Google introduces Gemini 4 Argon for long, complex professional work

Google announced a new frontier model, Gemini 4 Argon, on September 30, but it is not hitting the general API yet. According to the company, the rollout starts with a small group of trusted cyber defenders through its Fairwind Program, while Google works through the U.S. government's voluntary pre-release access process. Two numbers matter while it broadens: an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price, and an output limit Google calls industry-leading, at 1M tokens, up from 64K before.

Built for long, multi-step tasks

Argon is positioned for long-horizon professional work: multi-step software engineering, legal and finance knowledge work, and cybersecurity defense. The headline capability is the expanded output limit. Google says that when the model can generate hundreds of thousands of tokens in a single trajectory, it sustains deep reasoning on tough problems in one pass instead of a chain of short calls. Whether that translates into cheaper or more reliable production workflows is something you'll want to test for your own tasks when access opens up.

How it's being used at Google

The launch post leans on internal Google examples, all company-reported. In quantum computing research, Google says Argon optimized qubit-by-gates resources of subroutines, beating a published baseline by 40% in minutes. A team of Argon agents profiled data-center telemetry and applied memory optimizations that freed over 300 TiB, with an estimated 500 TiB to 1 PiB in total savings. On code, Google reports C/C++-to-Rust migrations scaling up to 800K+ lines, including the Fuchsia Zircon kernel, and a rewrite of 32K lines of SIMD code in libgav1 (an open-source video decoder) that runs 2.7x faster than the existing Rust port with identical output.

The benchmark picture

Google reports Argon at a new state of the art on DeepSWE v1.1 (77.9%), which measures long-horizon real-world software engineering tasks. It also claims the top spot on the Vals Index (which weights finance, coding, legal and tax work by U.S. GDP contribution), #1 on Zapier's AutomationBench with 51.3%, state of the art on LVBench for long video understanding at 91.7%, and a tie for first on CWE-bench v1 for vulnerability remediation at 68%. These are launch numbers from the vendor's own post; independent reproduction is still limited, so it's reasonable to treat them as a baseline to re-check rather than a settled ranking.

Cyber defense, and a guardrail-free release

Argon was trained for defensive cybersecurity: autonomously finding, validating and patching vulnerabilities. Notably, Google says it will release the model without cyber guardrails to trusted defenders and its own internal teams so they can use its full capabilities. Wiz is already using it in its Scan for Good program, and Google says an early demonstration surfaced a critical vulnerability in healthcare software exposing personal data, a risk previous frontier models missed. That is the most consequential part of the announcement for anyone running production systems: a frontier model distributed to third parties with its offensive-side defenses off.

Safety work before broad release

Google describes safeguards in four areas: refusing harmful cyber or CBRN requests while keeping legitimate dual-use research usable, under its Frontier Safety Framework; prompt-injection robustness, where it claims the lead on Gray Swan's indirect prompt injection benchmark; misalignment monitoring that watches the model's chain-of-thought and stops execution when needed; and hardened, sandboxed environments for testing. The rollout order afterward is paid API customers and Google AI Ultra subscribers first.

For teams planning around model APIs, the practical read is that access, not capability claims, is the bottleneck today: Argon is behind a phased rollout, and the $2/$10 per-million-tokens pricing sets the cost baseline. When it reaches the public API, the numbers worth verifying are the ones you can reproduce on your own workloads, not the launch table.