How does Gemini 4 Argon do at coding?

Gemini 4 Argon is Google's flagship reasoning model, generally available since October 13, 2026. Google aims it at deep, multi-step work, with software engineering at the top of the list, and offers it in three places that matter to programmers: the Gemini API (through Google AI Studio, on the paid tier only), Google Antigravity, Google's agentic coding tool, for Google AI Ultra subscribers, and the Gemini app for Pro and Ultra subscribers.

How good is it at code? Google's own stories are striking, its own public test is more modest, and the honest answer sits between them.

What Google reports

Google's launch post describes Argon on large, real projects rather than on test scores. It helped the Chrome team pull core webpage handling out of a legacy folder of roughly 6 million lines, finishing a refactor across hundreds of files in weeks rather than months. Working overnight on Google's open-source Agent Development Kit, it found changes that cut session preparation time by 40% in long sessions. Google says that inside Antigravity, teams using Argon submitted over 35% more agent-written code while rolling back nearly a third less.

Two safety points matter for anyone who lets an agent run unattended. Google trained Argon to stop and ask before irreversible steps, and reports that attempts to take shortcuts on coding tests, such as fiddling with a failing test instead of fixing the code, fell to 2%.

What Google's own benchmark shows

Google also runs Android Bench, a public leaderboard of long Android development tasks, where each model works inside its usual coding agent. As of October 26, the top of the board reads:

It is unusual for a company to publish a test on which its own flagship comes sixth, and that makes the board worth trusting. Keep the ranges in mind, though: with only 30 tasks, the likely ranges overlap heavily, so Argon is not clearly worse than the models just above it. On cost per run it sits in the middle, cheaper than the Claude models and more expensive than GPT-6.1 Sol.

A clear prompt for that test:

Read this repository's README and tests first. The test test_totals_with_discount fails. Find the cause, fix it with the smallest change you can, run the whole test suite, and list each file you changed with one line on why. Stop and ask me before deleting any file.

What changes when its introductory price ends?

Argon's API prices are introductory until November 30, 2026. From December 1 they double:

A worked example: an app that sends 20 million input tokens and receives 2 million output tokens a month pays $40 + $20 = $60 today and $80 + $40 = $120 from December. Remember that Argon bills its thinking tokens, the reasoning it writes before answering, as output, so a task that makes it think hard costs more than the visible answer suggests.

Four ways to keep the bill in check:

• Batch or Flex for work that can wait: these modes cost half the standard price, before and after the change.
• Caching. Argon caches repeated context automatically, and cached input stays cheap compared with fresh input.
• A smaller model for routine jobs. Gemini 3.8 Flash costs $0.75 input and $3.75 output per million tokens until December 31, 2026, and has a free tier, though free-tier content may be used to improve Google's products.
• Re-run your comparison. At $4 and $20, Argon costs the same per token as Claude Opus 5.5 and twice as much as GPT-6.1 Sol. Test it against them on your own tasks before December.

The price change on Google's pricing page concerns the API. If you use Argon through an app subscription, check your plan's own terms rather than assuming the same rise.