<- all tokdocs

Can You Run GPT-6 Astra Inside Claude Code? How the Codex Plugin Works

By , founder of Mavric and Apso

View on TikTok ->

Claude Code cannot run GPT-6 Astra as its model. What you can do is install OpenAI's codex-plugin-cc, which lets Claude Code call the separate Codex CLI, and Codex can be set to use Astra. A TikTok from @ai.honeycove calls this running Astra "inside of Claude Code" and shows a chart claiming it cuts usage to 38%, a figure with no measurement behind it.

Is GPT-6 Astra real?

Yes. OpenAI's announcement in its developer community, "Introducing GPT-6-Astra: The most intelligent and aligned model in the world," is dated September 3, 2026. It gives the API id gpt-6-astra, a 1,050,000-token context window, 128,000 tokens of maximum output and an April 30, 2026 knowledge cutoff. Pricing is $10.00 per million input tokens, $1.00 per million cached input tokens, $12.50 per million cache writes and $50.00 per million output tokens. The video, posted October 7, came 34 days after that announcement.

Claude Fable 5.1, the "boss" model in the video, is also real. Anthropic's September 2026 announcement keeps standard pricing at "$10 per million input tokens and $50 per million output tokens," with cache reads at $0.25 per million, and estimates Fable 5.1 will cost about 25% less than Fable 5 for typical workloads.

Can Claude Code use OpenAI models?

No. Anthropic's model configuration docs list Anthropic models only. The aliases are default, best, fable, sonnet, opus, haiku, sonnet[1m], opus[1m] and opusplan, and every full model id is a Claude model. The supported providers are the Anthropic API, Amazon Bedrock, Microsoft Foundry, Google Cloud's Agent Platform, the Claude Platform on AWS, the Claude apps gateway, and LLM gateways reached through ANTHROPIC_BASE_URL.

A plugin cannot change that. Anthropic's plugins overview describes a plugin as "a directory of skills, agents, hooks, MCP servers, or other components." A model is not one of the component types.

The video's own screen recording shows what happens. Claude Code runs a shell command, codex exec --sandbox read-only --ephemeral ... --output-last-message "...\codex_review.txt", and reads the result from a text file. OpenAI's README describes the plugin as a way to "Use Codex from inside Claude Code for code reviews or to delegate tasks to Codex." Codex runs as a separate process, on a separate account, with its own model.

How to set up the Codex plugin for Claude Code

The video shows the steps:

  1. Install the Codex CLI with npm install -g @openai/codex.
  2. In Claude Code, run /plugin marketplace add openai/codex-plugin-cc.
  3. Run /reload-plugins, then /plugin install codex@openai-codex.
  4. Add instructions to CLAUDE.md telling Claude when to hand work to Codex.

OpenAI's README requires a "ChatGPT subscription (incl. Free) or OpenAI API key" and Node.js 18.18 or later. The CLAUDE.md in the video is seven plain-English rules under "# AI Collaboration Rules," including "Do not use Astra or Codex for every small task" and "The main Claude Code session remains responsible for the final decision." It contains no model setting or routing logic, so Claude follows it the way it follows any instruction.

Which model does Codex actually use?

It depends on configuration the video does not show. OpenAI's README says "if you do not pass --model or --effort, Codex chooses its own defaults." The codex exec command on screen passes no --model. Codex CLI rust-v0.159.1, released September 29, "Added GPT-6.1 Sol as the default model," so a default run after that date reaches Sol, not Astra.

The video also loads a third-party skill, astra-advisor by PatrickSys, which does pin the model. Its README says it "spawns one native subagent with model=gpt-6-astra, reasoning_effort=low, and fork_turns=none." It requires Node.js 22+ and a Codex login with access to gpt-6-astra, and it is marked experimental at v0.1.1. The video never names this skill aloud. So one consult in the video plausibly reached Astra, and the separate review command did not pin a model.

Does this cut token usage by 62%?

Nothing supports that. The video's chart shows "CLAUDE ONLY ... 100%" against "FABLE + ASTRA ... 38%." OpenAI's plugin README makes no token-savings claim. The astra-advisor README says the release "does not establish that Astra Advisor improves coding quality, catches more bugs, saves money, or is independent in a statistical sense." No frame shows a measured Claude-only run to compare against.

Per token, the two models cost the same: $10 per million input and $50 per million output. Delegating to Astra moves the bill to a different vendor.

What does change is which quota you spend. Anthropic's support article says "Both Pro and Max plans have a five-hour session limit and a weekly limit," and Max plans add "a separate weekly limit for Fable." Sending work to Codex bills it to your ChatGPT account, so you draw on two quotas instead of one. Total tokens go up, because the handoff adds a written brief, a scratchpad file, a review request and a review result.

Plugins add a standing cost too. Anthropic's plugins page says the name and description of every skill, agent and command "are in Claude's context on every turn," and "Those tokens count toward your usage." The user in the video has 8 plugins, 83 skills, 6 agents and 35 plugin MCP servers loaded.

Is the Codex plugin new?

No. The video opens with "Most people don't know." OpenAI released codex-plugin-cc v1.0.0 on 30 March 2026, 191 days before the video, and v1.0.6 on 8 July, 91 days before. It is an official OpenAI repository.

Bottom line

Claude Code runs only Anthropic models. The Codex plugin lets it call a separate Codex process, which uses Astra only if you pin it, as the astra-advisor skill does. The setup spreads work across a Claude quota and a ChatGPT quota, but the 38% chart has no measurement behind it, and the two models cost the same per token.

Sources

Published October 7, 2026. Based on a TikTok by @ai.honeycove. Full video notes (markdown)