Google DeepMind has released Gemini 4 Argon to the public, opening access to a model built for long, multi-step work. Developers, enterprises, and paid Gemini subscribers can now use it after a limited period with early testers and a small group of cyber defenders.
Argon became broadly available on October 13, 2026, through Google AI Studio, the Google Gemini API, and Google Search. The model is a sparse mixture-of-experts system and supports multimodal language work across coding, finance, legal tasks, science, and cybersecurity.
The headline feature is its response size. Argon can write up to 1 million tokens in one response, while its model card describes a combined input and output window of up to 1 million tokens. At introductory pricing, a full 1 million-token response costs $10 in output tokens.
A large window aimed at complex work
That capacity gives Argon room to handle long codebases, large case files, extended research material, and multi-step technical tasks without splitting the work into many smaller exchanges. Google describes the model as delivering frontier reasoning and better performance across complex workflows, with less hedging, repetition, and chatbot-style cliches.
Google compares Argon with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. On Google’s benchmark table, Argon posts the top score on most rows, including 77.9% on DeepSWE v1.1, 68.90% on Vals Index, 51.3% on AutomationBench, and 19.6% on Harvey’s Legal Agent Benchmark.
It also scored 84.2% on GraphWalks, 68.0% on CWE-bench v1, and 87% on the ERQA embodied reasoning benchmark. Independent testing organizations rank Argon first on Vals Index and Finance Agent v2, though the model scores lower on other tests.
The results are not a clean sweep. Argon trails on five rows of Google’s own table: 55.0% on FrontierSWE v2, 57.4% on Terminal-Bench 4.0, 57.6% on Terminal-Bench Science 0.1, 45.3% on PostTrainBench, and 69.2% on OSWorld-2.0.
Sam Witteveen, CEO and co-founder of Red Dragon AI, described Argon as a very good model but said he does not consider it the best model available. He also said its strategic value is significant, which captures the broader picture: Argon does not lead every test, but its scale and professional focus give it a strong position.
Google points to internal gains
Google has used Argon inside its Antigravity development environment, where more than 100,000 Google employees use an internal version each month. Google says agentic code submissions rose by more than 35%, while post-submission code rollbacks fell by nearly a third, or 30%.
The company also reports a 40% reduction in session preparation time and a 21% improvement in cold imports. Inference throughput on an Nvidia A100 GPU increased from approximately 37,900 to more than 108,000 tokens per second, a change that raises the amount of work the system can process in a given period.
Some of the examples are large and practical. Argon helped reorganize a legacy folder containing 6 million lines of browser feature code inside Chromium’s roughly 40-million-line codebase. Google says the task took weeks instead of the months it expected. Argon also identified 97 stale or duplicated software bugs in under three hours.
Google used the model for cybersecurity vulnerability patching, optimizing YouTube comments, and debugging a drone physics simulator at Wing, Alphabet’s drone delivery company. These examples show how the company is testing Argon across software maintenance, security, content systems, and engineering simulations.
Safety and access now move into focus
Google first announced Argon on September 30, 2026, for a small group of cyber defenders. The company says it deployed safeguards that lower biological and cybersecurity risks before making the model broadly available.
That wider release puts Argon’s long context window and professional performance in the hands of developers, enterprises, and paid Gemini subscribers. It also gives users several ways to reach the model, including Google AI Studio, the Google Gemini API, and Google Search.
The launch arrives with a clear trade-off. Argon leads rivals on many professional benchmarks and handles very large tasks, but it trails on tests such as FrontierSWE v2 and Terminal-Bench 4.0. Its introductory price makes a full 1 million-token response possible for $10, yet the value will depend on whether users need that much output and whether the model’s strengths match their work.
For now, Google is positioning Gemini 4 Argon as a model for demanding workflows rather than a universal winner. Its public release expands access to a system already used across Google’s own teams, while the mixed benchmark results leave room for other models to compete.
Based on
