Codex models: every OpenAI model Codex runs, and the ones it dropped

The three models Codex runs in August 2026

OpenAI rolled the GPT-5.6 family out to ChatGPT, Codex and the API on 9 July 2026. It replaced a naming scheme that had produced a new Codex-suffixed model roughly every two months, and the new one has no Codex variant at all: the same three general models serve the chat window, the CLI, the IDE extension and Codex cloud.

All three take the same reasoning ladder: low, medium, high, extra high, max and ultra. Effort moves cost and latency far more than most people expect, so it is worth treating as a second dial rather than something you set once and forget. The CLI reads it from model_reasoning_effort, and you can override it for a single run.

The Continuum coding-agent card for the flagship is GPT-5.6 Sol.

gpt-5.6-sol: the flagship, and the one to use least

Sol is the strongest model in the family and the default in most Codex installs. It is also the most expensive by a wide margin, and the reason people burn a weekly cap in three days.

The gap between Sol and Luna is 25x on input and 25x on output. That is not a rounding difference between two similar models, it is the difference between a habit that costs $40 a month and one that costs $1,000. On a subscription the same ratio shows up in the credit rate card: 125 credits per million input tokens on Sol against 5 on Luna.

gpt-5.6-terra and gpt-5.6-luna: the two you should use more

Terra

OpenAI positions Terra as competitive with GPT-5.5, the previous generation flagship, at a fraction of Sol's price. That framing is the point: a model that matched the best available option four months ago is not a compromise, it is the sane default for feature work. It is also the migration target for anyone still on gpt-5.4.

Luna

Luna is the cheapest model in the family and the one most people never try. It is genuinely good at work where the thinking has already been done: renames across a repository, applying a spec you wrote, generating a commit message, converting a fixture, writing the boring half of a test suite. Handing that class of task to Sol is paying a research model to do typing.

# mechanical, complete spec, no judgement required
codex -m gpt-5.6-luna "rename Config.Timeout to Config.RequestTimeout across the repo"

# ordinary feature work
codex -m gpt-5.6-terra "add pagination to the /orders endpoint, follow the pattern in /invoices"

# the thing that already failed twice
codex -m gpt-5.6-sol -c model_reasoning_effort="high" "why does the worker deadlock under load?"

gpt-5.3-codex-spark: the last model with codex in its name

Spark arrived in research preview on 12 February 2026, a week after GPT-5.3-Codex itself, as a smaller text-only sibling built for near-instant iteration rather than depth. It is still listed in Codex today, and it is the only surviving member of the Codex-suffixed generation.

Which ChatGPT plan and which API tier serves each model

Codex has two independent rails and they do not carry the same models. The subscription rail is what the CLI uses after codex login; the API rail is what an API key uses, and what CI needs.

Every surface draws on the same plan: the CLI, the openai.chatgpt IDE extension, the ChatGPT desktop app in Codex mode, Codex cloud and the iOS app all count against one set of limits, not five. The API is the exception, and the only rail with no rolling window and no weekly cap.

How to pin a model in the CLI

Four scopes, from narrowest to widest. Verified against codex-cli 0.144.1.

# -m and --model are the same flag
codex -m gpt-5.6-terra
codex --model gpt-5.6-luna

# non-interactive, same flag
codex exec -m gpt-5.6-luna "write a conventional commit message for the staged diff"

# model plus effort, for one run only
codex -m gpt-5.6-sol -c model_reasoning_effort="xhigh"

model                  = "gpt-5.6-terra"
model_reasoning_effort = "medium"

# -p layers $CODEX_HOME/<name>.config.toml over your base config
# so ~/.codex/cheap.config.toml holds the mechanical-work settings
codex -p cheap "apply the rename in RENAME.md"

Inside a session, /model switches without losing the conversation, which is the right move the moment a task turns out to be more mechanical or more stubborn than it looked. The TUI shows the active model, so there is never a reason to guess which one is answering.

Codex 5.1: gpt-5.1-codex, gpt-5.1-codex-max and gpt-5.1-codex-mini

The 5.1 generation is where the Codex-suffixed naming peaked, and it is why so many search results still name models you cannot select. Three ids shipped in November 2025 and all three are gone.

Codex-Max is the one worth remembering, because its headline capability did not disappear with it. Compaction, the technique that let a single task run coherently across millions of tokens and several hours, is now ordinary behaviour rather than a model you select. If you came here looking for gpt-5.1-codex-max because you needed multi-hour sessions, you already have them.

Codex 5.3: gpt-5.3-codex, and where the 5.3 codex line went

GPT-5.3-Codex landed on 5 February 2026 and was, briefly, the most talked-about coding model of the year. It is also the clearest example of the thing this page exists to prevent: it was the recommended replacement in April, and it was gone from ChatGPT sign-in by summer.

1. 5 February 2026. gpt-5.3-codex ships, multimodal, initially Pro-gated.
2. 12 February 2026. gpt-5.3-codex-spark follows in research preview, text only.
3. 14 April 2026. The 5.1 and 5.2 Codex ids retire, and gpt-5.3-codex is named as one of their replacements.
4. Mid 2026. gpt-5.3-codex itself stops being offered to ChatGPT-signed-in Codex accounts. Spark survives on Pro.

The practical reading of the 5.3 story: search volume for a model name lags its availability by months. If you arrived here searching "codex 5.3" or "5.3 codex", the honest answer is that the thing you read about is either Pro-only preview software or gone, and the model you actually want is gpt-5.6-terra.

Every Codex model, and what replaced it

Questions people ask

Keep reading

Continue through OpenAI Codex CLI with OpenAI Codex CLI: the complete guide.