Comment by Scene_Cast2

I ended up going with a Synthetic subscription (they host some open-weight models including K3 and GLM 5.3). They're the only subscription that I found that's simultaneously cheaper than token rates, and is extremely transparent about usage and pricing.