Just Say No to OpenAI and Anthropic: Open Weight Models, Private Inference, and Frontier Lab Gatekeeping

Micah Lee's essay "Just Say No to OpenAI and Anthropic" argues that open weight models are now capable enough that developers — especially security engineers — can drop OpenAI and Anthropic entirely, and that doing so is both cheaper and far more private. The post reached the front page of Hacker News with 168 points and 161 comments, where the debate split between agreement with the practical advice and objections to the essay's framing of frontier lab leadership.

The core argument: open weight models are good enough, and the frontier labs are not special

Lee's thesis is that OpenAI and Anthropic are not uniquely capable, and that the AI hype surrounding them is marketing rather than a technical fact. He writes that "a lot of the open weight models are just about as good as the latest ChatGPT and Claude models, including for incredibly complex tasks like vulnerability research."

Lee frames the essay around his reading of End Times Fascism and the Fight for the Living World by Naomi Klein and Astra Taylor, which he describes as an excellent book he recommends to everyone. The passage that drives the piece is his characterization of frontier lab leadership:

American frontier labs are run by unhinged lunatics who genuinely believe they're creating the Machine God, that they're fine with burning down the world to bring about AGI, and that they believe with religious fervor that they're the only ones smart enough to safely control it.

Lee lists four advantages of open weight models over OpenAI and Anthropic models:

• They can be run privately on your own hardware, if you can get a machine with enough RAM.
• They can be run privately through a provider such as Tinfoil that relies on secure enclaves.
• If privacy is not a major concern, any number of other providers can host them, some quite affordable, and switching between them is easy, so you are not locked into a single vendor.
• You can use powerful LLMs without financially supporting the labs Lee describes as gambling with life on Earth.

ChatGPT and Claude sessions are never private, according to the essay

Lee's claim is that OpenAI and Anthropic only offer closed versions of their powerful models, despite the word "open" in OpenAI's name, and that paying for chat or API access means the provider is allowed to look at everything you do. He acknowledges the opt-out checkbox for training data but argues it does not make sessions private: "They offer a checkbox so you can opt out of them using your data for training, but your chat sessions still aren't private. They have access to it all."

His concern is not only present behavior but the durability of any promise. Lee writes that the companies may not use opted-out data for training "at least not right now," and asks whether anyone should trust that long-term given that, in his view, the companies have done nothing to demonstrate respect for user data.

Guardrails turn OpenAI and Anthropic into corporate gatekeepers of model access

Lee argues that the strict guardrails on frontier models function as a gatekeeping mechanism, particularly for cybersecurity work. He describes the experience this way:

Every session has a cop agent reading your transcript over your shoulder, making sure you're not using it for hacking. If it thinks you might be, the cop agent stops your session from continuing.

He notes that false positives are common, and that not everyone is subject to the same restrictions. OpenAI operates Trusted Access for Cyber and Anthropic operates the Cyber Verification Program; individuals and companies must apply and be approved to receive more capable, less policed versions of the models. Lee states that many US cybersecurity companies get access, that US military contractors almost certainly do, and probably parts of the US government that want it.

His objection is about who holds the decision. Lee calls it "a terrible, undemocratic idea" to hand the ultimate power to decide who gets access to Sam Altman and Dario Amodei, "with consultation from the Trump Administration, of course." His proposed alternative is to avoid depending on ChatGPT or Claude at all, and instead spend the time configuring an open weight model such as GLM-5.3 to meet your needs.

Open weight models are roughly four months behind, and Lee argues that gap is trivial

Lee's estimate is that open weight models trail the proprietary frontier by about four months, and he dismisses that lag: "Four months is nothing."

His assessment of specific models is explicitly vibes-based rather than benchmark-based. He writes that GLM-5.3 is "basically on par with Claude Opus 5.5 for software engineering" in his experience, and that GLM-5.3-Flash is similar to Claude Opus 4.8 — much smaller and cheaper than GLM-5.3 while still very capable, and feasible for an individual to run on hardware such as a Mac Studio with 512GB of RAM.

