Gemini 4 Argon: what Google launched, what it costs and who can use it in the Gemini app

Gemini 4 Argon is Google’s newest frontier model, and it is now in the Gemini app. Google announced it on September 30, 2026, first for a small group of trusted cyber defenders, and made it globally available on October 13. This report covers what Google says Argon does, what it costs, who can use it and what changed in the Gemini app, using Google’s own launch posts and model documentation. The benchmark and impact claims are Google’s.

Screenshot: the Gemini web app, signed out, captured by Universal Digital Assistant on October 17, 2026. Gemini 4 Argon itself needs a Google AI Pro or Ultra subscription in the app.

Gemini 4 Argon at a glance

What Google says it can do

Google describes Argon as built for “deep work in complex, real-world tasks” across coding, finance, legal, cybersecurity and science, and says it ranks #1 on the Vals Index, which measures performance across sectors such as finance, coding, legal and tax work. It also says Argon communicates more clearly, follows style instructions better than earlier Gemini models, answers more directly with less hedging, and avoids repeating itself or using clichés. Examples from Google’s post:

• Coding at Google. Argon in Antigravity helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. It helped the Chrome team separate core webpage handling from a roughly 6 million line legacy folder, finishing “months of work in weeks”.
• Performance engineering. On Google’s open-source Agent Development Kit, working overnight, Argon found optimisations that cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%.
• Science. Google says Argon made the xTB quantum chemistry package up to 2.3x faster at predicting molecular structures, and sped up its open Gemma 4 E2B model 2.85x on one GPU setup with no quality loss.
• Robotics. It scored 87% on the Embodied Reasoning (ERQA) benchmark, and Google describes a video-to-simulation pipeline in which a humanoid sank seven consecutive cornhole shots.
• Security. Argon is being expanded to hundreds of additional Fairwind Program partners. Wiz, for example, used it to find and fix a vulnerability that exposed a master key.

Google also quotes early customers such as Box (a 14% accuracy improvement over Gemini 3.1 Pro) and Kepler (matching analyst reference figures on more than 99% of numbers in its financial tests). These are vendor-reported results.

What changes in the Gemini app

Google’s Gemini app post says Argon delivers frontier reasoning and “improved performance across complex tasks”, and that for longer tasks it shares progress more often. Alongside it Google made three app changes:

1. Chat or tasks. A simple toggle moves between chat and tasks. When you choose tasks, Gemini can work for you even when you’re not at your computer. See What is Gemini Spark? for Google’s agent.
2. More than 40 connected apps, including Instacart, OpenTable, Adobe and Shopify, for jobs like ordering groceries and booking tables.
3. Antigravity in the macOS app. Google Antigravity’s agentic coding interface is now available directly in the Gemini macOS app.

Gemini 4 Argon is available to Google AI Pro and Ultra subscribers in the app. Our Google AI Plus vs Pro vs Ultra guide shows the plans, and skills, which are replacing Gems, are covered in Gemini skills replace Gems.

Safety

Google calls Argon its most comprehensive safety and security stack yet. It says it trained Argon to respect system boundaries, report negative findings honestly and pause for human confirmation before irreversible steps, and that shortcut attempts on coding benchmarks dropped to 2%. It reports that external bio red teaming could unlock less than 1% of the bio sub-tasks it tests, and that violations on an adversarial set fell nearly two-fold compared with earlier Gemini models.

How it compares

Argon arrives in the same month as Claude Sonnet 5.5 and Claude Haiku 5.5 from Anthropic and GPT-6.1 Sol from OpenAI. Its introductory API price of $2 and $10 matches Sonnet 5.5 and GPT-6.1 Sol, though Haiku 5.5 is far cheaper, and the labs publish different benchmarks, so test it on your own tasks. For the assistant overall, see the Gemini profile and Claude vs Gemini.