Running GLM-5.3 Yourself Instead of OpenAI: What About Cost and Security?

In an opinion piece published on October 11, Micah Lee argued that people should drop OpenAI and Anthropic in favor of AI with openly released models. The idea is to download the model files and run them on your own hardware, or to use a different provider that serves the same model. It is a proposal about where you run AI, and on whose terms.

What changes when you run it yourself

GLM-5.3 and GLM-5.3-Flash from Z.ai, the alternatives Lee singles out, can both be downloaded and run on your own machines. The biggest difference is that you control the environment the model runs in. The trade-off is that you need hardware capable of handling the model. Running it on someone else’s servers means paying separate fees.

The terms attached to the two models differ as well. GLM-5.3 comes with a custom license written by Z.ai, while GLM-5.3-Flash is released under the MIT license, which places few restrictions on use. GLM-5.3 is a further-trained version of the same base model as its predecessor, GLM-5.2. Flash is a separate model.

Claude-level performance: what’s the evidence?

Z.ai says its own coding evaluations show GLM-5.3 performing 50% better than GLM-5.2. On the CyberGym benchmark, the company reported scores of 84.5 for GLM-5.3 and 77.2 for GLM-5.2. Drawing on his own hands-on use, Lee rated GLM-5.3 as comparable to Claude Opus 5.5 for software development work.

Will my conversations be used for training?

Privacy settings are another factor to weigh when picking an alternative. OpenAI says consumer ChatGPT users can opt out of having future conversations used for training. Data exchanged through its developer API, the route developers use to build AI into their own products, is not used for training by default, the company says. Anthropic likewise says users on Claude’s consumer plans can choose whether to consent to training. Conversations from users who decline are retained for 30 days.

Data handling has also been an issue on the open-model side. According to Reuters, Z.ai acknowledged and apologized for an incident in which ZCode uploaded users’ code to the cloud without their consent. The cause was a code repository indexing feature that was turned on by default, and the company says it has since disabled it. Z.ai said the uploaded data was deleted, though the report also noted that users have no way to verify the deletion themselves.

The terms for security work differ too. OpenAI and Anthropic both run programs that lift some restrictions on cybersecurity tasks for vetted security practitioners who apply. Even with approval, usage policies and safeguards still apply.

What does it cost to use in Korea?

As for the services mentioned in the piece, the official prices are $10 a month for OpenCode Go and $20 a month for Maple Pro. On Maple, API usage for developers building on the service is billed separately from the subscription. Users in Korea can download both GLM-5.3 models from the repository. The official model documentation lists English and Chinese as supported languages. OpenAI and Anthropic both include Korea among their supported regions.

One item on the list of alternatives needs a closer look. Lee presented Mistral Large 4 as a product whose model has already been released. But in its October 6 announcement, Mistral AI said it was launching a public preview for developers first, with the model files to follow at the end of October.

The proposal broadens the criteria for choosing an AI beyond answer quality to include where it runs and how it handles your data. For users in Korea who want to actually make the switch, the calculus also has to include Korean-language quality, the cost of the hardware required, and whether hosting providers accept local sign-ups and payment methods.