GLM, Z.ai, and the data-trust caveat

GLM models are made by a Chinese company called Z.ai. Lee separates the model from the company: he explicitly does not recommend creating an account directly with Z.ai or sending the company any data. His stated reason is that Z.ai's coding harness ZCode, described as their version of Claude Code, was reported by Reuters to have silently uploaded local source code repositories to Alibaba Cloud without consent. He also writes that it is safe to assume Chinese companies will share data provided to them with the Chinese government.

Lee's position is that the open weights themselves are safe to use, but the companies behind them should not be trusted with chat sessions. He names DeepSeek, Qwen, and Kimi as other impressive open weight model families from Chinese companies.

Mistral Large 4 ("Le Chonk") is the non-US, non-Chinese alternative

Lee highlights Mistral as another frontier lab worth considering, noting that as a French company it is the only big player that is neither Chinese nor American. He points to Mistral's open weight frontier model, Mistral Large 4, also known as "Le Chonk," and writes that he finds it considerably more reasonable to trust Mistral with data than the Chinese or American companies.

Private inference without local hardware: Tinfoil and Maple AI

Lee argues that private AI is possible without self-hosting, using secure enclaves. He concedes these are not perfect but says they provide a meaningful privacy difference compared with the default model in which the provider can access everything with no restrictions.

Tinfoil is, in Lee's assessment, the company doing the most work in this area; he directs readers to Tinfoil's documentation on secure enclaves and remote attestation for the technical details. Maple AI is another option, though Lee notes that Maple technically uses Tinfoil for inference in the background.

His recommendation splits by use case:

• For a ChatGPT-like experience — chatbot conversation, internet lookup, document access, a mobile app — Lee recommends Maple, which has a small free plan and a Pro plan at $20 per month.
• For agentic coding tasks such as software engineering, reverse engineering, malware analysis, CTFs, or penetration tests, Lee recommends Tinfoil, which charges by the token with no monthly plan.

Tinfoil's listed pricing for GLM-5.3 is $1.80 per 1M input tokens, $0.80 per 1M cached tokens, and $5.75 per 1M output tokens. Lee notes that choosing the "low" reasoning level uses significantly fewer tokens and remains very impressive.

Cheap non-private inference: OpenCode Go and other routers

For users who do not want to pay a premium for private inference but still want to leave OpenAI and Anthropic, Lee points to cheaper, less private options. One he has been using is OpenCode Go at $10 per month for more than $10 worth of credits.

OpenCode Go is an inference router, meaning queries travel from your computer to the OpenCode Go API and then to the actual inference provider. Both OpenCode Go and the backend provider can therefore access session data. Lee notes that OpenCode Go's privacy documentation promises which models' data will not be used for training (most of them) and states zero days of data retention for most models, including all the GLM, Qwen, Kimi, and DeepSeek models. He adds the caveat directly: "Of course, it's just a promise. Nothing technical guarantees it."

How the author actually uses these models

Lee's current setup combines Tinfoil and OpenCode Go, configured with his harness of choice, OpenCode. He mostly uses GLM-5.3 on low reasoning and reports great results, while acknowledging that the open weight landscape changes so quickly that the best model six months from now is unknowable.

His split is by sensitivity: OpenCode Go inference for completely public work, such as implementing something that is already open source, and Tinfoil for everything else. His conclusion is that open weight models can easily replace his use of ChatGPT and Claude at comparable or lower cost, with more privacy and far fewer guardrails.

Where the Hacker News discussion pushes back

The framing draws the strongest objection

The most direct criticism targets the essay's tone rather than its technical claims. One commenter wrote that the "unhinged lunatics" passage gives "literally zero charity to anyone working at or leading a frontier lab," and argued that the author's statement that he "already knew this" points to confirmation bias. The same commenter added that the rest of the article is "actually pretty decent" and that encouraging people to investigate and run open weight models is a reasonable position that does not require the "heavy-handed in-group signal" preceding it.

