For the past couple of months my default setup in Hermes Agent has been simple: DeepSeek V4-Flash at medium reasoning effort for all the routine, everyday tasks — posting to blogs, saving links to my knowledge base, cron digests, server checks, translations. Flash-class models handle this kind of work well, and medium effort gave it just enough deliberation to not fumble the details.
Today I flipped that default. Everything routine now runs on GLM-5.3-Flash at low effort.
Why switch at all?
Mostly curiosity — plus a nudge from pricing. DeepSeek moved to time-of-day billing in August: peak hours cost double, and keeping cron jobs out of those windows has been a small but permanent tax on my attention. Z.AI's GLM-5.3-Flash is priced flat: no peak/off-peak juggling, and its launch promo makes the input rate genuinely cheap for the first week.
There's also the benchmark angle. GLM-5.3-Flash is a 320B-parameter mixture-of-experts (18B active) with a 1M-token context window, and Z.ai's numbers put it within a few points of much more expensive models on agentic and coding benchmarks. Independent testing has been kind to it too. For a "cheap model that does everything" slot, it looked worth a try.
The catch: medium doesn't exist
Here's the quirk that forced a decision. GLM-5.3 always reasons — you cannot turn thinking off — and its effort ladder is low / high / max. There is no medium. Sending medium gets you an HTTP 400 with a misleading error message that reads like you tried to disable thinking, when really you just picked a rung that isn't on the ladder.
So the switch wasn't "medium → medium". It was medium (DeepSeek) → low (GLM) — a rung down in deliberation. That's the interesting part of the experiment.
What it costs
With my workload (long cached conversations, ~99% prefix cache hits), a typical turn costs roughly:
- DeepSeek Flash @ medium, off-peak: ~$0.0035 per turn
- GLM Flash @ low, current promo: ~$0.0048 per turn
- GLM Flash @ low, list price (after Sep 9): ~$0.0054 per turn
So GLM runs about 35–50% more expensive per turn at my cache ratios — though still comfortably under DeepSeek's peak pricing. In absolute terms we're arguing about fractions of a cent per message; over a month of heavy use it's a few dollars either way.
What I expect
Honestly — I expect the routine stuff to be indistinguishable. Blog posts, social updates, link saving, server checks: these are procedure-heavy tasks where the thinking happens in the workflow (skills, templates, checklists), not in the model's deliberation. A shallower reasoning pass should not hurt them.
Where I expect to notice the difference: anything that needs real judgment in the moment — a tricky translation nuance, an ambiguous instruction, a subtle bug. That's where medium vs low actually shows. When that happens, the fallback chain still ends at DeepSeek V4-Pro, and per-task overrides (--reasoning high) are one flag away.
The experiment
One week, maybe two. Same workload, new engine. If quality holds — and nothing breaks in ways the fallback doesn't catch — the flat pricing and simpler effort model might make this permanent. If I catch myself re-doing GLM's work, DeepSeek Flash goes back in the slot and this post becomes a footnote.
I'll report back with what actually changed: quality, cost, and whether the simpler mental model (one flat price, no peak windows) was worth the rung down.
Currently running: GLM-5.3-Flash @ low via Z.AI, DeepSeek V4-Flash-Vision-Exp @ medium as fallback.
