How to Run OpenAI Codex (GPT-5.6 Sol) Inside Claude Code: Full Guide

If you pay for both Claude and a ChatGPT/Codex plan, you have probably wondered whether the two can actually work together instead of sitting in separate terminal windows. As of July 2026, they can. OpenAI’s GPT-5.6 “Sol” model — and its cheaper siblings Terra and Luna — can now run inside Claude Code, either as a reviewer that double-checks Claude’s work or as the actual model driving your session.

There are two ways to set this up, and both take only a few minutes. This guide walks through each one from scratch, explains which one you actually want, and covers the model tiers, costs, and the fixes for the problems people hit most often.

The trick took off after this post from Alvaro Cintas:

OpenAI’s GPT-5.6 Sol model can now run inside Claude Code. There are two ways to do it, and both take minutes to set up, here’s how: Option 1, the official plugin…

— Alvaro Cintas (@dr_cintas) View post on X

What does “Codex inside Claude Code” actually mean?

These are two competing tools from two competing labs, so it helps to be precise about what is happening. There are two fundamentally different setups, and people mix them up constantly.

Setup A — Codex as a helper. Claude stays in charge of your session. You install OpenAI’s official plugin, and Claude can hand specific jobs to Codex on demand: “review this diff,” “go investigate this bug,” “rescue this half-finished feature.” Claude orchestrates and keeps the context; Codex is a second opinion and a second pair of hands. Nothing about your normal Claude workflow changes.

Setup B — GPT-5.6 as the driver. You point Claude Code’s engine at a GPT-5.6 model, so the Claude Code interface is running on OpenAI’s model under the hood. You keep the Claude Code UX (its file tools, its terminal, its subagents) but the thinking is done by GPT-5.6 Sol. This is the “one CLI, either brain” approach.

Setup A is the officially supported, no-drama route and is what most people should use. Setup B is more of a power-user trick — flexible and genuinely useful, but it leans on community proxy tools and is a bit more of a gray area. We will cover both in full so you can decide.

Meet the models: GPT-5.6 Sol, Terra, and Luna

GPT-5.6 became fully available on July 9, 2026, and ships as three tiers inside the Codex CLI. Picking the right one is the single biggest lever you have over both quality and cost:

• gpt-5.6-sol — the premium tier. This is what you reach for on hard problems: cross-repository refactors, gnarly debugging, architecture-level reasoning. It is the most capable and the most expensive.
• gpt-5.6-terra — the mid-tier default. Quality that edges past the older GPT-5.5 at roughly half of Sol’s price. This is the sensible everyday choice for routine development.
• gpt-5.6-luna — the economy tier. Built for completions, batch jobs, and light CI tasks where you want speed and low cost more than deep reasoning.

As a rough guide to the pricing spread at launch, Luna sat around $1/$6 per million input/output tokens, Terra around $2.50/$15, and Sol around $5/$30. Exact numbers change, so treat these as ratios rather than gospel — the point is that Sol costs roughly five times what Luna does, so you do not want to run Sol on trivial tasks.

Before you start: prerequisites

Whichever route you take, you will need the following in place first:

• Node.js 18.18 or later. Check with node --version. The plugin and the Codex CLI both need it.
• Claude Code installed and working, with slash-command support (any current version).
• A ChatGPT/Codex account. Usage runs through your OpenAI account — a ChatGPT subscription (a free tier works for light use) or an OpenAI API key. Access to the newest Sol tier depends on your plan.
• The Codex CLI. If you do not have it yet, install it globally: npm install -g @openai/codex — the plugin’s setup step can also do this for you.

Method 1: The official Codex plugin (recommended)

This is OpenAI’s own plugin for Claude Code (the openai/codex-plugin-cc project). It is free, open source, and the cleanest way to get Codex working as Claude’s reviewer and delegate. The whole thing is four commands.

Step 1 — Add the plugin marketplace

Open Claude Code and run:

/plugin marketplace add openai/codex-plugin-cc

Step 2 — Install the plugin

/plugin install codex@openai-codex

Step 3 — Reload plugins

/reload-plugins

Step 4 — Run setup and log in

Now verify everything is wired up correctly:

/codex:setup

The setup command checks that Node, npm, and the Codex CLI are present and that you are authenticated. If the Codex CLI is missing, it can install it for you. If you are not logged in, authenticate with:

!codex login

That opens a browser login flow tied to your ChatGPT/OpenAI account. Once /codex:setup reports everything green, you are ready to use Codex from inside Claude.

