Google Released Gemini 4 Argon with Record-Breaking Specs

Google has introduced the Gemini 4 Argon model, designed for complex reasoning, code writing, and agentic tasks. The company has already integrated it into its internal processes.

In the DeepSWE v1.1 benchmark, which evaluates models' ability to solve tasks in real-world codebases, Gemini 4 Argon scored 77.9%. This result set a new record.

The new architecture supports an output limit of up to one million tokens, increased from the previous 64,000. The model demonstrated leadership in Zapier's AutomationBench test with a score of 51.3%.

Model metrics include 91.7% on LVBench for long video understanding, as well as leading positions in financial and legal agent tests without the publication of exact figures.

The model can independently find, verify, and patch security vulnerabilities. Using it, Wiz identified a critical vulnerability in medical software overlooked by other systems.

Access to Gemini 4 Argon is initially provided to cybersecurity specialists without safety restrictions. AI Ultra subscribers and API customers will receive access later.

At launch, the cost is $2 per million input tokens and $10 per million output tokens. After the promotional period, rates will double.

Media claims about Argon's superiority over GPT-6 Astra are journalists' interpretations. Google does not conduct direct comparisons of its models with competitors' products.