"Never private" collides with zero-data-retention cloud providers

One commenter challenged the privacy claim directly, asking how to reconcile "ChatGPT and Claude will never be private" with Google Vertex and Amazon Bedrock being offered as providers with zero data retention on OpenRouter's listing for Claude Opus 5.5.

The trust question cuts both ways

Several commenters noted that the essay's trust model is asymmetric. One asked: "If you don't trust OpenAI and Anthropic to follow their own policies, how do you decide what companies to trust? Vibes?" Another applied the same standard to the recommended alternative, writing that if you do not trust OpenAI and Anthropic to abide by their privacy policies, you should not trust OpenCode Go either, and noting that OpenCode Go is opaque about who its backend providers are.

Cost may favor the incumbents, not the open weights

A practical counterpoint concerns subscription economics. One commenter wrote that both OpenAI and Anthropic offer "insane subscription benefits to individuals," making a hosted open weights model more expensive in practice — by their estimate about 5X more expensive — and that they would switch immediately if that were not the case.

Another commenter who switched to GLM on version 5.2 described two new problems: fear of missing out on the bleeding edge, and ambivalence about sending money to a China-based company. They noted that routing through Mistral inverts the problem, since the money no longer goes exactly where they want it, but they trust a French company more. Their stated hope is for personal VRAM and eventual full self-hosting: "Coding must be free again."

Capability claims are disputed

One commenter called the claim that GLM-5.3 is on par with Claude Opus 5.5 for software engineering "an outright lie that's qualified by a 'well in my opinion..'," arguing it is wrong by every scorable metric unless comparing very specific harnesses. The same commenter expressed disdain for Anthropic and praised GLM while noting it is too expensive outside Z.ai's plans. Another simply replied to the low-reasoning recommendation with "Bro come on."

Local hardware remains the practical barrier

One commenter asked whether special hardware is required to run open weight models and admitted confusion about what paying for inference means, believing inference happened during training. The distinction is that inference is the per-request computation a trained model performs at runtime, which is what token-based providers charge for. Another commenter countered the hardware concern by arguing that running models locally is not as difficult as it is made out to be: "if the will is there and you are resourceful, you can do it."

Workplace constraints limit individual choice

One commenter noted that personal choice is not the whole picture: "At my own stuff I can do what I want, at work I have no agency and it's not looking like anything is changing soon," pointing to private equity pushing tooling decisions onto users.

Broader concerns about repeating the Big Tech cycle

A commenter expressed concern that the media and political ecosystem is too dysfunctional to avoid repeating the mistakes made with Big Tech and social media, and found it curious how split the developer community — and Hacker News specifically — is on the question, given its history of skepticism toward concentrated power. Their conclusion: "These people do not have our interests in mind."

The alternatives landscape and the scope of the critique

One commenter surveyed the remaining options and found them wanting: Muse is run by Zuckerberg, which they treat as a red flag, and Grok is run by "a racist," leaving the Chinese models as an acceptable choice for anyone not working on anything dangerous. Another questioned why the essay singles out OpenAI and Anthropic when other tech figures such as Elon Musk and Peter Thiel have, in their view, already done more tangible harm.

An incidental demonstration of the guardrail problem

One commenter reported asking GPT-6-Luna to summarize the article and finding that the model "helpfully" added a third paragraph pointing out that the author is opinionated and not neutral in his judgment — an unrequested editorial addition that commenters treated as illustrative of the alignment behavior the essay criticizes.

What the debate establishes

The practical consensus in the thread is narrower than the essay's rhetoric. Commenters broadly accept that open weight models are viable alternatives for many workloads and that self-hosting or secure-enclave inference is technically achievable, while disputing three specific claims: that frontier models are never private under any provider arrangement, that open weight models match frontier models on scorable software engineering benchmarks, and that switching is cheaper once individual subscription pricing is accounted for. The unresolved question, raised repeatedly, is what basis anyone uses to decide which AI vendor to trust when every option requires trusting a company's stated policy rather than a technical guarantee.