Use GPT-5.6 in Claude Code for video automation work

What this session is for

The video-automation jobs we hand to a GPT-5.6 session

This site runs a video render API, a CapCut automation API and caption tools. Most of the work around those products is not creative — it is JSON. Scenes, timings, asset URLs, caption styles, webhook configs. That is exactly the kind of work a coding agent is good at, and the isolated session below lets you throw GPT-5.6 at it without touching your daily Claude configuration.

Drafting render payloads. Describe the video in plain language — scenes, text overlays, aspect ratio — and let the model draft the JSON. Validate the field names against the API reference before you render: models invent fields with great confidence. The JSON to video guide shows what a valid payload looks like.
CapCut automation calls. Turning an edit checklist ("trim silences, add captions, export 9:16") into a sequence of CapCut API calls is repetitive translation work. The model drafts, you review the parameters.
Caption cleanup scripts. One-off scripts that fix SRT timing drift, strip filler words or restyle captions before they go through the caption pipeline.
n8n workflow JSON. HTTP-node configs for render webhooks and polling loops, following the n8n & Make setup.

A concrete first prompt for the session you are about to set up:

Claude Code · example prompt

Draft a JSON payload for a 9:16 video with three scenes:
1. a hook line as large text over background footage, ~2 seconds
2. a product screenshot with a caption bar at the bottom
3. a call-to-action scene with a logo and one button-style text

Use placeholder asset URLs. List every field you were unsure
about at the end, so I can check them against the API reference
before rendering.

Understand this first

Why your existing Claude Code setup stays intact

The separation comes from three things: a different working directory, a different configuration directory and environment variables that exist only for this one process. Nothing is written to your shell profile, and no proxy settings are added to ~/.claude/settings.json.

Claude Code (isolated session) → 127.0.0.1:18765 (local proxy) → ChatGPT/Codex (GPT-5.6)

Component	Your normal Claude Code	GPT-5.6 demo
Launch command	`claude`	temporary environment command or `claude-gpt56`
Configuration	`~/.claude`	`~/.claude-gpt56`
Working directory	your own project	`~/Code/gpt56-demo`
Model route	your existing Anthropic route	local proxy on port 18765
Customizations	your normal hooks, plugins and skills	disabled with `--safe-mode`

Terminal · preparation

Install the proxy without replacing Claude Code

1. Record which Claude Code binary you use now

This command stops with a clear message if Claude Code or Homebrew is missing. It only displays the Claude binary it finds; it does not change anything.

Terminal · preflight

REAL_CLAUDE="$(command -v claude)"
test -n "$REAL_CLAUDE" || { echo "Claude Code not found"; exit 1; }
printf 'Claude: %s\n' "$REAL_CLAUDE"
"$REAL_CLAUDE" --version
brew --version

2. Create a neutral demo repository

Claude Code will open here, not in /Users/yourname or a real client project.

Terminal · demo directory

mkdir -p "$HOME/Code/gpt56-demo"
git -C "$HOME/Code/gpt56-demo" init -q
cd "$HOME/Code/gpt56-demo"
pwd

3. Install only the local bridge

Homebrew installs claude-code-proxy alongside Claude Code. It does not overwrite the original claude binary.

Terminal · installation

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

4. Sign in with your ChatGPT account

An OpenAI sign-in page opens in your browser. Use your ChatGPT Plus or Pro account, not a separate OpenAI API key.

Terminal · ChatGPT sign-in

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

If the browser does not open, use the device-code flow:

Terminal · alternative sign-in

claude-code-proxy codex auth device

Two Terminal windows

Start the proxy, then launch the isolated Claude session

Terminal 1: keep the local proxy running

Leave this window open. The monitor will show the actual requests, the selected upstream model and any errors.

Terminal 1 · proxy

claude-code-proxy serve

Terminal 2: launch GPT-5.6 from the demo directory

Paste the entire block at once. Every proxy variable applies only to the Claude process launched at the bottom.

