Pattern: Coding-Model Competition Has Moved From Peak Benchmarks to Price-Performance Tiers

SpaceXAI·medium signal

The week's releases converge on cheap 'good-enough for the loop' tiers rather than top-of-leaderboard peaks: Grok 4.5 at $2/$6 (Cursor-trained), Sonnet 5 claiming ~97% of Opus 4.8 at ~60% price, and GPT-5.6 shipping explicit Terra ('balanced') and Luna ('cost-efficient') variants alongside flagship Sol. For agentic engineering — where you burn tokens across many subagent turns — the selection axis is shifting from 'smartest model' to 'lowest cost that clears the task bar.'