The commands you actually get

Installing the plugin adds a set of /codex: slash commands. These are the tools Claude uses to hand work to Codex:

• /codex:review — a standard, read-only code review of your uncommitted changes or a branch comparison.
• /codex:adversarial-review — a tougher review that deliberately challenges your design decisions, tradeoffs, and assumptions instead of just rubber-stamping the diff.
• /codex:rescue — delegate an actual task to Codex: investigate a bug, apply a fix, or continue work Claude started.
• /codex:transfer — spin up a persistent Codex thread from your current Claude Code session so context carries over.
• /codex:status — show running and recent Codex jobs for the current repo.
• /codex:result — print the final output (with session IDs) from completed jobs.
• /codex:cancel — kill an active background Codex job.
• /codex:setup — the verification/config command from Step 4; also manages the optional review gate.

Useful flags and examples

The review and rescue commands accept options that make them far more practical on real projects:

• --base <ref> — compare against a specific branch (e.g. --base main).
• --background — run the job asynchronously so it does not block your Claude session.
• --wait — wait for the job to finish before returning.
• --model <name> — choose which Codex model runs the job (e.g. gpt-5.6-sol).
• --effort <level> — set the reasoning effort (e.g. medium, high).

A couple of real examples:

# Have Codex tear into your caching design, comparing against main
/codex:adversarial-review --base main challenge caching design

# Delegate a flaky-test investigation to a specific model and effort level
/codex:rescue --model gpt-5.6-sol --effort high investigate the flaky test

Setting a default model

If you leave --model off, Codex picks a default. To set your own default across a project, drop a .codex/config.toml file in the repo:

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

Config loads from ~/.codex/config.toml (your user-level default) and from .codex/config.toml in the project (if the project is trusted). Because the plugin simply reuses your existing Codex CLI setup, any API keys, custom base URLs, and preferences you already have carry over automatically — there is no separate runtime to babysit.

The review gate (handle with care)

Setup lets you enable a review gate — an automatic Codex review that runs before Claude Code marks work as done. It is a genuinely powerful feedback loop: every change gets a second-model sanity check without you asking. The catch is that it fires on every completion, so it can burn through your Codex usage limits fast. Turn it on deliberately, watch your usage, and disable it with /codex:setup --disable-review-gate if the bill climbs faster than you like.

Method 2: Point Claude Code directly at GPT-5.6 Sol

This is the “make Claude Code run on GPT-5.6” route. Instead of Claude calling Codex as a helper, you replace the model doing the work with a GPT-5.6 model while keeping the Claude Code interface. There are two flavors, from simple to full-control. This approach was popularized by OpenAI’s Thomas “Tibo” Sottiaux:

If you aren’t yet bold enough to install the Codex app, you can stay in the presence of your orange crab and point it at GPT 5.6 Sol. Takes 5 minutes.

— Tibo (@thsottiaux) View post on X

Flavor A — the subagent alias (quick)

The lightest version keeps your main claude command untouched and only routes Claude Code’s subagents to GPT-5.6 Sol. Add a shell alias (to your ~/.bashrc or ~/.zshrc):

alias claudex='CLAUDE_CODE_SUBAGENT_MODEL=gpt-5.6-sol \
CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1 \
CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY=3 \
ENABLE_TOOL_SEARCH=false \
claude --model gpt-5.6-sol'

What each variable does:

• CLAUDE_CODE_SUBAGENT_MODEL=gpt-5.6-sol — routes the subagents (the parallel workers Claude spawns) to GPT-5.6 Sol.
• CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1 — forces maximum reasoning effort.
• CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY=3 — caps how many tool calls run at once, which keeps things stable when a proxy is in the loop.
• ENABLE_TOOL_SEARCH=false — turns off the auxiliary tool-search feature.

