Knowledge Islands

← All guidance

Prompting OpenAI GPT-5.6 (Sol / Terra / Luna)

How to prompt OpenAI's GPT-5.6 family in Codex CLI. Read the shared Prompting guides principles first, then Prompting OpenAI GPT-5.5 — GPT-5.6 inherits the GPT-5-line prompt contract (outcome-first, effort-as-lever, verbosity param, tool preambles); this guide carries only what is distinctive to the 5.6 tiers. The canonical mapping of the portable model types (ki-tokenomics, ADR-KI-HARNESS-009) is shared, while ki-tokenomics-codex owns Codex-hosted runtime evidence.

Limited preview (2026-07). GPT-5.6 is in limited preview (a small set of approved orgs; broader access expected shortly). Tier names, effort rungs, and pricing here are preview-sourced and volatile — reconfirm against OpenAI's own docs before relying on them (this file's Sources, and ki-tokenomics' sources.md, track the reconfirmation).

The two axes: tier × effort

GPT-5.6 exposes model choice and reasoning effort as two separate, independently-set controls — the same two-axis shape as Claude Code's model + effort (agent frontmatter FM-2/FM-8) and Claude Cowork's model + effort pickers.

Tier (model line):

Tier Positioning Portable type OpenAI's recommended routing
Sol Flagship — deepest reasoning reasoning (and frontier in Ultra mode) Architecture, security, hard refactors, novel design
Terra Balanced default standard Standard feature work, everyday engineering
Luna Volume / cheap fast Docs, formatting, mechanical and sub-agent tasks

Effort (per request): Light → Medium → High → Extra High, plus two special modes — Max (more thinking time on a single task) and Ultra (subagent-parallel work; consumes usage limits fastest).

Mapping the portable types

There is no separate Codex model for long-horizon autonomy the way Claude has Fable — frontier and reasoning both resolve to Sol, separated by mode/effort rather than by a distinct model name. This is a real cross-vendor asymmetry, not an error: Claude draws the frontier/reasoning line with a model family (Fable vs Opus); Codex draws it within Sol via Ultra mode.

Portable type Codex resolution
frontier Sol @ Ultra (subagent-parallel)
reasoning Sol @ High / Max
standard Terra @ Medium
fast Luna @ Light / Low

A repo running under Codex declares these in ["knowledgeislands/ki-agentic-harness:ki-tokenomics".model_tier_bindings] (e.g. reasoning = "gpt-5.6-sol"); a config that lists both Claude and Codex models per type (reasoning = "opus, gpt-5.6-sol") is portable across both runtimes, each resolving to the first model it supports.

Prompting notes

  • Re-evaluate effort downward first. As with GPT-5.5, a sharper prompt plus clear stop rules often recovers what a higher effort would buy — reach for Extra High / Max / Ultra only when the task genuinely needs it, since the top modes consume usage limits fastest.
  • Match tier to the step, not the session. Run a bulk fan-out on Luna and escalate only the irreducible reasoning step to Sol — the same barbell discipline ki-tokenomics prescribes for Claude tiers.
  • Everything in Prompting OpenAI GPT-5.5 still applies — the Responses API, phase field, verbosity param, and tool-preamble guidance are unchanged across the GPT-5 line.