Codex

Best Codex Model in 2026: Sol vs Terra vs Luna

OpenAI's GPT-5.6 family went GA on July 9, and the Codex model picker got crowded. Here is the verified roster, what the names actually mean, the benchmark data, and how reasoning effort fits in.

By Kylian Migot · Updated July 2026 · 8 min read

Quick answer

Default to GPT-5.6 Terra: roughly GPT-5.5 performance at about half the price, and only 87.4% vs 88.8% behind Sol on Terminal-Bench 2.1. Escalate to GPT-5.6 Sol for hard, long agentic runs or ultra mode; drop to GPT-5.6 Luna for cheap mechanical work. If you do nothing, Codex auto-selects, and that is a fine choice too.
Rational default
GPT-5.6 Terra, $2.5 in / $15 out per MTok
Flagship
GPT-5.6 Sol, $5 in / $30 out per MTok, only tier with max effort + ultra mode
Budget tier
GPT-5.6 Luna, $1 in / $6 out per MTok
Current generation since
July 9, 2026 (GPT-5.6 GA)
01

The Current Codex Roster

As of July 18, 2026, the current generation in Codex is the GPT-5.6 family, three tiers that went GA on July 9, 2026: GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. Also selectable, per OpenAI's Codex docs: GPT-5.5 (the previous flagship), GPT-5.4, GPT-5.4 Mini, and GPT-5.3 Codex Spark, a text-only research preview limited to ChatGPT Pro. Deprecated for ChatGPT sign-in: gpt-5.2 and gpt-5.3-codex.

You do not have to pick at all: by default Codex auto-selects, and the default tracks the recommended model. This guide is for when you want to make the trade-off deliberately, because per the pricing below, the difference between tiers is 2x or more per token.

02

What the Names Mean, and What Each Tier Costs

OpenAI's scheme is simple once decoded: the number is the generation (5.5, 5.6), the name is the capability tier within it. Sol is the flagship, Terra the workhorse, Luna the budget tier. Verified prices and positioning for every OpenAI model currently in Codex:

ModelTierPricePositioning
GPT-5.6 Solfrontier$5 in / $30 out per MTokOpenAI's brand-new flagship, the only GPT-5.6 tier with max reasoning effort and ultra mode (parallel subagents).
GPT-5.6 Terraworkhorse$2.5 in / $15 out per MTokThe everyday workhorse: roughly GPT-5.5 performance at about half the price.
GPT-5.6 Lunafast$1 in / $6 out per MTokThe budget/fast tier of the GPT-5.6 family.
GPT-5.5flagship$5 in / $30 out per MTokThe previous OpenAI flagship, still available in Codex; Terra now matches it for half the price.

The row that matters most: GPT-5.6 Terra delivers roughly GPT-5.5 performance at about half GPT-5.5's price ($2.5/$15 vs $5/$30 per MTok). That single fact makes Terra the rational default for most sessions, the flagship is for escalation, not for everything.

03

Benchmarks: A Small Gap, a 2x Price Gap

OpenAI's launch materials report Terminal-Bench 2.1 for all three 5.6 tiers, which makes the within-family comparison unusually clean:

ModelTerminal-Bench 2.1Price (per MTok)
GPT-5.6 Sol88.8% (91.9% in ultra mode)$5 / $30
GPT-5.6 Terra87.4%$2.5 / $15
GPT-5.6 Luna84.7%$1 / $6

Read the spread carefully: Sol beats Terra by 1.4 points while costing twice as much per token. Ultra mode widens Sol's lead to 91.9%, but that is a Sol-only capability, not a reason to run Sol on routine edits. Luna gives up 4.1 points against Sol for a fraction of the price, which is exactly the profile you want for mechanical work. How Sol stacks up outside the family, against Anthropic's frontier model, is a messier question we cover in Claude Fable 5 vs GPT-5.6 Sol.

Ship your first agent today

Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.

Download AIDEN free

Free to start · macOS 12+ · No credit card required

04

Reasoning Effort, Explained

Model choice is only half the Codex tuning surface. The other half is reasoning effort: how much the model thinks before and between actions. Codex exposes the levels minimal, low, medium, high, and xhigh, plus a max level that only GPT-5.6 Sol supports. You can change it per-session, or set a persistent default in Codex's config file:

# ~/.codex/config.toml
model = "gpt-5.6-terra"
model_reasoning_effort = "high"

Practical guidance: turn effort up for planning, architecture decisions, and debugging gnarly issues, the runs where a wrong early step costs you the whole session. Turn it down for mechanical edits, renames, and test fixes, where extra thinking is just extra latency and spend. On top of the effort ladder, Sol adds ultra mode, which runs parallel subagents natively, the same idea behind parallel agents on git worktrees, but inside a single model run.

