Google launched its new AI model, Gemini 4 Argon, on September 30, positioning the software to tackle complex programming tasks, cybersecurity challenges, and corporate data analysis. A company spokesperson noted that Argon is larger than the previous Pro series of advanced models. The release marks a major reorganization of the Google DeepMind lab, arriving after months of delays that saw the company fall behind competitors like Anthropic and OpenAI.
Gemini 4 Argon undercuts competing models in pricing
Gemini 4 Argon introduces a significant expansion in output capabilities, jumping from previous 64 thousand token outputs straight to 1 million tokens. The initial pricing structure is set at $2 dollars per million input tokens and $10 dollars per million output tokens. This undercuts competing top models like Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra, which are priced at $10 dollars for inputs and $50 dollars for outputs.
The model scored 77.9 percent in the DeepSWE v1.1 benchmark, which simulates software engineering tasks, and claimed the top spot in the Vals Index covering finance, programming, law, and tax consulting. The Vals Index weights individual sectors based on their share of the United States GDP. Google also recorded 51.3 percent in Zapier’s AutomationBench and 91.7 percent in LVBench for long video comprehension. However, Google’s press release noted that Argon still lagged behind competitors in two out of four tests focused on programming.
Google restricts Gemini 4 Argon availability for testing
Google is initially restricting public availability for Gemini 4 Argon. The company is distributing the model through its Fairwind program to select cybersecurity partners, including the security firm Wiz, for vulnerability testing and software defense. Google stated that the model is participating in a voluntary process of the administration of President Donald Trump.
A Google spokesperson stated that the model performs comparably to OpenAI’s Astra and Anthropic’s Opus in coding and cybersecurity evaluations. Argon has demonstrated the ability to independently locate software vulnerabilities, verify them, and generate patches, scoring 68 percent on the CWE-bench v1 benchmark. According to Google, this score ties for first place. The company also claims it has significantly improved the ability to find previously undetected vulnerabilities compared to earlier models.
Gemini 4 Argon rollout replaces cancelled Pro release
The rollout of Gemini 4 Argon replaces the previously announced Gemini 3.5 Pro model, which CEO Sundar Pichai had scheduled for a June release. The cancellation coincides with a leadership shakeup at Google DeepMind. Founder and CEO Demis Hassabis has stepped back from his role, and several senior executives overseeing the Gemini models have departed the company.
Despite previous criticism that Google was losing ground in the artificial intelligence race, Pichai rejected those assertions in July. The company has shifted its marketing strategy in recent months, emphasizing the cost efficiency of its AI products over cutting-edge capabilities alone. Through Argon, the firm hopes to recover the prominence lost to models from Anthropic and OpenAI in recent months.
Operational Implementation
Argon is not merely a laboratory experiment and is already rewriting portions of Google’s own code into Rust. Google indicates that completed optimizations are expected to free more than 300 TiB of memory, with total potential estimated between 500 TiB and 1 PiB.
Frequently Asked Questions About Gemini 4 Argon
Who can currently use Gemini 4 Argon?
The model is not available for public use. Google is limiting initial access to select cybersecurity partners and internal teams through the Fairwind program.
What happened to Gemini 3.5 Pro?
Google cancelled the release of Gemini 3.5 Pro, which had been originally scheduled to launch in June, choosing instead to introduce Gemini 4 Argon.
How does Argon’s pricing compare to competitors?
Argon costs $2 dollars per million input tokens and $10 dollars per million output tokens, compared to $10 dollars and $50 dollars for rival models like Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra.
Technical Specifications
- Output Token Limit: Up to 1 million tokens.
- Programming Benchmark: 77.9 percent in DeepSWE v1.1.
- Security Benchmark: 68 percent in CWE-bench v1.
- Initial Availability: Fairwind cybersecurity program participants including Wiz.
Google previously announced Gemini 3.5 Pro for a June release before ultimately restructuring DeepMind and launching Gemini 4 Argon.