Google unveils Gemini 4 Argon as its new frontier model for long-running coding, enterprise and defensive cybersecurity work. The company is limiting the first release to trusted cyber defenders while it tests safeguards before broader developer and consumer access.

 

Google has also disclosed introductory API pricing and a one-million-token output limit, but it has not given a public launch date. That combination makes Argon a consequential model announcement whose strongest performance claims still await independent testing.

 

The Gemini 4 Argon announcement establishes four immediate facts:

  • Trusted cyber defenders get the first external access.
  • The output limit rises from 64,000 to one million tokens.
  • Introductory API pricing starts at $2 per million input tokens.
  • Broader availability has no firm date.

 

Related Coverage

 

Google Unveils Gemini 4 Argon to Trusted Cyber Defenders

Google announced Argon on September 30 and said the model is rolling out through its Fairwind Program. The initial group consists of trusted cybersecurity defenders rather than the general public, giving Google a controlled setting to measure useful security work and possible misuse.

 

The company is also participating in the US government's voluntary process for pre-release access to advanced models. Google said feedback from early users will shape guardrails before Argon reaches developers, enterprises and consumers, beginning with paid API customers and Google AI Ultra subscribers.

 

The official announcement lists introductory prices of $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced at a 95% discount. Google did not publish a calendar date for general availability.

 

Gemini 4 Argon Extends Long-Horizon Coding Work

Argon's most distinctive published specification is a one-million-token output limit, up from 64,000 tokens in Google's previous model. An output allowance that large could let an agent sustain unusually long reasoning and implementation traces, although practical reliability across such trajectories remains unproven outside Google.

 

Google says its engineers have used Argon for debugging, algorithm design and large code migrations. The company described internal projects ranging from converting C and C++ systems to Rust to optimizing a Rust video decoder, while stressing that critical rewrites remain subject to automated and manual review.

 

On Google's reported evaluations, Argon scored 77.9% on DeepSWE v1.1, a benchmark for long-horizon software-engineering work. It also ranked first on AutomationBench with 51.3%, according to the company. These figures are useful signals, but they are not substitutes for reproducible third-party comparisons under identical tools and prompts.

 

Google positions the model beyond programming. It says Argon performed strongly on tests covering finance, law, tax and long-video understanding. That pitch places the model in competition for expensive professional workflows where accuracy, auditability and access to private data matter as much as headline benchmark scores.

 

Gemini 4 Argon Targets Defensive Cybersecurity

Cybersecurity is the center of the staged rollout. Google says Argon can find, validate and patch software vulnerabilities autonomously, and that approved defenders will receive access without the cyber guardrails applied to ordinary users. That exception raises the value of careful participant screening and monitored environments.

 

Wiz is testing the model through its Scan for Good initiative, which examines critical public infrastructure for high-risk exposures. Google reported that Argon identified a serious vulnerability affecting healthcare software that earlier frontier models had missed, but it withheld technical details that could expose affected systems.

 

Google also reports a 68% score on CWE-bench v1, tied for first place on that vulnerability-remediation test. The company says internal evaluations covered complex codebases in 20 programming languages and black-box testing of live web systems. Independent researchers have not yet verified those results publicly.

 

Argon's security capabilities create a dual-use problem: the same reasoning that helps defenders can assist attackers. Google's safeguards therefore cover cyber and chemical, biological, radiological and nuclear misuse, indirect prompt injection, agent misalignment and hardened sandboxes for high-risk testing.

 

Pricing Is Public, but the Release Date Is Not

Google's rollout strategy separates announcement, controlled access and broad commercial release. The published token prices help enterprises model future costs, yet customers still lack firm dates, rate limits, service guarantees and detailed regional availability.

 

The company says Argon has already supported internal quantum-algorithm optimization and data-center memory work. One internal deployment reportedly freed more than 300 tebibytes of memory, with larger savings estimated. Those examples show the scale Google wants Argon to address, not a promise that outside customers will reproduce the same outcomes.

 

For buyers, the next meaningful evidence will be access to the actual API, independent benchmark runs and disclosure of production constraints. Until then, Gemini 4 Argon is best understood as a controlled frontier-model launch with concrete pricing, unusually large output capacity and a cybersecurity-first testing program.