GPT-5.6 in Codex: Sol, Terra, and Luna

Codex already defaults to gpt-5.6-sol, the flagship, running at medium reasoning (models reference). If that’s what you want, there’s nothing to change. To drop to a cheaper tier for routine work, pass the slug you want when you launch:

codex -m gpt-5.6-sol      # flagship, the default
codex -m gpt-5.6-terra    # balanced everyday work
codex -m gpt-5.6-luna     # fast and cheap

GPT-5.6 went GA on July 9, 2026 across ChatGPT, Codex, and the API (changelog), one generation past gpt-5.5.

The three tiers

Same generation, sized differently. The models reference calls Sol the strongest for complex coding, computer use, and research; Terra the balanced model “competitive with GPT-5.5 at a lower cost”; and Luna the fast, low-cost option for clear, repeatable jobs like extraction and classification.

API pricing per million tokens, from the GPT-5.6 announcement:

Model	Input	Output
gpt-5.6-sol	$5.00	$30.00
gpt-5.6-terra	$2.50	$15.00
gpt-5.6-luna	$1.00	$6.00

Which one to pick

1. Sol for the hard stuff: ambiguous refactors, deep research, anything where a wrong answer costs you real time. Dial reasoning effort down when the task is easy so you burn fewer tokens on the parts that don’t need it.
2. Terra as your daily driver. Most feature work and bug-fixing doesn’t need the flagship, and Terra lands close to it for half the output price.
3. Luna for high-volume, well-defined runs: batch edits, classification, scripted codex exec jobs where you already know exactly what you’re asking for.

When you’re between two tiers, go lower first and bump up if the output isn’t good enough. That’s cheaper than the reverse.

Switching without restarting

1. Type /model inside a running session to open the picker and switch tiers on the fly.
2. To make it stick, set it in ~/.codex/config.toml:

   model = "gpt-5.6-terra"

3. For per-workflow pins, drop a profile file at $CODEX_HOME/<name>.config.toml and select it with codex --profile <name> (the file does nothing until you pass the flag). A ci.config.toml pinned to Luna, run as codex --profile ci, keeps scripted runs off the flagship meter (config sample).

Gotchas

• Old slugs are deprecated on ChatGPT sign-in. The models reference marks gpt-5.2 and gpt-5.3-codex deprecated in Codex when you’re signed in with a ChatGPT account. gpt-5.3-codex still answers over the API, but pin a 5.6 slug for anything current.
• Typo the slug and you get an error, not a fallback. A model id your provider doesn’t serve throws instead of quietly downgrading. See model not supported.
• Quota draining faster than you expect? Model choice plus reasoning effort is the whole cost equation. Sol at high effort moves the meter fast. See usage limit draining too fast.

Related

• /models/: the living tracker with every current slug, its status, and what’s deprecated, kept current
• model: the config key that sets your default
• model_reasoning_effort: the other half of the cost equation
• Usage limits: how the plan quotas that model choice eats into actually work