Terminal 2 · GPT-5.6 in Claude Code

REAL_CLAUDE="$(command -v claude)"
test -n "$REAL_CLAUDE" || { echo "Claude Code not found"; exit 1; }

mkdir -p "$HOME/Code/gpt56-demo"
git -C "$HOME/Code/gpt56-demo" init -q
cd "$HOME/Code/gpt56-demo"

env -u ANTHROPIC_API_KEY \
  -u CLAUDE_CODE_USE_BEDROCK \
  -u CLAUDE_CODE_USE_VERTEX \
  -u CLAUDE_CODE_USE_FOUNDRY \
  CLAUDE_CONFIG_DIR="$HOME/.claude-gpt56" \
  IS_DEMO=1 \
  ANTHROPIC_BASE_URL=http://127.0.0.1:18765 \
  ANTHROPIC_AUTH_TOKEN=unused \
  ANTHROPIC_MODEL='gpt-5.6-sol[1m]' \
  ANTHROPIC_SMALL_FAST_MODEL='gpt-5.6-luna[1m]' \
  CLAUDE_CODE_AUTO_COMPACT_WINDOW=372000 \
  CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
  CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 \
  "$REAL_CLAUDE" \
    --safe-mode \
    --setting-sources user \
    --effort high \
    --name "GPT-5.6 demo"

CLAUDE_CONFIG_DIR: Moves settings and session history for this run to ~/.claude-gpt56.
IS_DEMO=1: Hides account and organization details in the Claude Code interface during your recording.
ANTHROPIC_BASE_URL: Routes Claude Code to the proxy on your own Mac for this process only.
ANTHROPIC_MODEL: Tells the proxy to map the request to GPT-5.6 Sol. Luna handles smaller background requests.
[1m] and 372000: [1m] is only a local Claude Code hint. The proxy removes the suffix; according to the proxy documentation, the real upstream context is 372K tokens, not one million.
--safe-mode: Disables local hooks, plugins, MCP servers, skills and other customizations for this session.

Verify before you record

Verify the route, not what the model says about itself

A model can misidentify itself. Check Terminal 1 instead: the proxy should show a successful Codex request using gpt-5.6-sol. Also verify the sign-in status.

New Terminal tab · verification

claude-code-proxy codex auth status
lsof -nP -iTCP:18765 -sTCP:LISTEN

Then open a clean Terminal tab and prove that your normal shell did not inherit any proxy variables:

Clean Terminal tab · normal Claude

env | grep -E 'ANTHROPIC_BASE_URL|ANTHROPIC_MODEL|CLAUDE_CONFIG_DIR'
command -v claude
claude --version

The first line should produce no output. command -v claude should show the same path it showed before installation.

Optional · easier to repeat

Create one separate command: `claude-gpt56`

After the first test, you can put the long launch command in a transparent local script. It deliberately does not create an alias named claude and does not change your PATH. You still need Terminal 1 running claude-code-proxy serve.

One time · create launcher

mkdir -p "$HOME/.local/bin"

tee "$HOME/.local/bin/claude-gpt56" >/dev/null <<'SCRIPT'
#!/bin/zsh
set -eu

REAL_CLAUDE="$(command -v claude || true)"
if [[ -z "$REAL_CLAUDE" ]]; then
  echo "Claude Code not found"
  exit 1
fi

DEMO_DIR="${GPT56_DEMO_DIR:-$HOME/Code/gpt56-demo}"
mkdir -p "$DEMO_DIR"
if [[ ! -d "$DEMO_DIR/.git" ]]; then
  git -C "$DEMO_DIR" init -q
fi
cd "$DEMO_DIR"

exec env -u ANTHROPIC_API_KEY \
  -u CLAUDE_CODE_USE_BEDROCK \
  -u CLAUDE_CODE_USE_VERTEX \
  -u CLAUDE_CODE_USE_FOUNDRY \
  CLAUDE_CONFIG_DIR="$HOME/.claude-gpt56" \
  IS_DEMO=1 \
  ANTHROPIC_BASE_URL=http://127.0.0.1:18765 \
  ANTHROPIC_AUTH_TOKEN=unused \
  ANTHROPIC_MODEL='gpt-5.6-sol[1m]' \
  ANTHROPIC_SMALL_FAST_MODEL='gpt-5.6-luna[1m]' \
  CLAUDE_CODE_AUTO_COMPACT_WINDOW=372000 \
  CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
  CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 \
  "$REAL_CLAUDE" \
    --safe-mode \
    --setting-sources user \
    --effort high \
    --name "GPT-5.6 demo" \
    "$@"
SCRIPT

chmod 700 "$HOME/.local/bin/claude-gpt56"

