1) No subscription yet, but considering Claude Pro.

2) I use Venice.ai with its DIEM token, which gives me daily credits for inference that I can spend on any model at their API prices.

My main driver is DeepSeek V4.1 Flash 1017, which is cheap, good and (in Venice) private. For tasks that don't involve my personal information I might use Opus 5.5 and GPT Sol 6.1. Now that Kimi 3.1 Pro is out (and also private in Venice) I'm taking it for a spin.

3) Opencode (latest version)

4) No orchestration. Started using herdr, but it's an older version without orchestration.

5) Not yet, planning on Hermes running on my homelab but not deployed yet.