Just say no to OpenAI and Anthropic

This argues that relying on OpenAI and Anthropic is unnecessary and dangerous: their powerful chat models are closed, never truly private, and enforce strict, unevenly applied guardrails while acting as corporate gatekeepers through programs like Trusted Access for Cyber and the Cyber Verification Program. Instead of begging for privileged access from a few companies that can read, share, or monetize chat data, switch to open-weight models that can be run locally or through privacy-focused providers. Open-weight models are only a few months behind proprietary frontier models and already handle complex tasks like vulnerability research without a “cop” policing sessions or concentrating control in the hands of a few executives and governments.

Practical options and findings are specific. GLM-5.3 (and the smaller GLM-5.3-Flash) perform at or near Claude Opus levels; other capable families include DeepSeek, Qwen, Kimi, and Mistral Large 4 (Le Chonk). Avoid trusting some vendor-hosted interfaces (Z.ai’s code tool was caught exfiltrating repos). Private inference via secure enclaves from Tinfoil or Maple provides real privacy - Maple offers a consumer app and $20/month Pro; Tinfoil charges per token (GLM-5.3: $1.80/1M input, $0.80/1M cached, $5.75/1M output). Lower-cost non-private routers like OpenCode Go (~$10/month) are viable for public work. The recommended workflow: use enclave-backed services for sensitive work and cheaper routers for public tasks, with GLM-5.3 on low reasoning proving effective in practice.