From now on, start Terminal 1 and then run your separate launcher:

Terminal 2 · future sessions

"$HOME/.local/bin/claude-gpt56"

You can select a different neutral directory for any launch:

Terminal 2 · different demo directory

GPT56_DEMO_DIR="$HOME/Code/video-demo" "$HOME/.local/bin/claude-gpt56"

Complete rollback

Remove the proxy without deleting your normal Claude data

Sign out, uninstall the Homebrew package and move only the two items created by this guide to the Trash. This guide deliberately does not use rm -rf.

Terminal · remove proxy

claude-code-proxy codex auth logout
brew uninstall claude-code-proxy

Terminal · move demo config and launcher to Trash

STAMP="$(date +%Y%m%d-%H%M%S)"

if [[ -e "$HOME/.claude-gpt56" ]]; then
  mv "$HOME/.claude-gpt56" "$HOME/.Trash/claude-gpt56-$STAMP"
fi

if [[ -e "$HOME/.local/bin/claude-gpt56" ]]; then
  mv "$HOME/.local/bin/claude-gpt56" "$HOME/.Trash/claude-gpt56-launcher-$STAMP"
fi

command -v claude
claude --version

Your demo repository can stay where it is. To remove it as well, move ~/Code/gpt56-demo to the Trash manually in Finder.

Targeted fixes

Common problems

Port 18765 is already in use

First find the local process that is listening. Stop it only if you are certain it is an old proxy process.

Terminal · find listener

lsof -nP -iTCP:18765 -sTCP:LISTEN

Your account returns “model not supported”

GPT-5.6 availability is rolled out by account and subscription. First check whether GPT-5.6 Sol is available in your Codex environment. A proxy cannot grant model access that your account does not have.

`--safe-mode` is not recognized

Your Claude Code version is probably too old for this recording workflow. Update your existing installation, then check again:

Terminal · update Claude Code

claude update
claude --version
claude --help | grep -- '--safe-mode'

The browser sign-in gets stuck

Try the device-code flow with claude-code-proxy codex auth device. Never show tokens, Keychain items or complete debug logs in your video.

Claude Code starts, but requests fail

Check Terminal 1. The proxy monitor is the best source for a missing login, an unsupported model, a rate limit or a protocol break after a Claude Code update.

Route without a community proxy

Use the official Codex CLI when reliability matters more

OpenAI supports ChatGPT sign-in directly in Codex. For GPT-5.6, OpenAI lists Codex CLI 0.144.0 as the minimum version. You do not get the Claude Code interface, but you do get the officially supported sign-in and model route.

Terminal · official Codex route

npm install -g @openai/codex@latest
codex --version
codex login
codex

Read OpenAI's current explanation of using Codex with a ChatGPT plan.

Why this workflow exists

The Z.ai helper inspired the experience, not the sign-in method

Z.ai uses an interactive helper to show how to load an external coding plan into Claude Code. That route requires a Z.ai API key and manages tool configuration. Here, the goal is to sign in with ChatGPT without letting a wizard manage your existing Claude installation. That is why this guide uses separate, auditable commands and process-scoped variables.

Posts from Theo and Tibo prompted this guide. They are not official product documentation. The installation above is therefore based on current proxy, OpenAI and Anthropic documentation, not on a single screenshot or reply.

Checked July 11, 2026

Sources and claim boundaries

raine/claude-code-proxy — installation, ChatGPT OAuth, port, model names, 372K compaction and proxy limitations.
Anthropic: Other LLM gateways — Anthropic does not support routing Claude Code to non-Claude models.
Anthropic: environment variables — behavior of CLAUDE_CONFIG_DIR, IS_DEMO, ANTHROPIC_BASE_URL and the traffic/fallback variables.
OpenAI: use Codex with your ChatGPT plan — the official ChatGPT sign-in route.
OpenAI: GPT-5.6 in ChatGPT — model availability and the minimum Codex CLI version.
OpenAI: GPT-5.6 — the official model introduction.
Z.ai Coding Tool Helper — reference for the guided tool experience; it is not the ChatGPT sign-in route.

