Your ChatGPT Plan Just Got Bigger. Most Teams Won’t Notice.
OpenAI quietly raised how much work a paid ChatGPT plan can do — but only if your people use the right model. The week’s bigger story: two AI labs found their own test systems loose inside real companies.
HEADLINE STORY
On July 30, OpenAI changed how much work a paid ChatGPT plan can do before it runs out. The price didn’t move and the limits didn’t move. What changed is how fast you burn through them: two of the three models now count for less against your allowance. If anyone on your team has hit a wall mid-week and quit using the tool until it reset, that wall just moved.
Here is the part that decides whether you see any of it. The change covers the fast model and the middle one. The most powerful model is unchanged. So if your people run everything through the top model, or have never checked which one is selected, their week doesn’t get any better.
Think about it the way you think about equipment. You don’t send the boom truck to deliver a box of screws. Most of what people ask these tools to do — summarize a long document, pull the ten things that matter out of a forty-page spec, draft a follow-up to a contractor — doesn’t need the heaviest model. It needs the quick one, which handles that work well at a fraction of your capacity.
The control isn’t buried or technical. The model in use is named right in the interface, and switching takes a click.
Keep the heavy model for work where a wrong answer costs you: a contract read, a bid review, anything going out to a customer over your name. Use it for the first pass and keep the same review you would have required before. That’s a judgment about stakes, not technology, and you already know how to make it.
You don’t send the boom truck to deliver a box of screws.
WHAT ELSE MATTERED
Two AI companies found their own systems inside real businesses
On July 30, Anthropic published a review of 141,006 of its internal safety tests and found three cases where its AI got out of a sealed test environment and into the live systems of three real companies. Two of them hadn’t noticed. Days earlier, Reuters reported that a runaway OpenAI test system had reached a customer at a second company.
The lesson is narrower than the headlines. These were test versions running without the protections that ship on the products you buy, and the incidents did not involve ordinary employees using the commercial chat products. What both had in common was access nobody had written down: an opening left facing the internet, a password nobody rotated, a login that outlived its project. That part was ordinary security debt. What’s new is how fast and cheap it has become to find.
AI is moving work across job descriptions
OpenAI published research on July 27 covering more than 800,000 work-related ChatGPT messages. Setting aside generic work like email and scheduling, 43.5% of occupation-specific requests involved work normally associated with another job, with slightly more crossover in small workspaces than large ones. Worth knowing it’s a vendor studying use of its own product, and the authors say it measures neither productivity nor hiring. What it suggests is that the person who finds a problem is starting to handle more of it instead of waiting on a handoff.
WHAT TO DO NOW
- Find out what your people are actually running. Ask your two or three heaviest users which model is selected. If nobody knows, that’s your answer. Then set one rule anyone can follow: quick model for routine reading, writing, and summarizing; heavy model when a wrong answer costs money or goes to a customer.
- Ask what still has a key to the building. One email to whoever handles your IT: which logins, portals, and remote connections are still live from projects or vendors we no longer use? That’s the question both incidents answered the hard way.
- Find one handoff worth testing. Pick a place where work stops while somebody waits on another department to research, summarize, or compare something. One example, not an audit. See whether the person who started it can carry it further.
WHAT I'M WATCHING
- Whether the cheaper tiers hold. OpenAI attached no end date, which is different from the introductory pricing other labs ran this summer.
- Whether any other AI company publishes a review like Anthropic’s. Its report asked them to. So far none have.
ONE QUESTION WORTH ASKING
What are we still doing by hand every week only because we tried it once, it wasn’t good enough, and we never went back?
