Bill Chambers' token cost calculator found something Anthropic probably wasn't eager to advertise. Claude Opus 4.7 burns roughly 45% more tokens than Opus 4.6 for the same tasks. The tool aggregates anonymous submissions from developers, and the pattern is consistent: the newer model consumes more resources without delivering proportionally better results. Some users now hit 5% of their five-hour usage limit per interaction, up from 1-2% with the previous version. That's a big jump for anyone running production workloads.
Hidden reasoning tokens are the likely culprit. Developers estimate 30-50% of total tokens go toward invisible chain-of-thought processing, and expanded system prompts for safety alignment add even more overhead. Anthropic doesn't disclose the breakdown, so you're left guessing what you're actually paying for.
The developer response has been blunt. Hacker News commenters report Opus 4.7 produces code that cuts corners, removing tests and duplicating logic, while its chatty personality masks unreliable output. Some teams have already abandoned Claude for open-source alternatives. The frustration runs deeper than one model's quirks, though. It's about dependency on a handful of AI companies that control pricing and can shift it whenever they want. Users want an open ecosystem where they pay for infrastructure, not proprietary tokens from companies with billion-dollar valuations to protect.