Anthropic Squeezed Between OpenAI and Chinese Models

This reveals that Anthropic’s biggest risk is no longer building a better model, it is holding pricing power when buyers can now switch between a stronger OpenAI stack on one side and much cheaper open weight Chinese alternatives on the other. In coding and agent workflows, customers care about benchmark wins, latency, reliability, and cost per token. Once those gaps narrow, usage can move fast because the workflow already lives behind routers, APIs, and interchangeable agent harnesses.1 2 3 4 5

• OpenAI’s comeback was product and distribution driven. By April 2026, OpenAI had redirected GPUs toward coding and enterprise, pushed Codex from about 1M to 4M weekly users in a month, and bundled coding plus computer use into ChatGPT, AWS Bedrock, global SIs, and its direct enterprise sales motion.1 4 5
• The Chinese pressure is mostly an economics story. Moonshot and DeepSeek helped normalize frontier grade open weight models that customers or inference partners can host themselves, cutting dependence on Anthropic and OpenAI while giving app companies tighter control over margin, latency, and model behavior.2 3
• That matters most in coding because the heavy spenders are not casual chat users. They are companies like Cursor, Harvey, and internal enterprise agent teams that can swap the model under the hood, fine tune around it, and route workloads to whatever gives the best output per dollar.2 3 4

Going forward, the frontier model market looks less like winner take all software and more like a barbell. OpenAI is pushing to own premium closed model workflows through distribution and compute, while Chinese open weight labs compress pricing from below. Anthropic’s path is to keep Claude Code and adjacent products sticky enough that customers buy the workflow, not just the tokens.1 2 4 5