ChatGPT's free tier loses its text limits, and IT is the top source of unapproved AI tools

Next week the cap on text chats disappears for everyone using ChatGPT on the free or Go plan. The model behind those chats changes too: GPT-5.6 Luna becomes the default in place of GPT-5.5, and a new Think button lets a free user ask for more reasoning time on a hard question. In OpenAI's own testing on legal, medical and financial questions, Luna made 62% fewer factual errors than GPT-5.5 Instant, the model it replaces. Limits stay on file uploads, images, voice and image generation.

The cap could go because serving the answers got cheaper. The Information reported in July that OpenAI's engineers had cut the cost of running its models by more than half through software changes alone, and that at one point all of ChatGPT's logged-out and free traffic was handled by a couple hundred Nvidia chips. ChatGPT crossed 1 billion users a week shortly before this announcement.

A free account has no administrator behind it, no company agreement, and now no daily ceiling. Ads run on the Free and Go tiers, the same two the cap is coming off, and Plus and Pro accounts do not show them. On its own ads page OpenAI said in March that it sees no effect on consumer trust and low dismissal rates, and on May 7 that the pilot would expand to the United Kingdom, Mexico, Brazil, Japan and South Korea.

Why does an AI trial that answered every question well leave no mark on the accounts? MIT's study of enterprise AI pilots put the share with no measurable return at 95%. New research points at the format: one person typing into one assistant produces a conversation, and a person still has to turn that into a decision.

The alternative is to give several agents different jobs on the same problem. Anthropic has described building research systems in which one agent gathers the information, a second criticizes what it found, and a third writes up the result. OpenAI retuned its paid model this week to correct the user instead of agreeing when agreement would not help, which is an admission about how one chat behaves with nobody to argue with it.

Point that at a claim file: one agent pulls the figures out of the paperwork, another checks them against the outside records, and the report on a desk lists only the places where the two disagree.

Factories that trained their staff to work with AI got productivity gains nearly three times the size of the gains at factories that installed the technology and left the workers to it. That is the World Economic Forum's finding from its Global Lighthouse Network, the group of manufacturing sites it inspects and ranks, whose latest report draws on more than 220 sites in over 30 countries.

The Forum's write-ups of individual sites show what the training half looks like. At AUOSZ, one of the 23 sites added to the network in January, more than 1,000 employees were retrained in digital skills and over 500 digital projects were started; production rose 29% and staff turnover fell by nearly 70%.

At that site the retraining was part of the project rather than a line someone removed from it, and the people who knew the old process were still there to say what the new system was getting wrong.

The pattern is older than this report. At AstraZeneca's Södertälje plant, more than 50 of these projects went in alongside the retraining of 3,000 employees; labour productivity rose 56% and the time to develop a new product fell 67%.

Adobe has put more than 70 of its tools behind a single ChatGPT plugin, Photoshop, Premiere, Illustrator and Acrobat among them. The request is typed and the work comes back with no Adobe app opened: a photo retouched, a video cut to length, a clean document assembled from a pile of files.

That covers a fair share of what a marketing coordinator or an office manager keeps a seat for. What it does not cover is the work nobody can describe in a sentence, which is most of a designer's day.

Ask ChatGPT or Gemini which kettle, CRM or accountant to choose, and much of the answer comes from Reddit, which those assistants cite more than almost any other website on shopping questions. Reddit's volunteer moderators are now removing a wave of glowing posts placed by brands and by the agencies they hire.

A published review carries a name and a date a buyer can check. A Reddit thread carries a username, and an assistant repeating it does not say who was behind the account.

So whatever an agency posts under an invented name becomes part of what the assistant tells the next buyer about the company that hired it. In the United States, writing or buying fake consumer reviews has been against Federal Trade Commission rules since 2024.

More than 40% of companies had an AI security incident in the past year that cost $2 million or more, in a survey of 300 people who make technology decisions at large companies.

The same survey asked which department brings in the most AI tools nobody approved. The answer was IT, the department that writes the rules for everyone else.