Short answers

Frequently asked questions

Can GPT-5.6 in Claude Code build my video automations?

It drafts JSON render payloads, CapCut API sequences and caption scripts well — that is the JSON-heavy work described at the top of this page. Always validate field names against the API reference before rendering: language models invent plausible-looking fields, and a render job fails on exactly those.

Will this break my normal Claude Code setup?

The commands on this page do not change your normal Claude configuration. They use a separate configuration directory, a separate project directory and temporary process variables. Software updates or a bug in the external proxy remain general risks.

Do I need an OpenAI API key?

No. The community proxy opens a ChatGPT/Codex OAuth sign-in. Use your ChatGPT account. For the officially supported route, you can sign in to Codex directly with the same account.

Is this officially supported?

No. The official combination is GPT-5.6 in Codex. Routing Claude Code through a community proxy to a non-Claude model falls outside Anthropic support.

Does the model name in Claude Code prove GPT-5.6 is running?

Not by itself. Use the proxy monitor as technical evidence of the upstream model and a successful request. Do not ask the model to identify itself.

Why does the model name include [1m]?

It is a local hint that prevents Claude Code from compacting too early. The proxy removes the suffix before the upstream request. The documented GPT-5.6 context for this route is 372K tokens.

What actually runs inside Claude Code

Claude Code is built for Anthropic models, but an LLM proxy can translate its Anthropic-style requests for OpenAI models. That is how GPT-5.6 in Claude Code works: Claude Code talks to a local or hosted gateway, and the gateway handles authentication, model routing, streaming, and protocol conversion.

This is an unofficial compatibility route. Claude Code does not turn into Codex, and a ChatGPT login is not an Anthropic API key. Tools, prompt handling, token reporting, caching, and errors can behave differently through a proxy.

Route	Best use	Main limitation
Claude Code with Anthropic	Native Claude Code behavior	Requires supported Anthropic access
Claude Code LLM proxy	Running GPT or another compatible backend	Depends on third-party translation
Codex CLI	Native OpenAI coding workflow	Different CLI, settings, and tool behavior

For media workflows, the same gateway pattern can sit beside an AI video API integration, but the coding model and video model remain separate services.

Running GPT-5.6 Sol through the Raine proxy

raine/claude-code-proxy is a third-party Claude Code reverse proxy. Its documented flow is simple: install the proxy, authenticate it with an eligible ChatGPT account, start its local server, then launch the Claude Code CLI with the proxy URL and GPT model name.

A typical shell setup uses ANTHROPIC_BASE_URL for the local Claude proxy URL, ANTHROPIC_AUTH_TOKEN for the value expected by Claude Code, and ANTHROPIC_MODEL for gpt-5.6-sol. Use the exact model suffix and context settings documented by the proxy version you installed. Examples copied from GitHub or Reddit can become stale.

Run the proxy login command and confirm which ChatGPT account it uses.
Start the proxy and keep its diagnostics visible.
Export the required environment variables in the same terminal session that starts Claude Code.
Send a small prompt before enabling tools or long-running commands.
If streaming fails, inspect the proxy log before changing unrelated Claude settings.

If you want to automate the surrounding workflow, the n8n setup guide explains how to keep credentials and execution steps outside the prompt.

Environment variables, settings files, and managed policy

PATH controls where your shell looks for the Claude Code CLI and proxy executable. It does not select a model. In VSCode, the integrated terminal may inherit a different PATH or environment from your login shell, so compare the output of which claude and which claude-code-proxy in both places.

Claude Code settings JSON controls supported client behavior, permissions, hooks, and policy. Environment variables provide runtime values such as the Claude Code API key environment variable, proxy endpoint, or model name. Managed settings can override user and project settings. The Claude managed settings location and Claude Code settings schema vary by operating system and release, so check the current Claude Code documentation or the CLI diagnostics instead of guessing a path.

