Particle.news
Download on the App Store

US AI Labs Cut Mid‑Tier Prices to Counter Low‑Cost Chinese Models

Companies are using inference and infrastructure optimizations to drive higher volume and test whether bigger use can make up for thinner per‑token margins.

Overview

  • OpenAI announced large cuts on July 30 that lowered GPT‑5.6 Luna by 80% and Terra by 20% and said the moves reflected internal inference gains such as kernel rewrites, speculative decoding and prompt caching.
  • Anthropic responded by launching Claude Opus 5 at about half the cost of its prior top model and by cancelling a planned price increase for Sonnet 5 to keep mid‑tier rates competitive.
  • Usage data from OpenRouter and analysis by TD Cowen show Chinese models from DeepSeek, Moonshot and Z.ai have captured significant token share and that firms including DoorDash, Airbnb and Siemens are testing or using those lower‑cost options.
  • Early telemetry found token consumption rose faster than headline prices fell, with Luna usage spiking and short‑run revenue for some models increasing, but analysts warn this pattern may strain long‑term margins as volumes scale.
  • The market shift is driven by token billing and model routers that send tasks to the cheapest capable model, making cost‑per‑successful‑task the key comparator and giving enterprise buyers more leverage while US labs protect frontier pricing.