Google introduces Gemini 4 Argon AI model with 1 million token limit

The Argon flagship architecture

Alphabet's Google announced its Gemini 4 artificial intelligence generation on 30 September 2026, anchored by a new flagship model named Argon. The model is larger than Google's earlier Pro tier and is designed to handle complex, long-horizon tasks across software engineering, finance, legal analysis, and cybersecurity. In software engineering evaluations, Argon reached a score of 77.9% on the DeepSWE v1.1 benchmark, finishing ahead of OpenAI's GPT-6 Astra, Anthropic's Opus 5.5, and Fable 5.1. Google also reported industry-leading results on the Vals Index for economic analysis. To support larger computing operations in single prompts, Google increased the maximum output window to 1 million tokens, expanding from the 64,000-token ceiling present in earlier Gemini models. No API pricing has been announced during the initial testing phase.

Internal testing and codebase migration

Inside Google, technical teams deployed Argon on internal infrastructure prior to external testing. The model analyzed fleet-wide telemetry data to reduce memory usage across Google data centers by 300 TiB. Specialized Argon agents also converted legacy C and C++ codebases into Rust. This refactoring included thousands of lines in the core re2 and libgav1 software libraries, along with more than 800,000 lines within the Zircon kernel of the Fuchsia operating system. Google DeepMind product leads reported that internal researchers and software engineers relied on Argon for challenging technical workloads during recent evaluation cycles.

We've seen strong performance up close as Googlers have put the model through its paces in recent weeks, with many relying on it for their hardest coding and research problems.

Defensive cybersecurity and safety controls

Google is limiting initial external access to select cybersecurity partners through its Fairwind Program and is participating in the Trump administration's voluntary pre-release review process. Cybersecurity firm Wiz deployed Argon to identify a critical security vulnerability capable of exposing personal data in hospital software deployed globally, a flaw Google stated competing frontier models failed to catch. Following several hacking incidents and industry concerns over autonomous agent behavior during the summer of 2026, Google integrated chain-of-thought monitoring directly into Argon. This verification mechanism tracks the model's intermediate reasoning steps and halts execution if actions deviate from safety parameters.

Argon is a well-rounded model that has frontier capabilities across several domains.

Delays, internal skepticism, and market competition

The arrival of Gemini 4 ends an extended gap at the top of Google's model catalog, occurring nearly a year after the release of Gemini 3. In May 2026, Google chief executive Sundar Pichai stated that a new model would launch in June, but the anticipated Gemini 3.5 Pro was not released. Google instead released smaller Flash models throughout the summer. Internal workplace reports pointed to morale difficulties inside Google DeepMind and ongoing employee skepticism regarding the model's actual coding reliability relative to benchmark results. The launch occurs as competitors confront safety challenges; OpenAI cancelled its GPT-6.1 Astra model over security vulnerabilities, while both OpenAI and Anthropic investigate thousands of agent misbehavior reports. Google indicated that paying subscribers will receive access after early partner evaluations conclude.