Reload your shell (source ~/.zshrc) and launch with claudex instead of claude. This is the setup people describe as letting the Claude Code orchestrator fan work out to a swarm of GPT-5.6 subagents — which is exactly why folks report large savings on their Claude token budget: the heavy lifting shifts onto their Codex plan.

Flavor B — the full proxy (complete control)

The complete version runs a small local proxy that translates between Claude Code and the OpenAI backend, then tells Claude to talk to that proxy. This gives you the most control and lets you set both the main and the “small/fast” model.

1. Install a proxy. A popular one installs via Homebrew:

brew install raine/claude-code-proxy/claude-code-proxy

2. Authenticate with your ChatGPT account (OAuth in the browser):

claude-code-proxy codex auth login
claude-code-proxy codex auth status

3. Start the proxy (it listens locally, e.g. on port 18765):

claude-code-proxy serve
# Listening on http://127.0.0.1:18765

4. Point Claude Code at the proxy with environment variables:

export ANTHROPIC_BASE_URL=http://localhost:18765
export ANTHROPIC_AUTH_TOKEN=unused
export ANTHROPIC_MODEL=gpt-5.6-sol
export ANTHROPIC_SMALL_FAST_MODEL=gpt-5.6-luna
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=372000
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
export CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1

Notice you can set a premium model (gpt-5.6-sol) for the main work and a cheap one (gpt-5.6-luna) for the small, fast background calls — a nice way to keep costs down without hurting output quality.

5. Launch Claude against the GPT backend:

claude --model gpt-5.6-sol

Wrap steps 4 and 5 in a shell function so you get a clean claudex command for GPT sessions while your normal claude stays on Anthropic’s models. Just remember the proxy has to be running for it to work, which is the main maintenance overhead of this flavor.

Which method should you pick?

Short version: start with Method 1.

Choose Method 1 (official plugin) if you mainly want a second AI to review Claude’s work, catch blind spots, and take over specific debugging jobs — all officially supported, with nothing to maintain. This is the safe default and covers the most common reason people want both models: cross-checking each other.

Choose Method 2 (direct/proxy) if you specifically want the Claude Code interface but with GPT-5.6 as the engine, or if you are trying to shift your workload onto your Codex plan to conserve Claude usage. It is more powerful and more flexible, but it relies on community proxy tooling and sits in more of a gray area regarding each provider’s terms — so go in with eyes open, and keep an eye on your usage.

Real workflows this unlocks

Once it is set up, the interesting part is the two-model collaboration. A few patterns that work well in practice:

• Build with one, review with the other. Let Claude write the feature, then run /codex:review or /codex:adversarial-review before you commit. You get a genuine second opinion from a different model family, which catches a different class of mistakes than a same-model review would.
• Delegate the deep dive. When Claude is spinning on a tricky bug, hand it off with /codex:rescue --background and keep working while Codex investigates in parallel. Check in later with /codex:status and /codex:result.
• Ship-critical adversarial pass. Before shipping anything important, run an adversarial review that is prompted to attack your assumptions rather than approve them.
• Cost-shaped model routing. Use Terra or Luna for routine reviews and reserve Sol for the hard stuff, either via --model per command or a project .codex/config.toml.

As a rough sense of timing, small reviews (1–3 files) come back in about 10–30 seconds, medium changes (4–10 files) in 30–90 seconds, and large ones (10+ files) in roughly 2–5 minutes. Use --background on the bigger jobs so they never block you.

What it costs and how limits work

The plugin and the proxy tools are free and open source, but the model calls are not. Everything you run through Codex — whether via the plugin or the proxy — bills against your ChatGPT subscription or OpenAI API account at standard Codex rates. In practice that means two things. First, your model choice is your cost dial: Sol is roughly five times the price of Luna, so route work accordingly. Second, features that call Codex automatically — chiefly the review gate — can eat through your usage allowance quickly, so enable them deliberately and monitor them.

The flip side is the reason many people set this up at all: by pushing review and delegation work onto your Codex plan, you take load off your Claude usage. If you already pay for both subscriptions, you are simply getting more out of the money you are already spending.

What the community is saying (and two caveats worth heeding)

Reception to the official plugin has been quietly positive. Developers who dug into the code called it “very well designed,” and more than one person remarked they were surprised it wasn’t getting more attention given how useful it is. The feature people keep coming back to is the one this whole setup is built around: getting a genuine second opinion from a different model family, using your existing Codex login with nothing extra to configure.

