When the product was a flat monthly fee, the amount of errors and wasted tokens by the models was somewhat acceptable since even though it wasted time, it didn’t also waste money. Now, the users pay for all the errors, and poorly tuned models not following even basic instructions.
What is being done to address this?