Gemini 4 Argon: What Google's New Model Can Do, Cost and Access

01 / The story

Is Gemini 4 Argon real?

Yes. Before writing, we checked official Google sources. Google announced Gemini 4 Argon on September 30, 2026, rolling it out first to trusted cyber defenders through its Fairwind Program. [1] On October 13, 2026 Google said it was globally available to developers, enterprises and consumers. [2]

Google DeepMind's model card describes it as a new frontier model, not a fine-tune of an earlier one, built as a sparse mixture-of-experts transformer with native multimodal input. [3]

02 / The story

Who can use it and where?

According to Google, Argon is available through the paid tier of the Gemini API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Google Antigravity, and to Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. [2] [4]

The model ID for developers is gemini-4-argon. [5] Availability can still differ by country, account type and workspace settings, so check your own access.

03 / The story

What can it actually do?

Google positions Argon for long, multi-step work: large code refactoring, enterprise tasks in legal and finance, cybersecurity defense and scientific problems. [2] Those are use cases Google promotes, not independent proof of quality in your field.

The API documentation lists these specifications, which are more useful for planning than marketing language. [5]

04 / The story

How much does Gemini 4 Argon cost?

Google's developer guide lists introductory API pricing through November 30, 2026 of $2 per million input tokens and $10 per million output tokens, with cached input 95% off. From December 1 it lists $4 input and $20 output per million tokens. [4]

Simple example: a job with 200,000 input tokens and 10,000 output tokens costs 0.2 × $2 + 0.01 × $10 = $0.50 at introductory rates, and $1.00 at the December rates. Real bills also depend on thinking tokens, retries and tool use. Consumer access comes through Google AI Pro or Ultra subscriptions instead of per-token billing.

For a broader view of how token bills grow, see our guide to surprise AI API bills.

05 / The story

What do the benchmarks claim, and what do they not show?

Google says Argon ranked first on the Vals Index, which measures performance across finance, coding, legal and tax tasks, and that internal teams increased agentic code submissions by more than 35%. [2] These are Google's claims; Qlauson has not reproduced them.

A benchmark score reflects a specific test set, settings and date. It does not prove that one model is best for every task, and rankings change as rivals update. Test Argon on your own examples before switching. Our guide to evaluating AI models shows how.

06 / The story

How does it compare with alternatives?

Compare on what matters to your job: accuracy on your tasks, context size, tool support, price per finished task, data terms and where it is available. A larger context window helps with long documents and codebases, but stuffing everything into one prompt can still miss details.

For a plain comparison of popular assistants, see Claude vs ChatGPT in 2026 and our earlier Claude, ChatGPT and Gemini comparison. For why large models need so much hardware, see why AI needs computing power.

07 / The story

Limitations and safety notes

Google's model card includes sections on known limitations and safety evaluations; read it before using Argon for high-stakes work. [3] Like other large models, it can produce confident but wrong answers, so legal, financial and security outputs need expert review.

Computer use is still a preview feature. Give any agent built on Argon narrow permissions and require approval before it changes accounts, code in production or money.