Gemini 4 Argon: Google opens its new flagship model to developers and businesses

On October 13, 2026, Google opened Gemini 4 Argon, its most advanced AI model, to developers and businesses. The model was unveiled on September 30, but at that point it was limited to members of Fairwind, a Google program for governments and cybersecurity partners. It is the first model in the Gemini 4 family. Google promises the others without giving a date.

Who has access, and at what price

According to Google's blog post, Argon is available worldwide through the API, in AI Studio and on the Gemini platform for enterprises. Google AI Pro and Ultra subscribers can find it in the Gemini app and in the AI mode of search. The Antigravity development environment is open to businesses and Ultra subscribers. The post does not detail availability country by country, including in the European Union.

The launch price is $2 per million input tokens and $10 per million output tokens. It will rise to $4 and $20 on November 30, 2026. Cached input tokens get a 95% discount. Google states a limit of one million tokens, input and output combined.

Performance mostly measured by Google

Google presents Argon as a deep-work model, able to handle long tasks in programming, finance, law or cybersecurity. Most of the published scores come from Google or its partners, sometimes on private datasets. This is the case for the 87% claimed on the ERQA reasoning test, for which the post specifies neither the protocol nor the competitors.

One outside measurement exists. According to Engadget, the independent firm Artificial Analysis estimates that Argon matches OpenAI's GPT-6 Astra on its intelligence index, for about 60% of the cost per task at current promotional prices. It also gives Argon the lowest hallucination rate among major models, at 15%. These cost comparisons depend on the launch price, which will double.

The cybersecurity side

Google is extending Fairwind to several hundred more partners. Cybersecurity defenders will be able to request access to a version of Argon without the safeguards that limit offensive uses. According to Google, on Gray Swan's indirect prompt injection test, the attack success rate drops to 0.7% over 15 attempts. This figure is also published by the vendor.

Key takeaways

• Before adopting Argon, test it on your own use cases: published rankings are no substitute for an internal evaluation.
• Budget at the final price, not the launch price.
• Check the data processing terms and the hosting region before any use involving personal data.