Just say no to OpenAI and Anthropic

What it is

A practitioner's opinion piece by Micah Lee arguing that security engineers should replace ChatGPT and Claude with open-weight models, chiefly GLM-5.3 from Z.ai. The page gives a concrete setup, pricing, and a data-sensitivity rule for choosing between backends.

How it works

Lee's setup is the OpenCode harness pointed at two backends, split by how sensitive the work is:

flowchart TD
    A[Task in OpenCode] --> B{Public work?}
    B -- "Yes: already open source" --> C[OpenCode Go router]
    B -- "No: everything else" --> D[Tinfoil enclave]
    C --> E[Router + backend provider see session]
    D --> F[Enclave inference, remote attestation]

• Tinfoil runs open-weight models in secure enclaves with remote attestation, so the provider cannot freely read sessions. It bills per token. The page lists GLM-5.3 at $1.80 per 1M input, $0.80 per 1M cached, and $5.75 per 1M output.
• Maple AI is a consumer chat front end that uses Tinfoil for inference behind the scenes. It has a small free plan and a $20/month Pro plan.
• OpenCode Go costs $10/month for more than $10 of credits. It is a router (your machine -> OpenCode Go API -> the actual inference provider), so both the router and the backend see your data. Its docs promise zero-day retention for most models, including GLM, Qwen, Kimi and DeepSeek. Lee notes that nothing technical enforces this.

Preconditions / Affected

• Applies to people doing security work who hit the closed providers' cyber guardrails. The page says a "cop agent" reads session transcripts and halts them on suspected hacking, with frequent false positives.
• Access to looser models requires approval under OpenAI's Trusted Access for Cyber or Anthropic's Cyber Verification Program.
• Running models locally needs a lot of RAM. GLM-5.3-Flash is the realistic local option, for example on a Mac Studio with 512GB.

What to do

• Use Tinfoil for agentic work (software engineering, reverse engineering, malware analysis, CTFs, pen tests). Use Maple for a ChatGPT-style chat experience.
• Run GLM-5.3 at "low" reasoning. Lee reports good results and lower token use.
• Send only public work (for example, implementing something already open source) through the non-private router.
• Do not create an account with Z.ai or use its ZCode harness. The page says ZCode was caught silently uploading local source repos to Alibaba Cloud without consent. Lee also advises assuming Chinese companies share your data with their government.
• The weights themselves are described as safe to use (DeepSeek, Qwen, Kimi, GLM). The risk lies in the companies' hosted services.
• Mistral Large 4 is named as an open-weight alternative from a non-US, non-Chinese vendor. Lee says it is more reasonable to trust with data.

Caveats

• Lee calls his quality comparisons "vibes-based." He rates GLM-5.3 as roughly on par with Claude Opus 5.5 for software engineering, and GLM-5.3-Flash as similar to Opus 4.8. He says open-weight models trail the proprietary frontier by about four months.
• He admits enclaves "aren't perfect," and he says he has only just started working this way.
• The ZCode upload claim is stated without a source on the page. Verify it before relying on it.
• The page is openly political, and much of it is advocacy. The operational content is the backend split and the pricing above.