What Google released on 13 October
Gemini 4 Argon is Google's new flagship model. On 30 September it went only to selected security teams through the Fairwind Program, and since 13 October it has been generally available. Google positions it as a model for long, multi-step work: coding, agentic tasks, legal and financial analysis, and security. In the API the model ID is gemini-4-argon, with a combined input and output limit of up to 1M tokens (longer outputs use a special continuation protocol in the API, not the regular Gemini app).
Where you can find it:
- Gemini API on the paid tier, Google AI Studio, Android Studio and Firebase AI Logic.
- Gemini Enterprise (web app, mobile app and developer tools) on eligible subscriptions, and since 14 October also in Workflow Builder.
- Google AI Pro and Ultra in the Gemini app and AI Mode in Search; the Antigravity environment for individuals only on Ultra.
What it costs and when the price changes
Until 30 November 2026 an introductory price applies: USD 2 per million input tokens and USD 10 per million output tokens. Cached input tokens are 95% cheaper. From 1 December the rates double to USD 4 and USD 20. If you are testing the model in November, calculate costs at the December rates from the start, because a pilot priced at the promotional rate can easily understate the budget by half.
The second cost trap is reasoning. At a high thinking level the model can analyse a task for a long time before answering, and those tokens are billed too. In agentic work the bill depends more on the number of steps than on the price per token, so compare models on your own tasks and measure the cost of one completed task.
How to read the benchmark results
Google reports, among others, 77.9% on DeepSWE v1.1 (long-horizon coding) and 51.3% on Zapier's AutomationBench (end-to-end business tasks). These are vendor-reported results. From independent rankings: according to Vals AI (as of 12 October) Argon leads the Vals Index at 68.90%, while on the Artificial Analysis Intelligence Index it scores 53, below several competing models. On Terminal-Bench 4.0, according to figures cited by VentureBeat, it trails Claude Opus 5.5. The takeaway: it is a top-tier model without a clear lead, and the choice should rest on a trial with your own data.
Data, the EU region and GDPR
For companies in Poland the key information is in the Gemini Enterprise release notes: Argon runs in the global, us and eu multi-regions, and for eu Google states data residency and ML processing within that region. If your Gemini Enterprise deployment is set to the eu region, the new model does not force processing outside the Union. The API announcement only says the model is available "globally", so for API integrations check the location and contractual terms under which data is processed.
Note the difference between account types. Google AI Pro and Ultra are consumer subscriptions. If employees paste company documents into them, the data leaves the data processing agreement you have with Google under its business services. For customer personal data and confidential documents, use only company accounts with signed processing terms.
Under the AI Act, the obligations of a general-purpose model provider rest with Google. What stays with you is assessing what the model is used for (summarising emails carries different requirements than pre-screening job candidates) and ensuring staff have basic AI literacy, which has been required since February 2025.
What a Gemini Enterprise administrator should watch
- On 27 October the model is switched on by default in the Gemini Enterprise app. Until then the toggle is off. If you want to check costs and limits first, set this deliberately before that date.
- In developer tools the model is available immediately if the authorised models list uses the default selection. With a custom list, an administrator adds it manually.
- Since 22 October you can set a per-user limit so that one person cannot exhaust the shared quota pool of the whole edition. With a more expensive model this is a sensible safeguard.
- Google says Argon is its most resilient model against indirect prompt injection (0.7% attack success on the Gray Swan benchmark). That is still not zero: an agent that reads external emails, web pages and documents should have minimal permissions and require human confirmation before irreversible steps.
What is worth testing
A trial makes the most sense where tasks are long and multi-step: reviewing and refactoring older code, analysing long contracts or reports, automating processes across systems. For simple tasks, such as summarising an email or drafting a reply, cheaper models from the Flash family are usually enough. Prepare a few typical tasks from your company, run them through your current model and through Argon, and compare quality and cost per task. Such a test set will also help with the next launches, of which there is no shortage this year.