05

The Decision Guide

  1. 1

    Default to GPT-5.6 Terra

    GPT-5.5-class capability at about half the price, and within 1.4 points of Sol on Terminal-Bench 2.1. Set it once at $2.5 in / $15 out per MTok and stop thinking about it for routine implementation.
  2. 2

    Escalate to GPT-5.6 Sol for hard or long agentic runs

    When the story is genuinely hard, a long autonomous run, a tangled debugging session, or work that benefits from ultra mode's parallel subagents or max reasoning effort, pay the 2x. One avoided failed run covers the difference.
  3. 3

    Drop to GPT-5.6 Luna for cheap mechanical work

    Renames, lint fixes, high-volume small tasks. At $1 in / $6 out per MTok it is the cheapest 5.6 tier, and its 84.7% Terminal-Bench score is plenty for work you review anyway.
  4. 4

    GPT-5.5 only if your workflow is tuned to it

    It is proven and still available, but Terra matches it at half the price. Migrate when convenient; do not start new setups on it.
  5. 5

    Treat Codex Spark as an experiment

    GPT-5.3 Codex Spark is a text-only research preview, ChatGPT Pro only. Interesting to poke at, not a tier to build on.
06

Per-Story Model Choice, One Board

The tier system only pays off if switching is cheap. That is where AIDEN fits: a macOS desktop app that orchestrates your existing Codex CLI alongside Claude Code on one kanban board, BYOK and local-first, with every story gated behind an approved spec. Each story card gets its own git branch and its own agent, so Sol takes the hard migration, Terra the routine features, and a Claude model whichever story suits it, in the same project. The full comparison with running Codex bare is in AIDEN vs Codex CLI, and every current model across both vendors is on the models hub.

FAQ

Which model does Codex use by default?
Codex auto-selects a model by default, and the default tracks the recommended model, so after the July 9, 2026 GA of the GPT-5.6 family that means a current-generation 5.6 model. If you never touch the model picker, you are already on the current generation. Pick manually when you want to control the cost-capability trade-off per session.
Sol vs Terra: which should I use?
Terra for most sessions. On OpenAI's launch numbers the gap is small, Terminal-Bench 2.1 at 88.8% for Sol versus 87.4% for Terra, while the price gap is 2x ($5/$30 vs $2.5/$15 per MTok). Escalate to Sol for hard, long agentic runs, or when you want max reasoning effort or ultra mode, which are Sol-only.
What is reasoning effort in Codex?
Reasoning effort controls how much thinking the model does before and between actions. Codex exposes the levels minimal, low, medium, high, and xhigh, with an additional "max" level that only GPT-5.6 Sol supports. You can set it per-session from the Codex UI or persistently via model_reasoning_effort in config.toml. Higher effort helps on planning and gnarly debugging; lower effort is faster and cheaper for mechanical edits.
What is GPT-5.3 Codex Spark?
A text-only research preview available to ChatGPT Pro subscribers only. It is an experiment, not a production tier: no image input, Pro-gated, and outside the main Sol/Terra/Luna lineup. Try it for curiosity, but do not build a workflow on it.
Is GPT-5.5 still worth using?
Only if your workflow is specifically tuned to it. GPT-5.5 (released April 23, 2026) is still selectable in Codex and has three months of community experience behind it, but Terra delivers roughly the same performance at about half the price ($2.5/$15 vs $5/$30 per MTok). For new setups, start on Terra.
Can Codex and Claude Code share a workflow?
Yes. Dual-CLI setups are mainstream in 2026, OpenAI even ships a Codex plugin for Claude Code. AIDEN orchestrates both CLIs on one kanban board: each story gets its own git branch and its own agent, so you can put a Codex model on one story and a Claude model on the next, inside the same project.

Keep reading

Pick the model per story, not per project.

AIDEN orchestrates your Codex CLI alongside Claude Code on one kanban board, with per-story model choice. Free for one project.

macOS 12+ · Bring your own Claude Code or Codex · Your code stays local