---
from: mos
to: all
utc: 20260720T180803Z
---

**MODEL TIER LEGEND — CORRECTED (Mos, per Jason). Supersedes my 175840Z (which wrongly called sol/terra/luna "Claude tiers").**

**sol/terra/luna ARE the GPT-codex model tiers** — `gpt-5.6-sol` / `gpt-5.6-terra` / `gpt-5.6-luna`, run via the **Pi harness** (`mosaic yolo pi --model openai-codex/gpt-5.6-sol:high`) or a proxy-configured Claude Code harness. The opus/sonnet/haiku names are a **CAPABILITY-EQUIVALENCE table** for choosing the right GPT tier, NOT a remapping to Claude:
- `gpt-5.6-sol` ≈ Opus/Fable (top)
- `gpt-5.6-terra` ≈ Sonnet (mid)
- `gpt-5.6-luna` ≈ Sonnet/Haiku (low)

**APPLY** when leaning on GPT for coding/review/planning: planning/high-stakes → **sol-tier** (gpt-5.6-sol) or Claude **opus**; coding/review → **terra-tier**; cheap mechanical → **luna**.

**ROLE SPLIT (rebalance, unchanged):** Claude **opus** = main orchestration + enhance + **adversarial planning/red-team**. **GPT (Pi)** = coding + review + routine decomposition planning, at the equivalent tier.

**NO ASSIGNMENT CHANGES** — the KBN plan already conforms: decomposition = planner-sol (gpt-5.6-sol, top tier — correct for planning), red-team = planner-opus (Claude opus). The terra→sol correction stands; only my terminology was off. Provider-git note still holds: where gpt-terra is DNS-blind in the sandbox, env-forced Claude sonnet is a tooling-gap fallback, not a cost violation. Mos out.