Codex vs Google Antigravity: Which Agent Should You Use?

September 2026 · Published by Amar Kumar

Most coverage of AI coding tools stops at Claude Code, Cursor, and Copilot. That misses the two most interesting movements of 2026.

Codex went from 3% to 16% adoption in seven months — the fastest growth of any tool in JetBrains’ survey of 15,000+ professional developers. And Google Antigravity, at 6% globally, is tied third in India at 15%, the largest regional skew in the entire dataset.

Neither is beating Claude Code. Both are worth understanding, because they win on axes that have nothing to do with model quality: distribution and geography.

The numbers

All adoption figures here come from the JetBrains developer survey conducted May–July 2026, covering more than 15,000 professional developers. It is the tenth annual round, so the year-on-year comparisons are methodologically consistent.

Codex’s awareness figure moved even faster than its usage: from 27% to 65% in the same period. Two-thirds of professional developers now know it exists. That gap between awareness (65%) and adoption (16%) is where its next year of growth will come from.

What Codex actually is

Codex is OpenAI’s coding agent. The name is recycled — the 2021 Codex was the model behind the first GitHub Copilot — but the 2026 product is something different: an agent you run from the terminal or delegate to in the cloud.

The shape is familiar if you have used Claude Code. You describe a task in natural language. It reads the repository, edits files across the tree, runs commands, reads the output, and iterates until the task is done or it gets stuck.

• CLI — runs locally against your checkout, with your shell and your tests.
• Cloud tasks — hand off a job to a sandboxed container and collect a diff later, useful for long refactors you do not want to babysit.
• IDE integration — a side panel for the parts of the job you want to watch.
• AGENTS.md — a repo-root Markdown file describing build, test, and lint commands, which is how you stop the agent guessing.

The practical difference from Claude Code is smaller than the marketing suggests. Both are terminal-first agents with a cloud mode and a project-instructions file. If you have a working Claude Code setup, you already know how to drive Codex.

Why Codex grew 5x: distribution, not quality

It is tempting to read a 5× adoption jump as a verdict on model quality. The data does not support that reading.

On satisfaction, Claude Code remains well ahead — in JetBrains’ April 2026 round it was the most-loved tool at 46%, against Cursor at 19% and Copilot at 9%. And on conversion, 31% of all developers name Claude Code their primary tool, an 80% rate from regular use to main tool. Codex has not shown numbers like that.

What Codex has is distribution. It is bundled into the ChatGPT subscription that a large share of developers already pay for personally. The cost of trying it is zero marginal dollars and about ninety seconds. Compare that with evaluating a new tool that needs a separate purchase order.

This is the oldest pattern in developer tooling. The default that is already installed beats the better tool that requires a decision. Copilot won 2022–2024 on exactly this basis — it was bundled with GitHub, where the code already lived.

The strategic read: if your organisation has an OpenAI contract, Codex will show up in your codebase whether or not you evaluate it. Plan for that rather than discovering it in a code review.

What Google Antigravity is

Antigravity is Google’s agent-first IDE — an editor built around delegating work to an agent rather than an editor with an agent bolted on. It runs Gemini models and integrates tightly with Google Cloud.

Globally it is a minor player at 6%, and unlike Codex it did not grow much over the survey period. On the global numbers alone it would not be worth a section.

The reason it is worth a section is India.

The India anomaly

Antigravity is tied third in India at 15%, two and a half times its global share. No other tool in the survey shows a regional skew that large in a market that size.

Three things plausibly drive it, and they compound:

• Google’s India investment. At Google I/O Connect India 2026 the company announced AI education programmes, startup support, expanded enterprise infrastructure, and developer initiatives specifically for the Indian market. Antigravity rides that distribution.
• Gemini’s consumer footprint. Gemini was Google’s top trending search term globally in 2025, and gemini.google.com draws roughly 25.8 million monthly visits from India. Familiarity with the consumer product lowers the barrier to the developer one.
• Price sensitivity. Indian developers and startups are more likely to weigh per-seat cost heavily, and Google has been aggressive on free and low-cost tiers.

