Selected model
ModelGPT 5.6 Luna (high reasoning)
Very cheap and surprisingly efficient. Work that arrives well specified rarely needs more than this.
Model routing, in practice
Pick the model. Let the harness carry it.
Everyone comparing coding agents compares capability. In practice I am solving three things at once: what the model is good at, where its subscription is allowed to run, and whether I have burned that quota this week.
Routing log
01 · Model routes
Start with the work.
Choose a task to see the model I reach for now. When an option was rejected, the record names the axis that eliminated it.
Selected model
ModelGPT 5.6 Sol or Opus 5
I use the strongest available model with Matt Pocock's grill-me skill because every downstream phase inherits the spec.
Selected model
ModelOpus 5 plans; Sonnet 5 implements
The Claude models are simply better at this work, and splitting the plan from the implementation keeps the plan honest.
Selected model
ModelGPT 5.6 Sol, sometimes Opus 5
GPT 5.6 Sol is the usual reasoning model. I reach for Opus 5 when the higher ceiling matters more than the usage window.
Selected model
ModelA model that did not plan the work
Same principle as the build pipeline: the thing that made the decision does not get to grade it.
Two of those three axes have nothing to do with model quality. Subscriptions carry five hour and weekly limits, so part of the routing exists to spread work across those windows inside a small shared AI budget, and keep running out on one provider from stopping the work.
02 · Substrate
What each harness is for.
Access matters, but it is not the thesis. ChatGPT/Codex and work Bedrock both run through opencode and pi; Cursor inference now runs inside opencode too.
| Harness | Current role | What it is for |
|---|---|---|
| opencode | Daily driver | At work and at home, with openchamber as its GUI. |
| pi | Daily driver | At work and at home, with subagents in mux panes. |
| claude | Personal projects | This is where my Claude subscription applies. I do not usually use it directly for work. |
| codex | Subagent dispatch | I use it mainly to dispatch subagents, not as a primary driver. |
| cursor | Personal projects | Generous included usage and fast. Its models now reach me inside opencode too. |
| antigravity | Retired | Disabled 2026-07-04. Its headless path was unreliable in Phase 1 testing. |
This is a role log, not a harness comparison. A fuller comparison is coming.