by Robert Jackson, AI.RDN+ Skills Developer
2nd June 2026
Two weeks ago, my car flagged an issue with the front-left indicator. I did a few checks, and sure enough, the bulb had gone. I quickly used Gemini to check which replacement bulb I needed, giving it clear instructions about the make, model and trim level of my car. It replied that my vehicle had a complex LED array instead of a bulb, so it would need to go back to the manufacturer. I queried this, providing a photo of the light structure and what was obviously a bulb. It replied that it could understand why I would think that it was a bulb, but it was definitely a complex LED array, and the orange bulb I was seeing was definitely not a bulb, but LEDs behind an orange sheet. I opened up the light housing, removed the bulb, and sent another photo to Gemini. It finally conceded that my car did use a bulb and that it was thinking about a different model. A quick, cheap fix could have been a lot more expensive had I taken it at its word.
It would be easy to chalk this up to a random quirk that all chatbots have now and then, some more sinister than others (looking at you, Grok), but if you use a model even fairly regularly, you might have noticed they seem to be getting more frustrating to use. And it’s a concept we’re all familiar with.
Picture it. It’s Christmas Day, everyone around you is full of festive cheer, paper hats askew as they play a game or fall asleep in front of the TV. But you’re in the corner, angrily telling anyone who will listen that the Quality Street tub got smaller again. And you’d be right! Christmas 2025, the tub shed another 50 grams, taking it from 600g to a measly 550g, with prices rising by one pound on average. A 2.5kg tin from the 1980s can comfortably swallow four of today’s plastic tubs whole.
Shrinkflation. We all know it, we all hate it. Pay the same or more, get less. And now it appears to be happening with consumer AI.
The best example comes from Anthropic, with some engineering teams having reported that Claude has become less consistent in how it approaches problems. In one widely shared write-up from an AI lead at AMD, their team described the model reading less context before acting, rewriting entire files instead of making targeted changes, and skipping verification steps.
Elsewhere, there are indications that OpenAI is dynamically deciding whether your prompt gets processed by a high-end or low-end model based on real-time server load, which can impact performance. GitHub Copilot took steps to move away from the ‘unlimited’ usage models it was offering to a credit-based system. Basic text generation costs almost nothing, but if you use a high-end reasoning model for a multi-step purpose, it can consume dramatically more credits per response. Users report burning through their allowances by mid-month.
Pricing structures are also changing. Cheaper access, such as OpenAI’s ‘Go’ plan, has started testing advertisements, and most platforms are now pivoting to $100+ plans and credit-purchasing for the ‘all-you-can-eat’ access that a cheaper plan originally offered. I regularly get told to upgrade Gemini to an eye-watering $249 Ultra plan.
The why is simple enough; complex queries are expensive, and companies don’t consider their current business models sustainable (or more cynically, they’re just not making enough money for shareholders). We know that using an LLM is more expensive than a Google search, but as models become more sophisticated and agentic workflows are introduced, costs are rising astronomically, with some estimates suggesting that interactions with deep-reasoning models could cost up to $5 for a complex output.
So for now, the consumer will suffer. AI chatbots may not be shrinking in raw capability, but access to that capability increasingly is, unless you’re prepared to pay a lot for it. Enshittification by way of a digital chocolate tub.
*****************************
If you would like to share your thoughts on the article or share the article itself on LinkedIn, please click here.