Google presents Gemini 4 Argon – the company's most powerful model

Google is back in the lead: on 30 September, the parent company Alphabet presented the new flagship model Gemini 4 Argon – after months of delays and a phase in which Google had fallen behind in the AI race. The last "real" frontier model, Gemini 3.1 Pro, still dates from February 2026.

What Argon was built for

Google positions Argon as an all-rounder for coding, research and writing – with a special focus on defensive cybersecurity. According to Google, the model is supposed to "autonomously find, validate and patch critical software vulnerabilities". Google teams already use it internally for debugging and code migrations: among other things, 800,000 lines of the Zircon kernel (operating system Fuchsia) were migrated from C/C++ to Rust, and a data-center optimization is said to have freed up a total of 300 TiB of RAM.

How does it perform in benchmarks?

• Top placement in the emerging Vals Index – its creators highlight strong performance on financial questions, law, coding and security.
• The independent comparison by Artificial Analysis sees Gemini 4 on a par with OpenAI's GPT-6 Astra; only Anthropic's Claude Opus 5.5 performs better there.
• The lowest hallucination rate among the current top models: Argon would rather admit not knowing something than invent things.
• Best result on the Gray-Swan-Benchmark for Indirect Prompt Injection – i.e. protection against attacks via foreign content.

Availability: initially tightly limited

Is this mainly just an announcement for now: so far Google is only rolling out Argon to selected cyber partners via the "Fairwind" program, and the US government is allowed to examine the model in advance. Notably, prices are reportedly significantly below those of the direct competitors.

Assessment

The launch comes at a delicate time: OpenAI has just released Astra and paused the training of new models, while governments around the world are arguing about safety rules. For Google, Argon is an attempt to make up for the last months of delays – the Gemini app has counted more than one billion users since August. Whether the internal sceptics are right (Bloomberg reports on doubts about the coding performance) will become clear in the coming weeks.

Sources: TechCrunch, 30.09.2026 · Reuters, 30.09.2026 · derStandard.at, 01.10.2026