2025 - 16m tokens used on openrouter and 8+ Claude projects

The creator of Claude Code, which itself was a hackathon project to begin with, has this month said that he hasn’t committed a single code change in the last 30 days without using Claude. Obviously there’s an incentive to saying this. However there’s a weird cognitive dissonance going on in software right now, where a few luddites are stuck in a world where they believe LLMS are just statistical copy and paste engines (regardless of MoE, MCP servers, Caches, Agents and various other improvements in under a year), and to be avoided. Perhaps that’s a result of using ChatGPT/Grok and CoPilot (although Grok is really good with Go).

The last personal project I wrote that wasn’t Claude-first was tinycity, I’ve created about 8+ projects that would’ve normally taken weeks to write, feeding them a template C# project when they’re C#:

• https://github.com/yetanotherchris/consensus
• https://psychology-notes.yetanotherchris.dev/ (All the tools: Javascript + HTML)
• https://github.com/yetanotherchris/openwebui-importer (Python)
• https://github.com/yetanotherchris/crontab
• https://github.com/yetanotherchris/twowords (HTML)
• https://github.com/yetanotherchris/text-template (Go, I nudged it towards ANTLR)
• https://github.com/yetanotherchris/safeincloudviewer
• https://github.com/yetanotherchris/letmein (React upgrade)
• https://github.com/yetanotherchris/ask-llm

I’m hoping LLMS stop padding responses in 2026 to burn output tokens, which are 5x the price of input tokens. Claude has become noticeably less verbose and error-prone in the last 3 months though. At night time and with big contexts, it does run into problems, but overall I think the estimation of 1 year’s of work is probably close with my openrouter usage; most of its usage has been with consensus and Openweb UI.

I think one big advantage in Claude’s hybrid-planning mode is knowing where it’s going wrong. However if you don’t know what it’s doing, which a few times I haven’t, it can go down a big rabbit hole. It did this with Python and ML on an unpublished project, wasting around 5 evenings. But failing fast might be a big advantage of Claude or Grok coding in the next year - being able to trial and error 10x faster.

I'm Chris Small, a software engineer working in London. This is my tech blog. Find out more about me via Github, Stackoverflow, Resume