Claude Code, Codex Tokens Left Over? How to Use Multiple AIs Easily

3-Line TL;DR

• Paid personal plans for Claude Code, Codex and Antigravity all refresh usage on 5-hour and weekly schedules, yet managing three separate windows leaves two subscriptions sitting idle
• All three services include CLIs (command-line tools you run with text commands), and your main Claude Code session can call Codex and Antigravity directly through Python subprocess
• It does not save tokens, but it lets you combine what each AI does best from one window: research in Antigravity, images in Codex, control in Claude Code

Why Two of My Three AI Subscriptions Sat Idle

Made by HSL — data from Anthropic, OpenAI and Google Antigravity, checked Oct 20, 2026

• I currently pay for all three subscriptions: Claude, Codex (ChatGPT) and Antigravity (Google)
• For the first month I opened Codex and Antigravity one by one and managed each window myself, which was exhausting
• Gradually I ended up spending almost all of my tokens in Claude while the other two quotas went unused
• It felt like paying for three gym memberships and only visiting the one closest to home
• I considered keeping only Claude, though I hated losing Codex's image generation and the recently improved Gemini 4 Argon in Antigravity
• When Gemini 4 Argon first launched in Antigravity it was Ultra-only, and Google's current model table now lists it on paid non-trial AI Pro as well as Ultra in the app and CLI
• Paid tiers across all three services refresh on a 5-hour rolling window and a weekly cap (Claude limits, Codex pricing, Antigravity plans)
• Opening three separate windows by hand was the bottleneck

Item	Claude Code	Codex (ChatGPT)	Antigravity (Google)
Personal paid plans (US)	Pro $20/mo ($17 annual), Max from $100/mo (source)	Plus $20/mo, Pro $100, $200 or $500/mo (source)	AI Pro $20/mo, AI Ultra $100 or $200/mo (source)
CLI included	Pro and Max (not Free)	Plus and Pro	All plans; Gemini 4 Argon on paid Pro and Ultra (source)
Quota reset	5-hour session + weekly (source)	Plus: 5-hour + weekly; Pro: no 5-hour cap (source)	Pro and Ultra: 5-hour + weekly (source)
Quota check command	/usage (source)	/status (source)	/usage or /quota (source)
My current role	Overall control, development and fact-checking	Image generation, Reddit quotes and quick tasks	Research, drafting, text review and Python charts

Let Claude Code Call the Other Two Over CLI

Made by HSL — based on OpenAI non-interactive docs, Antigravity headless docs and author setup

• Then I realized Claude Code could drive both CLIs for me, and I wondered why I had not thought of something so simple sooner
• Claude Code has a built-in shell tool and isolated subagents that run side jobs in their own context and return a summary (Claude subagents)
• Codex CLI provides codex exec for headless mode (running without an interactive screen), sending progress to stderr and printing only the final answer to stdout (Codex docs)
• Adding - at the end of codex exec tells it to read the prompt from standard input (stdin), where another program pipes text in directly
• Antigravity CLI has agy -p to run one prompt and exit (Antigravity headless docs)
• When each CLI is signed in with your subscription account instead of an API key, runs draw from your monthly plan quota (Codex README, Antigravity headless docs)
• If you want a ready-made Codex bridge, OpenAI's official Claude Code plugin openai/codex-plugin-cc adds /codex:review, /codex:adversarial-review and /codex:rescue inside Claude Code

import subprocess

# Codex CLI: pass the prompt via stdin (-) so quotes never break
codex_res = subprocess.run(
    [
        "codex", "exec",
        "-m", "gpt-6.1-sol",
        "-c", 'model_reasoning_effort="high"',
        "-c", 'web_search="live"',
        "-",
    ],
    input=codex_prompt,
    text=True,
    encoding="utf-8",
    capture_output=True,
)

# Antigravity CLI: run Gemini 4 Argon for research, drafting, and text review
agy_res = subprocess.run(
    [
        "agy", "-p", argon_prompt,
        "--model", "gemini-4-argon-medium",
        "--dangerously-skip-permissions",
        "--print-timeout", "3300s",
    ],
    text=True,
    encoding="utf-8",
    capture_output=True,
)

Official Google Antigravity channel introduction to Antigravity CLI

How I Split the Work, and Five Lessons From Running It

Made by HSL — author workflow across Claude Code, Codex and Antigravity

• In my current setup, Antigravity (using Gemini 4 Argon) handles source research, text drafting, text review and the Python body charts
• Codex handles image generation and quick jobs, and it collected the Reddit quotes for this post
• Claude Code handles overall control, development, the thumbnail sketch and source checks
• This setup does not save tokens
• Claude spends its own share writing instructions and reading results, so total usage can be higher than doing everything in one AI
• What it buys you is picking the right AI for each job from one window
• OpenAI documents Codex's built-in gpt-image-2 skill ($imagegen) for interactive sessions and notes that images use included quota 3–5x faster on average than text turns (Codex image docs)
• On Windows, passing long prompts as PowerShell arguments broke on quotes and line breaks, so I switched to Python subprocess with stdin
• A helper AI returned dead Reddit links and a few numbers that differed from the source text
• Now Codex opens the original Reddit pages to collect quotes, and Claude Code checks every number against the primary source
• Having a helper AI review its own work and waiting for a second run only wasted time, and writing detailed instructions for a single run is faster
• Running all three at once caused one job to fail when two processes hit the same external login simultaneously, which an automatic retry loop fixed
• Giving the exact same task to the same model twice still produces noticeably different results

Community Reactions

• One user paying $20 for each prefers Claude for coding and sees little reason to keep ChatGPT beyond images (Reddit)
• A user who hands Claude's plans to Codex for review by hand asked how to automate the loop (Reddit)
• The developer of a tool where Claude Code drives Codex through codex exec says Codex runs on the normal subscription quota (self-promotion) (Reddit)
• A skeptic guesses that automating both burns more tokens than it saves (no measurements) (Reddit)

Q&A (Field Notes)

• Q. Does calling Codex and Antigravity from the CLI charge separate API fees?
  • Signing in to each CLI with your subscription account without setting pay-as-you-go API keys uses the quota included in your monthly plan (Codex README, Antigravity headless docs)
• Q. How fast does Codex image generation use up plan limits?
  • OpenAI's documentation states that built-in gpt-image-2 image generations consume included limits 3–5x faster on average than similar text turns (Codex image docs)
• Q. Does this still help if I only subscribe to two services?
  • Even with just Claude Code and Codex, installing openai/codex-plugin-cc or calling codex exec puts spare ChatGPT Plus or Pro quota to work on code reviews and side tasks