Claude Code vs Codex vs Google Antigravity (2026): Plans, Benchmarks and Cost Compared

Claude vs ChatGPT vs Gemini · Part 7 of 10

In this post we are going to cover the three AI coding agents: Claude Code from Anthropic, Codex from OpenAI, and Antigravity from Google.

A coding agent is more than autocomplete. You give it a task, and it reads the code, changes several files, runs the tests and fixes what broke. We go through seven questions, each with the short answer first and then the numbers.

Every price and limit in this post was checked on the official pricing and documentation pages on November 16, 2026. These numbers change often, so use the links at the end of the post to check them again before you pay for anything.

Part 7 of the Claude vs ChatGPT vs Gemini: The Complete 2026 Comparison series on TechBrothersIT.

1. What are Claude Code, Codex and Antigravity?

The short answer

They are coding agents. You describe a task, and the agent reads your repository, edits the files, runs commands and tests, and fixes its own mistakes. Each company has one.

An agent works in a loop. It reads the code it needs. It makes a plan and edits the files. It runs the tests or the build. If something fails, it reads the error and tries again. You review the result at the end.

Readthe repository
Editseveral files
Runtests and commands
Fixread the error, retry

Now the three products. Claude Code, from Anthropic, runs in the terminal, in the editor and in the Claude apps. Codex, from OpenAI, runs in the terminal, in the editor and inside ChatGPT. Google has Antigravity, its agent platform, and the older Gemini CLI for the terminal.

Claude
Claude Code
From Anthropic
Terminal, IDE and the Claude apps
Default model: Opus 5.5
ChatGPT
Codex
From OpenAI
Terminal, IDE and ChatGPT
Default model: GPT-6.1 Sol
Gemini
Antigravity
From Google
Agent platform, plus Gemini CLI
Model: Gemini 4 Argon
Default models as reported in October 2026. They change often.

2. Which subscription do I need for each coding agent?

The short answer

Codex is the easiest to start: it is on every ChatGPT plan, with limits on Free and Go. Claude Code needs the $20 Pro plan. Antigravity comes with Google AI Pro, and Gemini CLI now needs an API key or a business licence.

Here is the access table. Claude Code: not on Free. Included from Pro at 20 dollars, with 5 times the usage on Max. Codex: limited on Free and Go, expanded on Plus, and longer sessions on Pro. Antigravity: entry limits on AI Pro, and 5 or 20 times more on Ultra.

accessClaudeChatGPTGemini
Free planNoLimitednot listed
$20 planYesYes, expandedYes, entry limits
$100 to $200 plans5x · 10x usagelonger sessions5x · 20x limits
With an API keyYesYesYes Gemini CLI
Sources: claude.com/pricing, chatgpt.com/pricing, gemini.google/subscriptions, checked November 16, 2026

One change to know about. Google ended the free consumer sign-in for Gemini CLI on June 18, 2026. Today it works with a paid API key or a Gemini Code Assist licence, and individuals are pointed to Antigravity.

Keep in mind

Gemini CLI lost its free consumer sign-in on June 18, 2026. Individuals now use Antigravity.

3. How do I give the agent my project rules, and are these tools open source?

The short answer

Each one reads a markdown file from your repository: CLAUDE.md, AGENTS.md or GEMINI.md. Codex and Gemini CLI are open source. Claude Code is not.

You do not repeat your rules in every prompt. You write them once in a file. Claude Code reads CLAUDE.md. Codex reads AGENTS.md. Gemini CLI reads GEMINI.md. You put things like the build command, the test command, naming rules, and what the agent must never touch.

CLAUDE.mdClaude Code
AGENTS.mdCodex
GEMINI.mdGemini CLI
# Project rules Build: make build Test: pytest -q Never edit files under migrations/ Use snake_case for table names

On licence. Codex CLI and Gemini CLI are both published under the Apache 2 licence, so you can read the code and change it. Claude Code is a proprietary tool, and it works with Claude models only.

ClaudeChatGPTGemini
Rules fileCLAUDE.mdAGENTS.mdGEMINI.md
LicenceProprietaryApache 2.0Apache 2.0 Gemini CLI
Models it can useClaude onlyOpenAI, and compatible endpointsGemini
Licence and model support as reported by independent comparisons in 2026

4. How much of my code base can each agent keep in view?

The short answer

All three reach about 1 million tokens. So the size is not the difference. The difference is what a very long session costs, because OpenAI charges more above 272K tokens.

A million tokens is a lot of code. Google describes it as about 30 thousand lines. Claude Code's current models support up to 1 million tokens. Codex's GPT-6 models document 1.05 million, with about 922 thousand for input. Gemini 4 Argon has 1 million.

Claude Code
up to 1M
Codex
1.05M
Antigravity / Gemini
1M
Sources: Anthropic, OpenAI and Google model documentation, November 2026

But a full window is expensive and slow. Good agents do not load everything. They search, read only the files they need, and summarize the session when it gets long. So in practice, how the agent manages context matters more than the maximum.

