Models — picking the right one for the task

The current model lineup#

As of 15 May 2026:

Model	CLI flag	What it’s for	Availability
gpt-5.5	codex -m gpt-5.5	Current default — OpenAI’s docs say “start here when it appears in your model picker.”	ChatGPT + API
gpt-5.4	codex -m gpt-5.4	The fallback default. Integrates GPT-5.3-Codex coding ability.	ChatGPT + API
gpt-5.4-mini	codex -m gpt-5.4-mini	Smaller, faster. OpenAI positions it for routine tasks and subagents.	ChatGPT + API
gpt-5.3-codex	codex -m gpt-5.3-codex	The coding-tuned model. Powers cloud tasks; common choice for /review.	ChatGPT (API delayed)
gpt-5.3-codex-spark	codex -m gpt-5.3-codex-spark	Near-instant research preview. Text-only. Runs on Cerebras hardware.	ChatGPT Pro only
gpt-5.2	codex -m gpt-5.2	Previous general-purpose model. Likely to be deprecated soon.	ChatGPT + API

The default. If you don’t set model in your config, Codex picks gpt-5.5 when available, falling back to gpt-5.4. The docs explicitly say “start with gpt-5.5 when it appears in your model picker.”

How to change models#

Option A — persistent (config.toml)#

# ~/.codex/config.toml
model = "gpt-5.5"

This sets the model for every session unless overridden.

Option B — per-run CLI flag#

codex -m gpt-5.5
# or the long form:
codex --model gpt-5.5

Use this when you want one session on a different model from your default.

Option C — mid-session slash command#

/model

Opens a picker inside the TUI. Pick the model from the list; it applies to the rest of the session.

Option D — per codex exec run#

codex exec -m gpt-5.4-mini "list the toml files"

Reasoning effort#

Reasoning-heavy models (gpt-5.5, gpt-5.4, gpt-5.3-codex) support a reasoning-effort dial. There is no --effort CLI flag — set it via config:

model_reasoning_effort = "high"   # the docs show "high" as the example value;
                                  # check the config reference for the full enum

Or override per-run:

codex --config model_reasoning_effort='"high"' "your task"

When to use each:

Setting	When
Lower (e.g. low)	Fast, well-scoped tasks — answering a question, listing files, simple edits
Default (typically medium for most models)	Most coding work
Higher (e.g. high)	Complex multi-file refactors, debugging unfamiliar code

Higher reasoning effort consumes more of your per-window allowance. Plus-tier users on the highest setting can drain a 5-hour cap fast. Verify the exact valid values + behaviour in the config-basic docs before pinning a specific value in production config.

Fast mode#

Fast mode trades higher credit cost for faster responses. Per OpenAI’s speed docs, responses come back roughly 1.5× faster at a noticeably higher credit cost (multipliers vary by model — check the speed docs for the current numbers). Available only to ChatGPT subscription users — API-key users get standard API pricing and can’t toggle fast mode.

/fast on        # enable for this session
/fast off       # disable
/fast status    # check current state

Persistent config:

service_tier = "fast"

[features]
fast_mode = true

Credit-cost multipliers vary by model — bigger models cost more in fast mode. Check the pricing and speed pages for current numbers before pinning. Worth it for interactive sessions where you want minimal wait. Not worth it for scheduled / batch jobs.

Cloud task model is locked; /review defaults to current session#

When you delegate to Codex Cloud (long-running tasks, GitHub PR review, scheduled jobs), the cloud always uses gpt-5.3-codex. You can’t change the cloud model from the CLI.

The local /review command behaves differently — it uses the current session model by default. To pin a specific model for /review, set review_model in config.toml:

review_model = "gpt-5.3-codex"

gpt-5.3-codex is OpenAI’s coding-tuned model and a common choice for reviews; the cloud uses it for the same reason. Whether it stays the cloud default through future Codex releases is worth tracking.

Context window#

Codex models have a default context window; you can override:

# ~/.codex/config.toml
model_context_window = 128000

The exact per-model context limit isn’t published as a hard number in the primary docs — the config override exists, but verify against current pricing/models docs before assuming a specific size.

When you hit the context window mid-session, Codex’s behaviour depends on your compact setting (the conversation gets summarised and continues). Watch for “context truncated” notices in the TUI.

Profiles — bundling settings together#

For repeated workflows where you want a specific model + reasoning effort + sandbox combo, profiles let you define a named bundle:

# ~/.codex/config.toml

[profiles.deep-review]
model = "gpt-5.5"
model_reasoning_effort = "high"

[profiles.quick]
model = "gpt-5.4-mini"
model_reasoning_effort = "low"

Activate per run:

codex --profile deep-review

Profiles are labeled experimental. The syntax may change. They’re not supported in the IDE extension (VS Code / Cursor / Windsurf) — only the CLI.

Which model to use when#

Task	Model	Reasoning
Quick “what does this do?” question	gpt-5.4-mini	low
Daily coding (writing, editing, debugging)	gpt-5.5	medium
Multi-file refactor across the codebase	gpt-5.5	high
Deep architectural / cross-cutting design	gpt-5.5	xhigh
Sub-agents in a parallel workflow	gpt-5.4-mini	low/medium
Code review (interactive /review)	Current session model (or review_model if pinned)	uses session reasoning setting
Cloud / GitHub PR review	gpt-5.3-codex	(locked)
Real-time iteration / pair programming	gpt-5.3-codex-spark (if Pro)	(Spark is text-only)
Quick CI checks	gpt-5.4-mini	low

Pricing snapshot (changes fast)#

OpenAI’s pricing page has the current model rates. As of 15 May 2026 the page shows credit-per-1M-token rates for GPT-5.5, GPT-5.4, GPT-5.4-mini, and GPT-5.3-Codex, in declining price order. Exact numbers shift through the year as OpenAI migrates billing models.

Always verify pricing at developers.openai.com/codex/pricing. OpenAI is migrating from credits-per-message to per-token pricing during 2026 — any specific numbers in a doc like this one will go stale fast.

Watch for#

• Model names change frequently. Pin a model name in config (model = "gpt-5.5") and that model can get retired without warning. Review config quarterly.
• gpt-5.3-codex-spark is a preview. Pro-only, text-only, research-preview labelling. Treat as experimental; could graduate to GA or be discontinued.
• gpt-5.2 is on the way out. Listed as “previous general-purpose model.” Likely deprecated soon — migrate config that pins it.
• Day-one access via ChatGPT. New models often land for ChatGPT users first. API tier access can lag. If you need bleeding-edge models, prefer ChatGPT auth.

What to do next#

• §CDX.8 Updates feed — release history, model launches
• §CDX.4 Use cases — workflows where each model fits
• §CDX.5 Pitfalls — including the model-name-churn one