Use GPT-6 Astra in Claude Code on Windows: A Follow-Along Guide for Preserving Sessions

The installation itself was not the hardest part. Claude Code binary detection, a stale model picker, isolated configuration directories, and an invisible 303 MB session transcript all became separate problems.

/ 01 · SECTION

Why I considered this setup

I currently subscribe to both Claude Code Max and ChatGPT Pro while developing an application. The two services already had distinct roles in my workflow.

The combination worked well until the intensity of development increased. My Claude Code weekly allowance was regularly exhausted by Thursday.

I tried purchasing additional usage credits so that development could continue. However, USD 400 in extra credits was consumed in roughly one day. The issue was no longer an occasional overage. Both the cost and the interruption to the main development flow had become unsustainable.

Why could $400 disappear faster than a $200 Max subscription?

The two amounts do not purchase usage in the same way. Anthropic describes Max 20x as providing 20 times the usage of Pro within each five-hour usage window. Extra usage credits, on the other hand, are consumed at standard API rates after the included plan allowance is exhausted.

Anthropic does not publish the included Max allowance as a fixed token quantity or an equivalent API-dollar balance. It is therefore not possible to calculate what exact percentage of a Max allowance $400 in usage credits represents. The meaningful observation from my own workload was simpler: the weekly Max allowance was gone by Thursday, and another $400 lasted about one more day.

Would two Max 20x subscriptions solve it?

In terms of capacity, that looked close to what my workload needed. The obvious solution would have been to combine two Max 20x allowances in the same development workflow.

That was not practical in my setup. There was no way to stack two Max 20x subscription allowances inside one Claude Code account and the same long-running session. Using a separate account would separate authentication and session history, defeating the requirement that mattered most: preserving context.

Could I keep the existing Claude Code session and interface, but use a Codex model through my ChatGPT subscription after the Claude Max weekly allowance ran out?

That question led to the experiment documented below.

node --version
claude --version
codex --version

codex login

npm install -g codex-for-claude-code@latest

cd C:\path\to\your-project
claude-codex --model gpt-6-astra

$env:CLAUDE_CODEX_CLAUDE_BIN="$HOME\.local\bin\claude.exe"
claude-codex --model gpt-6-astra

/resume

/status

/ 02 · SECTION

Verified environment

OpenAI documents two authentication paths for local Codex work: signing in with ChatGPT for subscription access, or using an API key for usage-based access. This experiment used the first path. Usage limits still depend on the subscribed ChatGPT plan.

/ 03 · SECTION

Codex authentication and bridge installation

I initially considered broader routing solutions such as OmniRoute and LiteLLM. My requirement was narrower: preserve the Claude Code experience, reuse an existing Codex CLI login, and avoid moving this workload to a separate API-key billing path.

codex-for-claude-code was the most direct match. It is important to note that it is a third-party open-source bridge, not an official OpenAI or Anthropic integration.

1. Verify the Codex CLI login

codex login

You can also launch Codex once:

codex

Choose Sign in with ChatGPT during the login flow. If the bridge later returns 401 Unauthorized, the project README recommends launching Codex once to refresh the OAuth token before retrying.

2. Check the prerequisites and install the bridge

node --version
claude --version
codex --version

The bridge requires Node.js 20 or later, Claude Code, and a logged-in Codex CLI.

npm install -g codex-for-claude-code@latest

The normal launch command is:

claude-codex

3. Fix Claude Code binary detection on Windows

In my environment, the wrapper could not automatically locate the Claude Code executable. Setting its exact path solved the problem:

$env:CLAUDE_CODEX_CLAUDE_BIN="$HOME\.local\bin\claude.exe"

cd C:\path\to\your-project
claude-codex

A first-run theme screen appeared. This was not an error. The bridge uses an isolated Claude configuration directory, so it initially behaves like a fresh environment.

/ 04 · SECTION

If Astra is missing from the model picker, pass it directly

After Claude Code opened, I ran /model. The bridge exposed several Codex options, but GPT-6 Astra was not among them.

Codex gpt-5.5
Codex gpt-5.4
Codex gpt-5.3-codex-spark
Codex gpt-5.4-mini
Codex gpt-5.3-mini

Claude Code supports selecting a model by its full name with --model. With a custom ANTHROPIC_BASE_URL or LLM gateway, the gateway can define the model identifiers that Claude Code passes through.

claude-codex --model gpt-6-astra

After launching it this way, the picker displayed:

gpt-6-astra    Custom model

/ 05 · SECTION

The larger problem: my existing session was missing

I ran /resume to reopen my long-running development session. Only the bridge test sessions appeared.

The reason was configuration isolation:

Standard Claude Code: ~/.claude
Bridge environment:   ~/.claude-codex

The bridge uses a separate CLAUDE_CONFIG_DIR so its Codex model-picker state does not leak into normal Claude Code windows. Its README describes shared configuration assets, but the existing transcript did not appear automatically in my Windows environment.

