Imagine hiring a contractor who reads every blueprint in the building before touching a wall, and then writes you a long, detailed report on everything they did. The work is good. The report is thorough. But if you pay by the word, you notice the bill.
Gemini 4 Argon is that contractor. It is Google’s first Gemini 4 model and, for the first time in a while, a Google model that competes with the best coding agents. It is also one of the wordiest.
What it is, and where you’ll find it
Google announced Gemini 4 Argon on September 30, 2026, and opened it widely on October 13, 2026. For coding, there are two ways in.
Through the Gemini API. Any developer on the paid tier can call gemini-4-argon from AI Studio, the Gemini API, or tools that let you plug in your own key. You choose a low, medium (default), or high thinking level. A higher level lets the model reason longer before it answers, which takes more time and more tokens.
Through Antigravity. Antigravity is Google’s coding agent, available as a desktop app and a command-line tool (CLI). This is where the access story gets awkward: Argon in Antigravity requires Google AI Ultra, Google’s top consumer plan, or an eligible Gemini Enterprise plan. Google AI Pro subscribers get Argon in the Gemini app, but not in Antigravity. It isn’t offered in the older Antigravity IDE at all. A launch-day thread on Google’s AI Developers Forum shows users asking about exactly this.
How good is it at coding?
Artificial Analysis, an independent testing firm, measures coding agents rather than bare models, because a model’s results depend heavily on the agent software around it. Its Coding Agent Index averages three tests: DeepSWE v1.1 (complex tasks in real codebases), Terminal-Bench 4.0 (long jobs at the command line), and SWE-Atlas-QnA (answering questions about large codebases).
On that index, Gemini 4 Argon in the Antigravity CLI scores 64. Claude Sonnet 5.5 in Claude Code leads at 68. GPT-6.1 Sol in Codex scores 63, one point behind Argon. In other words, Google’s agent now trades blows with OpenAI’s value champion.
Google’s own launch table is more flattering. On DeepSWE v1.1, it reports 77.9% for Argon, against 74.1% for GPT-6 Astra, 74.2% for Claude Opus 5.5, and 67.4% for Claude Fable 5.1. That is the top score in Google’s comparison. Treat it with some care: no third party had reproduced the table at launch, and reviewers noticed that two of Astra’s numbers in it were higher than OpenAI’s own charts showed.
On Artificial Analysis’s broader Intelligence Index, Argon scores 53, tied with GPT-6 Astra.
Where it shines
- Big repositories. One million tokens of context is enough to load a large codebase, its documentation, and its test logs together. That space is shared by input, thinking, and output, so very long sessions need some planning.
- Very long outputs. Most models stop after tens of thousands of tokens of output. Argon can keep writing up to about a million tokens, continuing across responses, which helps with big migrations, generated code, and long technical documents.
- Cautious with facts, at least on a knowledge test. Artificial Analysis measured the lowest hallucination rate of any leading model on its AA-Omniscience knowledge test: of the questions Argon did not answer correctly, only 15% got a confident wrong answer, against 51% for GPT-6 Astra. Two caveats: the test covers general facts, not code, and Argon also answered fewer questions correctly than Astra (50% versus 63%). It may also invent fewer library options or functions, the bugs that take longest to track down, but nobody has measured that yet.
- Reading screenshots and recordings. It accepts images, audio, and video along with text, so you can give it a screenshot of a broken layout or a screen recording of a bug.
The honest catch
Long terminal sessions are still a weak spot. On Terminal-Bench 4.0, where an agent works for a long time in a real command line, Artificial Analysis measured about 57%. Claude Sonnet 5.5 scores 64%, and Claude Opus 5.5 and GPT-6 Astra about 59.6%. Google’s own table also shows Opus 5.5 ahead on this test (66.4% to 57.4%).
It burns tokens. Artificial Analysis needed 110 million tokens to run its Intelligence Index with Argon, against a median of about 85 million. On that index, the average cost per task was about $1.99 at the launch price, versus $0.72 for GPT-6.1 Sol. On coding work the gap is wider: Artificial Analysis’s Coding Agent Index, checked October 14, 2026, puts Argon in Antigravity CLI at $5.84 per task and GPT-6.1 Sol (xhigh) in Codex at $1.04. These are average API costs per benchmark task, not subscription charges. A low price per token does not make it cheap per task.
The launch price ends November 30. The API costs $2 / $10 per million input/output tokens until then, and $4 / $20 after. Cached input goes from $0.10 to $0.20.
Front-end work needs a guide. A hands-on review by Stork.ai found its designs cramped next to rivals’, though in the same review it kept working through a long overnight agent task. It is better at the logic than at the look.
Who should use it
| If your project looks like… | Reach for | Why |
|---|---|---|
| A huge repository, long generated outputs, Google Cloud stack | Gemini 4 Argon | 1M context, very long output, cautious on an independent knowledge test |
| Long autonomous terminal jobs | GPT-6 Astra or Claude Opus 5.5 | Ahead on Terminal-Bench 4.0 |
| The highest coding-agent score | Claude Sonnet 5.5 in Claude Code | Leads Artificial Analysis’s Coding Agent Index |
| Lots of tasks on a tight budget | GPT-6.1 Sol | Similar Coding Agent Index score at about a sixth of the API cost per task |
The coding verdict
Gemini 4 Argon puts Google back in the coding conversation. It is independently strong, has a context and output length few rivals match, and, on an independent knowledge test, is unusually careful about facts. If your team already pays for Google AI Ultra or builds on Google Cloud, it deserves a real trial. If you are on Pro, you’ll need to use it through the API instead of Antigravity. Either way, measure the cost per task, not the price per token, and budget for December.
