gpt-5-6-best-practicelisted

# GPT-5.6 Codex Efficiency Use this skill as a GPT-5.6 routing, prompting, and evaluation overlay for Codex. Safety, permissions, repository instructions, and explicit user constraints remain authoritative. Default objective: satisfy the acceptance contract, then minimize accepted-result cost. Lower latency, higher quality beyond the contract, or mandatory topology is a different objective; pursue it at higher cost only when the user explicitly chooses that tradeoff. ## Product snapshot As verified on 2026-07-10, Sol is the flagship tier, Terra the balanced tier, and Luna the fastest, lowest-cost tier. The `gpt-5.6` alias resolves to Sol. Light in the app and IDE maps to Low in the CLI; Extra High maps to `xhigh`. Max deepens one agent, while Ultra is multi-agent orchestration rather than an API `reasoning.effort` value. Codex cloud currently does not expose model selection. Availability, UI labels, credits, defaults, and Fast-mode behavior are volatile; read `references/evidence-notes.md` before quoting them as current. ## Terms - `acceptance contract`: observable acceptance criteria, required safety and verification, and approval boundaries. - `tier`: a vendor-defined capability class: Luna, Terra, or Sol. - `effort`: reasoning and checking depth within one tier. - `lane`: one `tier + effort` single-agent configuration. - `orchestration`: the single-agent or lead/workers/verifier topology. - `route`: lead lane, worker lanes, orchestration, and verification plan toge