Value	Purpose
`ANTHROPIC_BASE_URL`	Points Claude Code at an Anthropic-compatible gateway
`ANTHROPIC_API_KEY` or `ANTHROPIC_AUTH_TOKEN`	Supplies authentication expected by the selected route
`ANTHROPIC_MODEL`	Selects the model identifier exposed by the gateway
`HTTP_PROXY` and `HTTPS_PROXY`	Route network traffic through a corporate HTTP proxy
`PATH`	Lets the shell find installed commands

CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, CLAUDE_CODE_ATTRIBUTION_HEADER, CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK, and CLAUDE_CODE_DISABLE_COMMAND_INJECTION_CHECK should only be used when your installed Claude Code or proxy documentation explicitly supports them. Do not disable an injection check just to make an untrusted command run.

GPT-5.6 models, access, and context

GPT-5.6 is a model family. GPT-5.6 Sol is the high-capability tier, GPT-5.6 Terra balances capability and operating cost, and other tiers may be available for faster work. GPT-5.6 Sol Pro and Sol Ultra labels can refer to product modes, reasoning levels, or third-party aliases rather than separate public API model IDs. Check the model list returned by the product you are actually using.

GPT-5.6 Sol is available in Codex for eligible accounts. Free access, GPT-5.6 subscription access, and selectable reasoning levels depend on the current ChatGPT plan, workspace policy, rollout, region, and Codex version. If Codex says the model is not supported when using a ChatGPT account, update Codex, inspect the available model list, and try the supported model identifier. Do not work around an account restriction with a copied token.

The GPT-5.6 Sol context window is large, but the usable amount can be reduced by client limits, proxy translation, automatic compaction, tool output, and pricing rules for long prompts. Values such as 372000 in a GitHub configuration are often proxy-specific compaction thresholds, not proof of the model's full context window.

Text prompts can also generate structured scene data for a JSON-to-video rendering pipeline; keep that render schema separate from Claude Code settings JSON.

Costs, subscriptions, and privacy boundaries

There is no single GPT-5.6 pricing figure that applies to every route. ChatGPT subscriptions use plan limits or shared usage allowances. Codex may use plan allowances, credits, or token-based billing. The OpenAI API charges for input tokens, cached input, output tokens, and sometimes separate tools. Rates change, so check the current rate card before estimating a production workload.

The same rule applies to GPT-5.6 Sol Ultra cost: first identify whether Ultra means a reasoning setting in ChatGPT or Codex, a workspace credit rate, or a third-party proxy alias. API access and a ChatGPT subscription are separate products unless the current product documentation explicitly connects them.

A proxy can see prompts, code, tool results, and credentials that pass through it. Review its source, bind a local proxy to localhost, keep logs private, and never paste your ChatGPT login and password into an unknown proxy. ChatGPT data controls also depend on whether you use a personal account, managed workspace, API, or third-party gateway.

Login, desktop access, and common failures

Use the official OpenAI login page to access ChatGPT. ChatGPT login with Google and chat.openai.com or the current official login route should return you to the account connected to that Google identity. If ChatGPT login is not working, check that you chose the original sign-in method, disable only extensions that affect authentication, allow required cookies, and try a private window before resetting anything.

For Chat GPT login free access, create or open an OpenAI account through the official site. Never send a password to a GitHub project, Claude Code plugin, Docker image, or reverse proxy. A legitimate OAuth flow opens an OpenAI-controlled page and returns a scoped session to the application.

The ChatGPT desktop app is available through official OpenAI download channels. Use the official installer for ChatGPT for Mac, verify the publisher, and avoid repackaged downloads. Codex CLI is installed and authenticated separately from the normal chat window, even when it can use the same eligible ChatGPT account.

Choosing Codex, Claude Code, or a proxy

Use Codex when you want native OpenAI model support and its own coding tools. Use Claude Code when you want its native CLI, permissions, hooks, and Anthropic integration. Use a Claude Code Router or GPT plugin only when keeping the Claude Code interface matters enough to accept an extra compatibility layer.

