- Google DeepMind presents Gemini 4 Argon, geared towards programming, business work and cyber defense.
- The model raises the output to one million tokens and is being deployed first through Fairwind for security teams.
- Google publishes results such as 77,9% in DeepSWE v1.1, 68% in CWE-bench v1 and 91,7% in LVBench.
- Access will later be available to API and Google AI Ultra customers, with an initial price of $2/10 per million tokens.
Google DeepMind has unveiled Gemini 4 Argon, a cutting-edge artificial intelligence model designed to handle complex professional tasks. Programming, business intelligence, and cybersecurity are its three main pillars. The company positions it as a further step in its Gemini family and asserts that its internal test results are benchmark-leading.
The rollout won't be widespread from the start. Initial access is restricted to trusted partners through the Fairwind program, with a focus on cybersecurity teams. Later, it will be available to paying API customers and Google AI Ultra subscribers. There's no specific date yet for Spain and the rest of Europe.
A model designed for long, multi-step tasks
The main technical innovation is its output limit of up to one million tokens , compared to 64.000 in previous generations. This allows for sustained reasoning and the production of large volumes of code, documentation, or analysis without constantly fragmenting the response.
Google guides you towards multi-step workflows. It's not just about answering individual queries, but about chaining actions together over extended periods , reviewing documentation, researching, and completing specialized tasks.
The company has also highlighted its internal use. Thousands of employees have tested it for debugging, research, and code migration. C and C++ projects to Rust are part of the testing, with human review and automated testing before production.
In terms of infrastructure, Argon-based agents analyzed data center telemetry and identified optimizations. Google claims that over 300 TiB of memory were freed up and that the savings could be even greater. These figures were provided by the company and are pending external verification.
Results that Google presents as a reference
Google has released figures that place Argon at the forefront. In DeepSWE v1.1 , a test focused on long-term software engineering, it scored 77,9%. The company presents it as having the highest score in that evaluation.
In professional work, it tops the Vals index, which weighs finance, programming, law, and taxes. It also achieves 51,3% in AutomationBench , a Zapier benchmark for business processes, and 91,7% in LVBench for long videos.
In cybersecurity, Argon achieves 68% in CWE-bench v1 , tying for first place, according to Google. However, Claude Opus 5.5 comes out on top in FrontierSWE and Terminal-Bench. GPT-6 Astra surpasses Argon in OSWorld 2.0 offline with 72,6% compared to 69,2%.
These comparisons are largely based on manufacturer data. Since this data is not generally available, independent evaluators cannot yet replicate the tests. This necessitates interpreting the rankings with caution.
Cybersecurity: the first field of deployment
The model has been trained to find, validate, and fix vulnerabilities. According to Google, it can analyze exposed code and systems, even without access to the source code. Cybersecurity is the reason its release is limited.
The Fairwind program provides access to governments, public entities, and security partners. Priority groups include government agencies, healthcare organizations, and telecommunications companies. The goal is for defenders to be able to use it before attackers.
Google cites a collaboration with Wiz in which Argon reportedly detected a critical vulnerability in healthcare software used in hospitals. The flaw could expose personal data. Previous models would not have detected it , according to the company.
To prevent misuse, Google mentions monitoring, prompt injection resistance, and isolation of sensitive environments. A version without some cybersecurity barriers is reserved for accredited defenders. General access would maintain stricter filters.
Availability, prices, and outstanding questions
The rollout will be gradual. First Fairwind, then paid API customers and Google AI Ultra subscribers. There's no firm date for the general public , nor for Spain or the rest of Europe. Google says it will be as soon as possible.
The introductory price will be $2 per million entry tokens and $10 per million exit tokens. Afterward, it will increase to $4 and $20, respectively. Cached entry tokens will be discounted by 95%.
Google also participates in the U.S. government's voluntary process of reviewing advanced models before making them more widely available. This is a practice already followed by other manufacturers. The pre-release review has become part of the standard schedule.
The arrival of Argon intensifies competition between Google, OpenAI, and Anthropic. It also raises questions about costs, oversight, and enterprise integration. Benchmarks are part of the picture , not the whole picture.
Gemini 4 Argon is presented as a cutting-edge model for long-term projects, code, enterprise, and cyber defense. Its large token offering, test results, and restricted deployment define its strategy. Fairwind, Google AI Ultra, and the payment API determine access. Pending dates and independent verification, Spain and Europe will have to be content, for now, with following the announcement from afar.