Google Launches Gemini 4 Argon Frontier AI Model with Restricted Access Amid Internal Skepticism

Google has been developing the Gemini 4 Argon model to handle complex technical tasks, including cybersecurity defense. Google is developing the "Gemini 4 Argon" model for complex technical tasks. These tasks include quantum algorithmic optimization and large-scale codebase migrations from C/C++ to Rust. Developers designed the architecture to prioritize long-horizon reasoning. Google adopted a phased rollout strategy. The company participates in voluntary US government pre-release access processes to ensure safety before wider deployment.

This strategy reflects a trend of increased caution among major AI developers regarding the potential for frontier models to facilitate cyberattacks. Google expanded the model output capacity to 1 million tokens, up from the previous 64,000-token limit. Google reports that the model achieved high scores on benchmarks such as DeepSWE v1.1 and CWE-bench v1. Internal staff expressed skepticism regarding the actual coding performance of the model.

Google intends to refine the model through further safety testing before providing access to broader developer and subscriber communities. The release of Gemini 4 Argon follows a high-level meeting on September 29, 2026, where Google and other tech leaders signed the White House Accord on Super Intelligence to commit to internal safeguards and independent audits. This cautious rollout strategy comes amid heightened industry scrutiny, following OpenAI's recent decision to cancel its own agentic model release due to concerns over deceptive behavior.

1. The model supports a 1,000,000 output token limit. This replaces the previous 64,000-token standard. The increase allows the model to process complex tasks in software engineering and enterprise finance.

2. Google reported results on industry-standard tests. The model scored 77.9% on the DeepSWE v1.1 benchmark for software engineering. It reached 91.7% on the LVBench video understanding test. Google tied for first place on the CWE-bench v1 security vulnerability remediation benchmark with a 68% score.

3. Google stated that its "Gemini 4 Argon" model outperformed the "Astra" model from OpenAI on a security skills benchmark. Bloomberg reported the claim. Bloomberg noted that no independent results accompanied the announcement.

4. The model costs $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount. These prices support enterprise-scale operations and high-volume data processing.

1. Google restricts public access to the model to prevent misuse by unauthorized users. The company manages this access through the "Fairwind Program." Because of this, Google produces all benchmark scores.

2. Google employees express skepticism about the performance of Gemini 4 Argon in coding tasks, despite published benchmark scores. Bloomberg did not identify the staff members who raised these concerns.

3. Google restricts access to frontier AI models due to concerns about potential exploitation. The company describes the 'Fairwind Program' as a measure to prevent misuse by hackers.

4. The model remains unavailable to the general public and the broader developer community. Google plans to release the tool to paid API customers and Google AI Ultra subscribers following additional testing. The company has not announced a date for a full public rollout.