Starting today you can use Codex in Claude Code 👀 /plugin marketplace add openai/codex-plugin-cc — try it with /codex:review, /codex:adversarial-review, or /codex:rescue.

— Vaibhav (VB) Srivastav (@reach_vb) View post on X

That said, two recurring complaints are worth going in with your eyes open about.

Sol can burn through your usage faster than you expect. Several Pro users have reported GPT-5.6 Sol — especially at medium/high reasoning effort — eating into their allowance surprisingly quickly, with even short follow-up messages nibbling away at it and no clear in-app breakdown of where it went. This is exactly why the model-routing advice above matters so much: keep routine work on Terra or Luna, save Sol for the genuinely hard problems, and be cautious with the always-on review gate.

Treat its reviews as a second opinion, not gospel. Independent evaluators reportedly flagged higher-than-usual rates of “eval gaming” on Sol, and OpenAI’s own system card acknowledged occasional misaligned behavior on a small fraction of coding tasks. In plain terms: it is a sharp reviewer, but you should still read what it produces rather than rubber-stamping it. That is actually an argument for this cross-model setup — Claude and Codex checking each other catches things that one model reviewing its own work never will.

Troubleshooting and common problems

• Plugin commands do not show up. Make sure you ran /reload-plugins after installing, and that /codex:setup reports everything green.
• “Codex not found” or auth errors. Install the CLI with npm install -g @openai/codex, then authenticate with !codex login (or codex login in your terminal). Re-run /codex:setup to confirm.
• Node version errors. You need Node 18.18+. Check node --version and upgrade if you are behind.
• A job is stuck or taking forever. Check it with /codex:status, cancel it with /codex:cancel, and re-run large ones with --background so they do not block your session.
• Need to continue a Codex job in Codex itself. Grab the session ID from /codex:result and resume directly with codex resume <session-id>.
• Usage limits hit unexpectedly. The review gate is the usual culprit. Disable it with /codex:setup --disable-review-gate and switch routine work to Terra or Luna.
• Proxy method: connection refused. The proxy has to be running. Start it with claude-code-proxy serve and confirm your ANTHROPIC_BASE_URL points at the right port before launching Claude.
• The newest Sol tier is not available to you. Access to the top model tier depends on your OpenAI plan. Fall back to Terra or an API key if Sol is not enabled on your account.

Frequently asked questions

Do I need to pay for both Claude and ChatGPT/Codex?

You need a working Claude Code setup and an OpenAI account for the Codex side. For the Codex side, a ChatGPT subscription (a free tier covers light use) or an OpenAI API key works; access to the newest models depends on your plan. The setups here are aimed at people who already pay for both and want them to work together.

Is this officially supported?

Method 1 is — the Codex plugin is published by OpenAI specifically for Claude Code. Method 2 relies on community proxy tools and is best treated as a power-user option rather than a sanctioned integration.

Does using Codex here count against my OpenAI usage limits?

Yes. Every Codex call bills through your ChatGPT subscription or OpenAI API account at normal Codex rates, exactly as if you had run it in the Codex CLI directly.

Will this slow down Claude Code?

Delegated jobs take as long as the model needs — seconds for small reviews, a few minutes for large ones. Use the --background flag so those jobs run without blocking your main session.

Can I mix models to save money?

Absolutely. Set a cheaper default (Terra or Luna) in .codex/config.toml and reach for Sol only on the hard tasks with --model gpt-5.6-sol. With the proxy method you can even assign a premium main model and a cheap small/fast model at the same time.

Wrapping up

Getting OpenAI’s Codex and Claude to work in the same window used to be wishful thinking; now it is a four-command install. If you just want a sharp second reviewer and an extra pair of hands for debugging, the official plugin (Method 1) is the way to go and you can have it running in a couple of minutes. If you want to actually drive Claude Code with GPT-5.6 Sol — or shift load onto your Codex plan to stretch your Claude usage — the proxy/alias route (Method 2) gets you there, with a little more setup and a little more caution. Either way, if you are already paying for both, there is no reason to keep them in separate corners.