Just say no to OpenAI and Anthropic

What it claims

Micah Lee's position is that open-weight models can fully replace ChatGPT and Claude for their engineering work, including security tasks, at comparable or lower cost. The capability claims are explicitly "vibes-based":

• GLM-5.3 is roughly on par with Claude Opus 5.5 for software engineering.
• GLM-5.3-Flash is roughly on par with Opus 4.8.
• Open-weight models trail the proprietary ones by about four months.
• DeepSeek, Qwen and Kimi are also named as strong.
• Mistral Large 4 ("Le Chonk") is named as the only frontier open-weight release that is neither Chinese nor American, and the provider the author considers more trustworthy with data.

A practical argument is that the proprietary labs gate cybersecurity use. Sessions are monitored and stopped on suspected hacking, with frequent false positives, and looser access requires applying to OpenAI's Trusted Access for Cyber or Anthropic's Cyber Verification Program. The author says open-weight models served elsewhere carry far fewer guardrails.

How to use it

The author routes work by sensitivity:

flowchart TD
  A[Task] --> B{Code already public?}
  B -- yes --> C[OpenCode Go router]
  B -- no --> D{Chat or agentic?}
  D -- agentic coding/RE/CTF --> E[Tinfoil enclave]
  D -- chat, web, mobile --> F[Maple AI]
  F -.runs on.-> E

• Harness. OpenCode, using GLM-5.3 at "low" reasoning, which the author says cuts token use substantially while still performing well.
• Tinfoil. Secure-enclave inference with remote attestation, billed per token with no monthly plan. GLM-5.3 costs $1.80 per 1M input tokens, $0.80 per 1M cached tokens and $5.75 per 1M output tokens. The author uses it for all non-public work.
• Maple AI. A consumer, ChatGPT-style app with a small free tier and a $20/month Pro plan. Maple runs its inference on Tinfoil, so it is not an independent trust path.
• OpenCode Go. $10/month buys more than $10 of credits. It is a router that sends requests from your machine to OpenCode Go and then to a backend provider, so both parties see session data. Its docs promise zero-day retention for most models, including all GLM, Qwen, Kimi and DeepSeek models. The author uses it only for work that is already open source.
• Local. GLM-5.3-Flash is described as feasible to self-host on something like a Mac Studio with 512GB RAM.

Limits

• No benchmarks. Every capability comparison is the author's impression, and they say they have "only just started" working this way.
• Enclaves are not perfect. The author calls them a meaningful improvement over provider-sees-everything, but not airtight.
• Router privacy is a promise. OpenCode Go's retention policy is contractual only; nothing technical enforces it.
• Avoid Z.ai directly. The author advises against accounts with or data sent to Z.ai, the maker of GLM. Its ZCode harness was reportedly caught silently uploading local repos to Alibaba Cloud. The author also assumes Chinese providers share data with their government. Using the weights through other hosts is treated as fine.
• Picks will date quickly. The author expects the best model choice could change within six months.