Gemini 4 Argon: Google's new model can write 1 million tokens at a time

What's new

Google launched Gemini 4 Argon on October 13, 2026 (some sites list October 14). It is the first model in the Gemini 4 generation. Before the launch, Google had spoken publicly about Gemini 4 only twice: in a blog post in late July, where it wrote that it was in the middle of its most ambitious pre-training run to date (the large-scale base training of the model), and on an earnings call for the second quarter.

Argon is available in the Gemini app for Google AI Pro and Ultra subscribers. Developers can use it through the Gemini API, Google AI Studio, Google Antigravity and Google's enterprise platform. The sources do not say whether the model is available in Denmark and the EU, and the price of the Pro and Ultra subscriptions has not been disclosed.

What's smart

According to Google, Argon answers more directly with fewer caveats, repeats itself less and avoids clichés. The model can also automate tasks across more than 40 connected apps, including Instacart, OpenTable, Adobe and Shopify. The app gets a new chat-to-tasks feature, and Google Antigravity for macOS gets an agentic coding interface where the AI can work its way through a coding task on its own.

The model can deliver up to 1 million output tokens (chunks of text) in a single response, up from the previous 64,000. Google also describes built-in safety: defenses against harmful requests, resistance to prompt injection (attempts to trick the AI with hidden instructions) and sandboxed test environments.

Google says internal teams have used Argon for quantum optimization (a 40 percent improvement over the baseline), to free up more than 300 TB of memory in data centers and to migrate more than 800,000 lines of C++/Rust code.

Cheaper or better?

In the API, Argon costs $2 per million input tokens and $10 per million output tokens during an introductory period. The period ends around November 30 or December 1, 2026, after which the price rises to $4 and $20. According to HackerNoon, the introductory price is about half that of Claude Opus 5.5 ($4 and $20) and about one fifth that of GPT-6 Astra ($10 and $50). Afterward, the price matches Claude Opus 5.5.

On performance, Argon leads on 14 of 19 benchmarks according to one review (other sites say 13), including 77.9 percent on DeepSWE v1.1. Google's own figures include 57.1 percent on Terminal-Bench 4.0, 69 percent on Agents' Last Exam v1.1 and an 84.2 percent F1 score on GraphWalks with 256k to 1M context. Keep in mind that the figures come from Google itself and have not been independently verified. DataCamp's independent tests showed mixed results: Argon led on some measures, while Claude Opus 5.5 performed better on Terminal-Bench 4.0 and terminal-heavy coding tasks.

What it's good for

Google highlights complex tasks in coding, knowledge work, cybersecurity and science. On cybersecurity, Google reports 68 percent on CWE-bench and 42 percent on ExploitGym. If you run a company with a large codebase, the long outputs and the coding agent could be of interest. Everyday users may benefit from the app's ability to handle small tasks across apps, such as ordering and shopping.

Researchers can get free top-tier access in Google Antigravity for six months, and cybersecurity professionals can sign up for Google's Fairwind program. The sources disagree on how broadly the API is open: some say access is broad, while another describes it as limited to the Fairwind program, with broader access later and no date given.