Direct resume by session ID also failed

I obtained the original UUID from /status in standard Claude Code and tried:

claude-codex `
  --model gpt-6-astra `
  --resume <SESSION_UUID>

The result was:

No conversation found with session ID: <SESSION_UUID>

The UUID was valid. Its JSONL transcript existed under ~/.claude, but not under the configuration tree used by the bridge.

/ 06 · SECTION

Safely copying one session

1. Locate the transcript

$sessionId = "<SESSION_UUID>"

$src = Get-ChildItem "$HOME\.claude\projects" -Recurse -File |
  Where-Object { $_.Name -eq "$sessionId.jsonl" } |
  Select-Object -First 1

$src | Select-Object FullName, Length

My transcript was approximately 303 MB.

2. Create the matching destination directory

if (-not $src) {
  throw "Session file not found: $sessionId"
}

$projectDirName = Split-Path $src.DirectoryName -Leaf
$dstDir = Join-Path `
  "$HOME\.claude-codex\projects" `
  $projectDirName

New-Item -ItemType Directory -Force -Path $dstDir |
  Out-Null

3. Copy and verify the file

$dst = Join-Path $dstDir $src.Name

Copy-Item `
  -LiteralPath $src.FullName `
  -Destination $dst

Get-Item $src.FullName, $dst |
  Select-Object FullName, Length

This creates a point-in-time copy, not live synchronization. Standard Claude Code and the bridge will develop separate histories after the copy.

/ 07 · SECTION

Launch GPT-6 Astra with the copied session

$env:CLAUDE_CODEX_CLAUDE_BIN="$HOME\.local\bin\claude.exe"

cd C:\path\to\your-project

claude-codex `
  --model "gpt-6-astra" `
  --resume $sessionId

The session opened. Running /status showed the essential evidence:

Session kind:       interactive
Anthropic base URL: http://127.0.0.1:3099
Model:              gpt-6-astra

Before publishing screenshots: redact email addresses, Windows usernames, sensitive repository paths, and the real session UUID.

Why this guide does not append [1m] to Astra

OpenAI lists a 1.05M-token context window for GPT-6 Astra. However, the current codex-for-claude-code 0.2.6 source does not explicitly register Astra in its context-window table, and its README does not document an Astra-specific [1m] suffix path.

How this differs from API-key billing

OpenAI distinguishes ChatGPT subscription authentication from API-key, usage-based access. This setup used the OAuth credentials created by Codex CLI; I did not enter a separate OpenAI API key into the bridge.

That does not mean unlimited free usage. The usage limits attached to the ChatGPT plan still apply, and this third-party bridge is not an officially supported integration.

/ 08 · SECTION

Warnings and limitations

1. Outdated permission rule syntax

Write(~/.ssh/**) is not matched by file permission checks
Use Edit(~/.ssh/**) instead

The warning indicated that the current Claude Code version expected Edit(path) for file-change permission rules. It did not block startup, but a non-matching protection rule should not be ignored.

2. Missing agents target

agents target missing

The shared agents directory did not exist. This was non-blocking because I was not relying on custom agent definitions in this test.

3. Large CLAUDE.md

Large CLAUDE.md will impact performance

The project instruction file exceeded the recommended size threshold. The session still launched, but a large instruction payload can affect latency and context consumption.

/ 09 · SECTION

The command I use now

The copied transcript already exists under ~/.claude-codex, so it does not need to be recopied for each launch.

$env:CLAUDE_CODEX_CLAUDE_BIN="$HOME\.local\bin\claude.exe"

cd C:\path\to\your-project

claude-codex --model "gpt-6-astra"

Then use /resume to select the bridge-side session, or provide the UUID directly:

claude-codex `
  --model "gpt-6-astra" `
  --resume "<SESSION_UUID>"

/ 10 · SECTION

Conclusion

What looked like a one-command model switch became a sequence of independent compatibility problems: executable detection, model catalog lag, configuration isolation, and session discovery.

Model missing from picker
→ Pass it directly with --model

Existing session missing
→ Check ~/.claude versus ~/.claude-codex

Resume by UUID fails
→ Copy only the required JSONL transcript

Final verification
→ Check proxy URL and model in /status

When a tool claims model support or session interoperability, verify the real configuration path, transcript store, and active model in your own operating environment.

On this day, /status, one filesystem path, and a single error line told me more than the happy-path installation instructions.

/ 11 · SECTION

References

• codex-for-claude-code on GitHub
• Claude Code CLI reference
• Claude Code model configuration
• OpenAI Codex authentication
• Using Codex with a ChatGPT plan
• OpenAI: GPT-6 Astra model specification
• OpenAI: ChatGPT Work and Codex · Astra CLI requirement
• Anthropic: What is the Max plan?
• Anthropic: Usage credits for paid plans