A million output tokens is a lot of room to think, though room alone does not guarantee good judgment. Google’s new Gemini 4 Argon is designed to use that capacity on long, multi-step work, from software engineering and financial research to legal drafting and cybersecurity defense. For now, access is limited while the company tests the model and its safeguards.
Built for work that takes more than one step
Google says Argon can generate up to 1 million output tokens in a request, up from a previous limit of 64,000. The larger limit is meant to let the model sustain longer tasks, rather than repeatedly breaking work into smaller exchanges.
That distinction matters most when the task involves several stages: examining a codebase, tracing a problem, testing a possible fix, then explaining what changed. A larger token budget gives the model more room to carry out that sequence. It does not, by itself, establish that every step is correct, so the claims about Argon’s performance still need to be read alongside the tests and examples Google has shared.
Google says the model is already being used inside the company for specialized coding, research and writing. Its examples include work on quantum algorithms, large software migrations and memory use in Google’s data centers.
Google’s reported results
The company reports a score of 77.9% on DeepSWE v1.1, a benchmark for longer-horizon software engineering tasks. It also says Argon ranks first on AutomationBench, with a score of 51.3%, and reached 91.7% on LVBench, a test of long-video understanding.
For vulnerability remediation, Google reports a 68% score on CWE-bench v1, tied for the top result. The company also says Argon leads on the Vals Index, which assesses economic impact across areas including finance, coding, legal work and tax. It reports leading results on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark, without giving scores for those evaluations in its announcement.
These figures describe Google’s reported tests, not a universal ranking. Benchmark results depend on the test’s version and method, so comparisons with other models are meaningful only when those conditions match. The numbers are useful evidence of what Google says it has measured, not a substitute for seeing how the system performs in ordinary work.
Code migrations and memory savings
Google’s examples from internal projects offer a more tangible view of the model’s intended role. Argon agents are helping move C and C++ codebases to Rust, including work on the Fuchsia Zircon kernel, which the company says contains more than 800,000 lines of code. Google says these changes are subject to automated and manual audits, emulation testing and review before they reach production.
In another project, agents working on libgav1, Google’s open-source video decoder, replaced 32,000 lines of SIMD code in an existing Rust port. Google says the resulting version runs 2.7 times faster than that earlier Rust port while producing identical video output. The company describes the result as bringing the Rust version closer to the optimized C++ implementation. That is a narrower, more useful claim than saying Rust has somehow become faster than C++ in general.
Google also says a group of Argon agents analyzed profiling data from its data-center fleet and identified memory optimizations. The changes freed more than 300 TiB of memory once rolled out, according to the company, which estimates total potential savings of 500 TiB to 1 PiB. In a separate quantum-computing example, Google says Argon improved a published baseline for a subroutine’s resource use by 40% in a matter of minutes.
Those are company-reported results from internal work. They show the kinds of tasks Google is assigning to Argon, while leaving the usual practical questions open: how much human review the work needs, how reliably it transfers to other organizations, and what the cost looks like at scale.
Cybersecurity access starts with trusted defenders
Argon is designed to find, validate and help patch software vulnerabilities. Google says the first external users are trusted cyber defenders participating through its Fairwind Program. The company is also working with Wiz, whose Scan for Good initiative uses the model to look for and remediate high-risk exposures in critical public infrastructure.
In an early example, Google says Argon found a critical vulnerability that exposed sensitive personal information in healthcare software used by hospitals worldwide. The company says earlier frontier models had not detected it. Google also reports that Argon outperformed its previous 3.8 Flash Cyber model on an internal black-box penetration-testing benchmark, where models examine live web systems without access to source code.
For trusted defenders and internal teams, Google says it will provide Argon without cyber guardrails so they can use its full cybersecurity capabilities. That access is limited to those groups, not a general release. The distinction is central to the rollout: a model capable of finding and fixing serious flaws can also create risks if the same capabilities are used against systems without permission.
Safety work before wider release
Google says it is strengthening safeguards against harmful cyber activity and chemical, biological, radiological and nuclear misuse. It has also tested the model against indirect prompt injection, where malicious instructions hidden in material a model processes attempt to redirect its behavior. Google says Argon leads on Gray Swan’s Indirect Prompt Injection benchmark.
The company describes additional measures that monitor the model’s reasoning and actions for signs of misalignment, with systems that can stop execution when needed. It says similar monitoring was used during training, with alerts sent to a dedicated incident response team. Google also says it is isolating and securing sandbox environments before high-risk training and evaluation runs.
These are safeguards Google says it has built and tested, not a guarantee that every failure mode has been found. The company plans to keep gathering feedback from early testers and refining its protections before wider availability. It is also participating in a voluntary U.S. government process for pre-release model access.
Pricing and availability
Google says Argon is rolling out first to trusted cyber defenders and testers. Broader access is planned for developers, businesses and consumers, beginning with paid API customers and Google AI Ultra subscribers. The company has not yet opened general public access.
The introductory API price is $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced 95% below the standard input rate. The reported post-promotion rates are $4 per million input tokens and $20 per million output tokens.
A long output limit makes Argon suited to work that can stretch across many steps. It also makes careful testing more important, because a model can now spend much longer working through a task before anyone has a chance to interrupt it. Google’s first release is therefore less a public launch than a controlled trial, which is probably the sensible order when the product’s selling point is that it can keep going.