Google Gemini 4 Argon is here: 1M-token output, advanced coding and cyber skills

Google has announced Gemini 4 Argon, a new frontier AI model designed to handle complex, long-running tasks across software engineering, enterprise work and cybersecurity.

The model is initially being made available to a group of trusted cybersecurity defenders through Google's Fairwind Program, with broader access planned after further testing and refinement of its safety measures.

Google said Argon is being introduced gradually as it continues to work with early testers on guardrails. The company said it plans to make the model available to developers, enterprises and consumers in the future.

At launch, Gemini 4 Argon will be priced at $2 per million input tokens and $10 per million output tokens. Cached input tokens will be priced at 95 per cent below the standard input-token rate, according to Google.

1 million-token output limit

One of the major upgrades in Gemini 4 Argon is its expanded output capacity. Google said the model can generate up to 1 million tokens, a significant increase from the previous 64,000-token limit.

The company said the larger context for generation is intended to allow Argon to work through lengthy and complicated tasks in a single trajectory, giving it more room for extended reasoning and problem-solving.

Google said its engineers are already using Argon for tasks ranging from debugging and algorithm design to large-scale codebase migrations. The model recorded a 77.9 per cent score on DeepSWE v1.1, a benchmark for real-world, long-horizon software engineering tasks.

The company also highlighted Argon's performance in enterprise knowledge work spanning finance, legal and tax-related tasks, as well as end-to-end business automation. On AutomationBench, Google's stated score for Argon is 51.3 per cent.

Google says Argon is already powering internal work

Google said thousands of its employees are using Gemini 4 Argon for specialised coding, deeper research and writing-related tasks.

In one example, Argon helped Google's quantum computing researchers optimise subroutines by reducing the spacetime resources required. Google said the model beat a published baseline by 40 per cent within minutes.

The company also said Argon agents analysed data-centre profiling telemetry and identified memory optimisations that could free more than 300 TiB of memory once deployed, with estimated total savings of 500 TiB to 1 PiB.

Argon is also being used in Google's efforts to migrate C and C++ codebases to Rust. The work ranges from tens of thousands of lines of code to more than 800,000 lines in the Fuchsia Zircon kernel, with Google saying the migrations are subject to automated and manual audits, emulation testing and reviews before production deployment.

Focus on cybersecurity

Cybersecurity is another major focus of Gemini 4 Argon. Google said the model has been trained to identify, validate and patch critical software vulnerabilities autonomously.

For trusted cybersecurity defenders and Google's internal teams, the company said it will release Argon without cyber guardrails to allow them to use its full cybersecurity capabilities.

Google said cybersecurity company Wiz is already using Argon through its Scan for Good initiative. In an early demonstration, the model reportedly identified a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide — a risk that Google said previous frontier models had missed.

On CWE-bench v1, which measures the ability to remediate security vulnerabilities, Argon tied for first place with a score of 68 per cent, according to Google.

Google adds safeguards before wider rollout

Despite the model's expanded capabilities, Google said it is strengthening safeguards before making Gemini 4 Argon broadly available.

The company said its safety work covers misuse prevention, protection against prompt injection attacks, monitoring for potential misalignment and hardening the environments used to test frontier models.

Google said Argon has been designed to reject harmful requests linked to cyber or chemical, biological, radiological and nuclear attacks while supporting legitimate dual-use scientific research. It is also being tested against indirect prompt injection attacks, in which malicious instructions attempt to manipulate an AI model's behaviour.

Google further said it is deploying systems to monitor Argon's reasoning and actions and stop execution when necessary if the model moves beyond a user's intended task.

The company said wider availability will follow its phased rollout as it gathers feedback and continues refining the model's safety systems.