Keep in mind

A bigger window is a safety net. Good agents read only what they need.

5. Which one writes better code, according to the benchmarks?

The short answer

On most published tests, Claude Fable 5.2 is first, GPT-6 Astra is second, and Gemini 4 Argon is third. Gemini leads on one test, DeepSWE, in Google's own table. Always check who published the number.

Four tests. Frontier S W E version 2, hard engineering tasks: Fable 5.2 scores 69.9, Astra 65.5, Argon 55. FrontierCode 2, which asks whether real maintainers would accept the change: Fable 49.1, Astra 42.8. Terminal-Bench 4, run independently by Vals: Fable 65.2, Astra 59.6. And DeepSWE, from Google's table: Argon 77.9, with Claude Opus 5.5 and Astra both near 74.

benchmarkClaudeChatGPTGemini
FrontierSWE v2 vendor tables69.9% Fable 5.265.5% Astra55.0% Argon
FrontierCode 2.0 Anthropic49.1% Fable 5.242.8% Astra—
Terminal-Bench 4.0 Vals, independent65.2% Fable 5.259.6% Astra—
DeepSWE v1.1 Google74.2% Opus 5.574.1% Astra77.9% Argon
Each company's table shows the tests where it does well. The Vals result is independent.

Two honest notes. The gaps are a few points, not a different league. And a benchmark is not your code base. The agent that follows your rules file and your tests best is the one that will feel better.

Keep in mind

The gaps are a few points. Try your own task: one bug fix and one small feature on each.

6. What does one coding session cost if I pay by the token?

The short answer

It depends on the model far more than on the tool. A typical session costs about $3 on GPT-6.1 Sol, about $5 on Claude Opus 5.5 or Gemini 4 Argon, about $10 on Fable 5.2, and about $15 on GPT-6 Astra.

Take a realistic session. The agent reads a lot and writes a little. Say 4 million input tokens, with 90 percent of them cached, and 150 thousand output tokens. On GPT-6.1 Sol, that is 2 dollars 66. On Claude Opus 5.5, 5 dollars 32. On Gemini 4 Argon at the December rate, the same 5 dollars 32. On Claude Fable 5.2, 9 dollars 92. On GPT-6 Astra, 15 dollars 10.

GPT-6.1 Sol
$2.66
Claude Opus 5.5
$5.32
Gemini 4 Argon (Dec 1 rate)
$5.32
Claude Fable 5.2
$9.92
GPT-6 Astra
$15.10
4M input tokens (90% cache hits) + 150K output tokens at standard rates. Simplified: no cache-write fees and no long-prompt rates.

Now compare that with a subscription. Ten sessions a week on Opus is more than 200 dollars a month by the token. That is why heavy users buy a plan instead: the 20 or 100 dollar plan is much cheaper than the same work on the API.

Keep in mind

Daily coding: a subscription is far cheaper than paying by the token. The API is for automation.

7. So which coding agent should I use?

The short answer

Pick Claude Code for the highest published scores and large refactors. Pick Codex if you already pay for ChatGPT or want an open source tool. Pick Antigravity if your team is on Google Cloud.

Here is how I would choose. Claude Code: the strongest published results, a mature rules system, and it is included from 20 dollars. Codex: it is already in your ChatGPT plan, it is open source, and GPT-6.1 Sol is cheap to run. Antigravity: the natural choice on Google Cloud, with a 1 million token window and Gemini's strength on long inputs.

Claude
Pick Claude Code if
You want the top published scores
You do big multi-file changes
You want an open source tool
ChatGPT
Pick Codex if
You already pay for ChatGPT
You want open source
You want low cost per session
Gemini
Pick Antigravity if
Your team is on Google Cloud
You read very large code bases
You need the free Gemini CLI sign-in

Many developers use two: one for writing the change and another for reviewing it. Whichever you choose, write a good rules file and keep your tests fast. That improves every agent. In the next video, we compare images, video and voice.

Keep in mind

The best upgrade for any agent: a clear rules file and fast tests.

Key points to remember

  • All three are agents: they read your repository, edit files, run commands and fix their own errors
  • Claude Code starts on Pro at $20. Codex is on every ChatGPT plan, limited on Free and Go. Antigravity is on Google AI Pro and Ultra
  • Rules live in CLAUDE.md, AGENTS.md and GEMINI.md. Codex and Gemini CLI are open source
  • On most published coding tests Claude Fable 5.2 is first. Gemini 4 Argon leads Google's DeepSWE test
  • One session costs about $3 to $15 on the API, depending on the model

Sources

These are the official pages the numbers came from, read on November 16, 2026. Benchmark scores are named in the post together with who published them.

Next in this series

Part 8: Images, video and voice

This is an independent comparison. Prices are US list prices before tax, read on November 16, 2026. Plans, limits and models change often. Benchmark scores are reported as published, and each one names who ran it.

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.