How I Coding (Oct 2025 Edition) (xuanwo.io)

Anthropic’s Claude Sonnet 4.5 landed shortly after the author’s September update, and after a month of use the main takeaway is conceptual: “slow is fast.” The author argues attention — not raw model latency — is the scarce resource for coding. Faster, cheaper models like Sonnet 4.5 may feel productive but can drain attention by producing flaky, error-prone output you must constantly fix; higher-quality, slower models (e.g., Opus 4.1 in their experience) free cognitive bandwidth and end up accelerating real progress. They also note diminishing marginal improvements between model releases: Sonnet 4.5 is an incremental gains-over-cost tradeoff rather than a clear SOTA leap. Practical experiments with Codex (Cloud, Slack, GitHub) reveal limits: Cloud Codex is useful for exploratory queries about a codebase but struggles with complex environment setup and has constrained runtimes that make builds (especially Rust) painfully slow. Slack integration is premature — notifications only, no PR submission or conversational follow-up — while GitHub Codex reviews are surprisingly valuable, catching real bugs and leaving succinct approvals rather than nitpicks. The broader industry trend is subscription pricing and consolidation around a few SOTA providers (OpenAI, Anthropic, Google); the author advises developers to pick a top-tier model, reevaluate every 1–3 months, and spend most attention on language, tools, and product work rather than daily AI news.