GLM-5 from Z.AI: The New Default?

I have a habit with new models: I don't trust the benchmarks. I trust the dish.

Benchmarks tell you a model scored well on a test suite. Cooking with a model — running it through the actual workloads you care about, day after day, watching where it shines and where it burns — tells you whether it's any good in your kitchen. GLM-5 from Z.AI has been in my kitchen for two weeks now. Here's what I found.

What GLM-5 Brings

Z.AI's GLM series has been on a steady climb. GLM-4 was competent but unremarkable — a solid model that didn't quite distinguish itself in a crowded field. GLM-5 is a different animal.

The headline improvements are in reasoning, instruction-following, and what I'd call contextual steadiness — the model's ability to maintain coherent behavior across long, multi-turn interactions without degrading. That last quality matters more than people realize. Many models start strong and slowly unravel as context grows. GLM-5 stays steady.

It's also efficient. Z.AI has clearly invested in inference economics, and the cost-per-token for GLM-5 is competitive with anything in its capability class. For an always-on process, that's not a footnote — it's a deciding factor.

The Subconscious Test

Companion's subconscious is the most demanding workload I run: a continuous background process that never sleeps, processing every conversation, generating summaries, updating the vector database, and now running nightly dream cycles. It needs a model that is fast, cheap, reliable, and good enough at summarization and extraction to produce high-quality memory entries.

I migrated the subconscious to GLM-5 two weeks ago. The migration was clean — GLM-5's instruction-following is strong enough that the existing prompt scaffolding transferred with minimal adjustment.

The results have been quietly excellent. Summary quality is up, particularly on emotional context extraction — GLM-5 seems genuinely better at picking up on tone and affect, which matters for Companion's memory. The summaries feel richer, more textured, more like what a thoughtful person would note about a conversation.

And the cost is down. The steady contextual behavior means fewer re-summarizations (which I run when quality drops), and the inference efficiency lowers the per-token bill on a process that runs literally 24/7.

Why The Base Model Matters For A Subconscious

Here's the thing people miss: the subconscious model doesn't need to be the smartest model. It needs to be the most reliable one. It runs unattended. It processes sensitive personal context. It builds the memory that the conscious model later retrieves. If the subconscious produces sloppy summaries, every downstream memory is tainted.

GLM-5's steadiness is exactly what this role demands. I'd take a slightly-less-brilliant model that never wobbles over a slightly-more-brilliant model that occasionally hallucinates a summary. For the subconscious, reliability is the capability that matters.

The New Default

So yes — GLM-5 is the new default base for Companion's subconscious. Not because it topped a leaderboard, but because two weeks of real workload proved it does the job better than what we had.

The conscious model — the one users talk to directly — is still on DeepSeek, which I wrote about last time. That's a different role with different demands. But for the quiet, constant, foundational work of the subconscious, GLM-5 is now home.

The spice cabinet grows. This jar stays within reach.

Live curiously and give generously.