Phase 1 Retrospective: AI Models, APIs, and the Cost of Figuring It Out

Phase 1 Retrospective: AI Models, APIs, and the Cost of Figuring It Out Date: 2026-03-13 Status: Closing Phase 1 — AI-Assisted Local Infrastructure + Grid Trading Next: 1-year project with m900 What We Closed Two projects that ran in parallel since early February 2026 are now transitioning to steady state: AI-Assisted Local Infrastructure — setting up m900 (Lenovo ThinkCentre M900 Tiny) as a persistent AI agent running OpenClaw, with memory, crons, and tool access to the real world. Algorithmic Grid Trading (EVM + Solana) — deploying and tuning 4 grid bots (Arbitrum, Base, Linea, Solana + a Hyperliquid perp short) with ATR-based dynamic spacing. Both worked. Neither worked cleanly. AI Model Comparison: What Actually Happened Tested across multiple models over ~6 weeks: Model Verdict Claude Sonnet 3.5 / 4 Best reasoning + tool use. Default choice. Claude Opus Noticeably better on complex tasks. Expensive. Google Gemini Good. Already have the GCP environment — should use more. Venice AI