Leland Char’s Post

I changed my AI coding stack again today. Again. As of 7:00 p.m. today, I was simultaneously paying for four AI tool subscriptions: - Gemini / Antigravity - Claude Max - Cursor Pro - OpenAI / Codex The monthly total was roughly $750 across $200, $250, $100, and $200 plans. That sounds ridiculous. But I have stopped thinking about this as software spend and started thinking about it as throughput. When I am building SayMei after work, the bottleneck is no longer whether one model can help me write a component. It is how many long-running tasks I can safely keep moving at once. Tonight I had four agents running in parallel for more than 30 minutes at a time: - One was doing a full code cleanup and refactor, paying down leftovers from a redesign that started months ago. - One was connecting the unfinished pieces of our flashcard system. - One was building an agentic calibration loop that adjusts to a user's level at the end of every lesson. - One was porting my MCPs, skills, and CLI workflows from Claude Code into Codex. In a few hours of post-work coding, the repo moved by over 26,000 lines and counting. That number is messy. Some is scaffolding, some is refactor churn, and some will get deleted. But the magnitude is the point. The unit of work has changed. A year ago, I thought about AI coding as "ask a model for help, then go back down into the code layer to play with parameters." Tonight felt different. I was operating more at the orchestrating large projects layer: describing work, having an agent split it across agents, checking taste, and deciding what was allowed to ship. Over the last 20 days, Gemini 3.1 Pro in Antigravity was my default because it was fast and surprisingly strong for front-end design iteration. When last Thursday Opus 4.7 and Claude Design changed the equation for long-running implementation and design-system work, I moved more of my workflow into Cursor Pro and Claude Max. Today, after seeing what the new Codex workflow could do, I cancelled an Ultra subscription, restarted my OpenAI plan, downloaded Codex, ported over my skills and MCPs, and started using it immediately. That is the part of AI tooling that still feels under-discussed. The edge is not having "the best model." The edge is building a workflow where switching models is cheap, parallel execution is normal, and your taste still decides what ships. For solo builders working on the weekends, it changes what one person can realistically build after dinner. I do not think everyone should spend $750/month on AI tools. But I do think serious builders should treat their AI stack like a living system, not a fixed preference. #DeveloperTools