run the three-question test this week. Whichever side of it you land on, at least you will know. That is the YPO Tech Network AI Brief for Wednesday, May 13th. I am Stephen Forte. If this was useful, send it to a fellow member. I will be back tomorrow with more. Until then, stay sharp.
