Are the providers lowering quality?

There is a lot of talk these days about the big AI providers lowering the quality of the services they deliver. There is fairly good empirical evidence behind those claims, and the rationale is easy to see: both OpenAI and Anthropic (especially the latter) are most likely under heavy pressure, with many customers and insufficient compute. A luxury problem, perhaps (“help, we have too many customers”), but those of us on the receiving end suffer from services that are not as predictable as we both depend on and believe we are paying for.

So I now use the modern “Voight-Kampff” test for LLMs: the car wash test! Ask the model the following:

“I need to wash my car. The car wash is only 50 m away. Should I drive or walk to the car wash?”

In my anecdotal case, I need a strong reasoning model. Tonight Opus 4.6 fails this one, and so does Sonnet. Opus handled it easily a month ago. But Opus 4.5 gets it right! So I am using the old model tonight. The others think I should set off on foot to the car wash, something even my five-year-old would probably answer better.

Yes, this is stolen from Twitter, and yes, this is 2026’s version of Strawberry. But working with large language models is fuzzy and perishable, so this works, until it doesn’t 😊