GLM 5.2 in line with Opus 4.6, Codex 5.4
Z.ai’s GLM 5.2 model feels like it is in-line with Claude Opus 4.6 and Codex 5.4. Thus far I have tested GLM 5.2 on some html and css files for personal Websites. So far, so good.
I spent approximately $1 to review an 1,100 line file 3 or 4 times. I wish I recorded all of the details, but I forgot to do so. I can tell you that in my most recent coding session, 57K tokens were sent, 15K were received, cost was $0.10 for the last message in the session, the total session cost was $0.28, probably one-tenth the cost of Opus 4.6.
Recall Anthropic’s latest models - Mythos and Fable 5 - were yanked several weeks ago, three days after launch. Anthropic’s current model in the public domain is Opus 4.8. I like 4.8 better than GLM 5.2, which makes sense as GLM is likely 6 months behind the latest Anthropic models.
OpenAI’s current model in the public domain is GPT 5.5, which I use inside of Codex terminal. GLM 5.2 feels like the prior model, GPT 5.4. OpenAI’s GPT 5.6 model will release in mid-July once the Trump Admin finishes reviewing OpenAI’s limited customer list (100 enterprise firms will have access to GPT 5.6).
The big difference between GLM 5.2 and Opus/Mythos/Fable and GPT is that GLM truly is a model. It has limited tooling (if any), whereas Anthropic’s “models” and OpenAI’s “models” are more than models, they are AI harnesses - models with lots of tooling around them.
Therefore, if we were to bring GLM into Kilby in order to offer Kilby at an ultra-low price, we will have to build the generic tool layer (read PDFs, ingest PDFs, read web pages, etc.), that sits between the model and our proprietary Investment Management-specific tools and datasets. I need to learn more about what type of effort would be required to build that tooling and to maintain it.
I will report here as I use the open source models and learn as I go.