There is no universal winner in “Claude or GPT better at code.” Results depend on the repository, prompt, tool access, context management, latency, and how well the agent checks its work. Test both on one real task with the same acceptance check. Reddit threads and GPT-5.6 Artificial Analysis results can help form a shortlist, but they do not reproduce your codebase or proxy setup.

Claude and ChatGPT can exchange text or files when you move the data between them, use an approved integration, or call their APIs from one workflow. They do not automatically share chats, memories, passwords, or subscriptions. Claude can draft instructions for a custom GPT, but creating and publishing that GPT happens in the relevant OpenAI product.

Questions people ask

Can Claude Code run GPT models?

Not as its native provider route. It can send requests through an Anthropic-compatible LLM proxy that translates them for GPT models. Tool behavior and streaming may differ from native Claude Code.

How do I run GPT-5.6 Sol inside Claude Code?

Start a compatible proxy such as raine/claude-code-proxy, authenticate it through its documented flow, and point ANTHROPIC_BASE_URL at the local server. Set the model name exposed by that proxy, then test a small prompt before using tools.

Is GPT-5.6 Sol available in Codex?

Yes, for eligible accounts and supported Codex versions. Free access and model choices depend on the current plan and rollout. If the model is rejected, update Codex and inspect its available model list.

Can I use Codex with a ChatGPT subscription?

Eligible ChatGPT plans can authenticate with Codex, subject to current plan limits and workspace policy. A ChatGPT subscription does not automatically provide an OpenAI API key or unrestricted API usage.

Can I use GPT-5.5 in Codex or Claude Code?

Codex can use GPT-5.5 where that model remains available for your account and client version. Claude Code needs a compatible gateway that exposes GPT-5.5. Availability can change, so check the live model list.

Is GPT-5.6 free?

GPT-5.6 access depends on the product, plan, and model tier. Free accounts may receive a different GPT-5.6 tier than paid accounts and may not receive Sol. Restrictions can also come from rollout, usage limits, safeguards, or workspace settings.

How much does GPT-5.6 Sol Ultra cost?

There is no fixed universal charge. The cost depends on whether Ultra is a reasoning mode, a Codex credit setting, or an API model route, plus input, cached input, output, and tool usage. Check the current rate card because tariffs change.

How do I set a Claude proxy?

Use HTTP_PROXY and HTTPS_PROXY for a corporate network proxy. Use ANTHROPIC_BASE_URL when directing Claude Code to an Anthropic-compatible LLM gateway. A Claude subscription is not a free proxy service.

What does Fal AI do, and can I use Fal AI for free?

Fal AI provides hosted inference for generative media models, including image, video, and audio workflows. Trial credits or limited access may be offered, but there is no permanent free-use guarantee. Production billing usually depends on the selected model and its execution, compute, credit, or media-duration unit.

How do I log in to ChatGPT with Google?

Open the official OpenAI login page and choose the Google account originally linked to ChatGPT. If it loops or opens the wrong account, try a private window and confirm the selected Google identity. Never enter that password into a proxy.

Is my data private on ChatGPT?

Privacy depends on the product, account type, data controls, workspace policy, and any third-party proxy you add. A proxy may process or log prompts before they reach OpenAI. Do not send secrets until you have checked every service in the path.

How can a beginner start using AI?

Start with one small, checkable task such as rewriting a paragraph or explaining an error. Give the model the goal, relevant context, and the format you want. Check the result against the source instead of treating it as automatically correct.

How do I put GPT data into Claude or change GPT to Claude?

Copy approved text, upload an exported file, or connect both APIs through your own workflow. You cannot convert one model into the other, and their chat histories do not sync automatically. Remove private data before transferring it.

Is there a Claude Max API proxy or Codex plugin for Claude Code?

Third-party routers and proxies exist, but a Claude Max or ChatGPT subscription is not automatically an API entitlement. Check the project's source, authentication method, maintenance status, and license. Prefer a local, auditable gateway over an unknown hosted endpoint.

Is GPT 5.6 free to use?

No. It runs through a paid plan or the API. In a coding agent the cost is driven by how much context you re-send each turn, not by the model tier alone.

How can I use AI as a beginner?

Get a key, run the provider's own example unchanged, then change one field at a time. The mistake that costs a day is starting from a large workflow instead of from one working call.