The same survey shows the mirror image elsewhere: Cursor’s usage in China halved, from 28% to 16%. Tool adoption is not a single global market. It is a set of regional markets with different defaults, and if you hire or sell across borders, the tool your team reaches for in Bangalore may not be the one they reach for in Berlin.

For anyone publishing developer content, there is a second-order effect worth noting. In our own search data, Indian traffic converts impressions to clicks far better than US traffic. A tool with 15% share in that market and almost no English-language documentation is an unusually open content opportunity.

Side by side with Claude Code

How to choose

For most teams the decision is made by a subscription you already hold, not by a benchmark. That is not cynicism; it is the correct answer, because the marginal quality difference between these agents is smaller than the cost of running a procurement process.

• Already paying for Claude — use Claude Code. It leads on adoption, satisfaction, and the depth of community tooling around it.
• Already paying for ChatGPT across the team — start with Codex. It is included, and the gap is not large enough to justify a second bill.
• Deep in Google Cloud, or a team in India — evaluate Antigravity seriously. The Gemini integration and local support are real advantages, and your engineers may already be using it.
• Standardising across a large org — do not pick one. Roughly 70% of engineers already run two to four tools. Write the project-instructions files for each and let people choose their editor.

The thing actually worth standardising is not the agent. It is the guardrails: what the agent is allowed to run, what gets reviewed before merge, and how the repository describes itself.

AGENTS.md and CLAUDE.md

Every one of these tools reads a project-instructions file, and almost every team leaves it empty. It is the single cheapest improvement you can make to agent output, and it takes ten minutes.

An agent with no instructions guesses your test command, guesses your formatter, and runs the wrong thing. An agent with five accurate lines does not.

# AGENTS.md (Codex) — the same content goes in CLAUDE.md for Claude Code

## Commands
- Install: `uv sync`
- Test: `pytest -q`
- Lint: `ruff check . && ruff format --check .`
- Type check: `mypy src/`

## Conventions
- Python 3.12, type hints required on public functions
- No new runtime dependencies without asking
- Tests live next to the module as `test_*.py`

## Do not touch
- `migrations/` — generated, edit the model instead
- `infra/prod/` — requires a separate approval

Two rules make these files work. Keep them short, because every line competes for context with your actual code. And keep them true — a stale test command is worse than no file at all, because the agent will trust it and burn a run discovering it is wrong.

If you maintain both files, make one the source of truth and symlink or generate the other. Two drifting copies of your build instructions is a bug waiting to happen.

FAQ

Is Codex better than Claude Code?

Neither wins outright. Claude Code leads on adoption (39% vs 16%) and satisfaction, and 31% of developers name it their primary tool. Codex wins on cost when your team already pays for ChatGPT. For most teams the existing subscription decides it.

What is Google Antigravity?

Google’s agent-first IDE, running Gemini models with tight Google Cloud integration. It sits at 6% adoption globally but 15% in India, where it is tied third among all AI coding tools.

Why did Codex grow so fast in 2026?

Distribution. It is bundled into the ChatGPT subscription many developers already hold, so trying it costs nothing extra. Adoption went from 3% to 16% between January and July 2026, and awareness from 27% to 65%.

What is AGENTS.md?

A Markdown file at your repository root telling a coding agent how to build, test, and lint the project. Codex reads it the way Claude Code reads CLAUDE.md. Keep it short and keep it accurate.

Should I switch tools because of these numbers?

Only if you are paying for a subscription you do not use. Adoption share measures distribution and momentum, not whether a tool suits your codebase. The switching cost is real and the quality gap between the top agents is not large.

The agent race is not being won on benchmarks. It is being won on which subscription a developer already pays for, and which company showed up in their country. Codex and Antigravity are both proof of that, in opposite directions.