Gemini 4 Argon fuehrt die Benchmarks an, Opus 5.5 punktet bei Terminal-Bench

Gemini 4 Argon is ranked first on the Vals Index, ahead of every Claude model on that leaderboard. On Terminal-Bench 4.0, the agentic terminal benchmark, Google's own comparison has Claude Opus 5.5 beating Argon 66.4% to 57.4%, as VentureBeat reports.

Google made Argon generally available on October 13, 2026 ( Google ). If you have Google AI Ultra, it's already in Antigravity CLI. If you pay for Claude Code, Opus 5.5 is one command away. That's enough to stop arguing about leaderboards and run a test on your own code: the same ticket, from the same commit, handed to both agents side by side.

Google's launch post makes Argon available to developers through the Gemini API paid tier in AI Studio and through the Gemini platform, where the model ID is gemini-4-argon ( Google Cloud ). Consumers get it in the Gemini app and AI Mode for Ultra and Pro subscribers ( Engadget ). Before October 13, it had been limited since September 30 to vetted cyber defenders in the Fairwind Program ( VentureBeat ).

The coding-agent route is narrower. Antigravity's models page lists Argon for Google AI Ultra and Enterprise plans only, not for Free, Plus or Pro. It also says Argon is "Not available in Antigravity IDE": you get it in Antigravity 2.0, the Antigravity CLI and the IDE extensions.

If you'd rather call it from your own code, the per-token rates are on the pricing page ($2 / $10 per million input / output tokens through November 30, then $4 / $20 from December 1). The API also gives you a 1M-token unified input-output limit, per Google.

Each number measures something different:

Then there are Google's internal numbers. Google says teams using Argon in Antigravity raised agentic code submissions by more than 35% while "reducing code rollbacks by nearly a third," and that shortcut attempts on coding benchmarks "dropped to just 2%" ( Google ). Those are Google's own figures about Google's own engineers, and VentureBeat notes they have not been independently replicated.

The same article quotes Sam Witteveen calling Argon "a really, really, really good model," while also saying, "I still don't think, for example, Antigravity is anywhere near as good as Claude Code or Codex" ( VentureBeat ). That is the model and the harness pulling in opposite directions. When you race Argon in Antigravity against Claude Code, you test both at once. That's fine, because model plus harness is what lands in your pull request.

Both worktrees start from the same commit, so the only variable is the agent.

The race is only fair if both agents start from the same files. Claude Code has a --worktree flag, but by default it branches from your repository's default branch on the remote, not from your local HEAD ( Claude Code docs ). Create both worktrees by hand from the same commit instead, the way the docs' "manage worktrees manually" section shows:

cd ~/code/app git switch main git pull git worktree add../race-argon -b race/argon git worktree add../race-claude -b race/claude Enter fullscreen mode Exit fullscreen mode Each git worktree is its own checkout on its own branch, so the two agents never write to each other's files.

If you don't have agy yet, this is Google's installer for macOS and Linux ( Antigravity docs ). Read the script before you pipe it into a shell:

curl -fsSL https://antigravity.google/cli/install.sh | bash Enter fullscreen mode Exit fullscreen mode Write the ticket once, in a file both agents will get word for word:

Fix <the bug or feature, with the issue link>. Only touch files under src/ and tests/. Run the test suite before you stop. Do not commit. End with a short summary of what you changed and why. Enter fullscreen mode Exit fullscreen mode "Do not commit" keeps the result in each worktree's working tree, which makes the comparison below one command.

In the first terminal, list the models your account can use, then start Antigravity on Argon. A third-party cheat sheet tested on CLI 1.2.4 documents agy models and --model <slug>. Google's getting-started page doesn't list these flags, so check agy --help on your build and copy Argon's slug exactly as agy models prints it:

cd ~/code/race-argon agy models agy --model "<Argon slug from agy models>" Enter fullscreen mode Exit fullscreen mode In the second terminal, start Claude Code and pick your usual model with /model ( Claude Code docs ):

cd ~/code/race-claude claude Enter fullscreen mode Exit fullscreen mode Paste the same ticket into both. Match reasoning effort as closely as the two tools allow, and don't help one agent more than the other.

If you want to separate the model from the harness, add a third lane. Antigravity's models page lists Claude Opus 5.5 (Thinking) among its models. If agy models shows it for your plan, run it in a third worktree. That gives you Argon versus Opus inside the same harness, and Opus inside two different harnesses.

When both agents stop, compare what they left behind from the main checkout:

cd ~/code/app for w in race-argon race-claude; do echo "== $w " git -C "../ $w " status --short git -C "../ $w " add -N. git -C "../ $w " diff --stat done Enter fullscreen mode Exit fullscreen mode add -N marks new files as intent-to-add without staging their content, so files an agent created show up in diff --stat too. If you ran the third lane, add race-opus to the loop.

Then run your own test command inside each worktree. Score each run on things you can check:

Four checks per run: tests, diff size, interruptions, and whether the summary matches the diff.

Before you pick a winner, run three tickets that look like your real backlog: a bug fix, a small feature and a refactor. Keep the branch you trust, and when you're done, delete the race worktrees with git worktree remove ( git docs ).

I haven't run Argon myself, so I can't tell you who wins.

I build Hermes IDE, so weigh this section with that in mind. Hermes runs Antigravity CLI as a built-in agent next to Claude Code and Codex, each in its own real terminal ( README ). Its Antigravity model picker reads the list live from agy models, so once Antigravity offers Argon to your Ultra account, Argon appears in Hermes with no Hermes update. The effort picker for Antigravity offers low, medium, high and max.

Press ⌘N (Ctrl+Shift+N on Windows and Linux) and each task gets its own hermes/<name> branch and worktree ( README ). You can start the same task in Antigravity and in Claude Code side by side on one repo. Since 2.2.4, launching a task on two agents at once respects your running-agents limit. The Review Desk (⌘G) shows each agent's changes by turn or by file, and the attention inbox tells you which of the two is waiting on you.

Hermes doesn't show an estimated cost for Argon yet, because it doesn't price that model. Check usage in your Google account. Agents sign in with their own CLI logins, and Hermes never asks for your tokens. It's free to download, source-available under BSL 1.1, and runs on macOS, Windows and Linux. Hermes IDE is not affiliated with Google, Anthropic or Nous Research.

Which one would you bet on for your repo's next bug fix?

Download Hermes IDE, give the same ticket to Antigravity and Claude Code in two worktrees, and keep the branch that earns it. If the side-by-side setup is useful, star it on GitHub. If you build agents yourself, the AI Agents Pocket Guide covers the loop, tool and guardrail patterns behind tools like these.