Max Woolf traces June–October 2026 pricing moves by OpenAI and Anthropic. Examples: GPT-5.6 Luna cut 80% in July to 1.20/4/0.20 with less compute to serve than Opus 5.
Thesis: tit-for-tat competition plus open-weight pressure (DeepSeek Flash V4.1 on OpenRouter with 29 providers, 2% cache-read ratio) and Kimi K3.1 Pro nearing Opus 5.5 benchmarks at half price when weights drop. Models behave as commodities with low switching cost via OpenRouter/OpenCode/Zed.
IPO angle: both firms rumored near IPO (Anthropic months away, OpenAI 2027); price cuts despite losses may reflect healthy inference margins (~80% cited with caveats) and need to grow usage. Subscription caps (Codex $200 tier usage cut 50% when GPT-6.1 Sol launched) partially offset API savings.
Vercel AI gateway data: token costs fell to 43.3% of April 2026 while aggregate spend rose because capability improved faster than prices fell. Author expects open inference providers to keep pricing near marginal cost, limiting sudden price hikes feared by AI skeptics.