After two years of widespread adoption of AI assistants in the development industry, a reversal has occurred. Companies began to cut budgets for tools like GitHub Copilot and Claude, shifting corporate subscriptions to more modest rates with tighter limits. As a result, the monthly quota of tokens ends within one-and-a-half weeks, and tasks that were recently closed in minutes are again stretched out by hours.

The main reason for the setback is economic. On large and legacy code bases, a $100-$200 personal subscription can be completed in just one business day, and an experienced engineer can burn up to $1,000 per day at full operation. When the bill is tied to actual consumption of tokens, spending becomes unpredictable and the finance department responds by cutting budgets. Critics of this approach consider it shortsighted, as losing development speed is more expensive than saving subscriptions.

There is also a bright side to the return to manual work: developers find that their skills of analysis, debugging, and programming have not gone away. Many point out that they once again feel full control over the architecture, whereas the assistant sometimes made incorrect assumptions about critical cases. For those in the restricted mode, it is better to manage the tokens: use expensive models like Opus for planning and analysis, and give execution cheaper Sonnet and Haiku. The biggest benefit of a neurose is when reading code, breaking an unfamiliar codebase, and summarizing documentation rather than writing new code.