Google today announced Gemini 4 Argon, its new frontier model. In a post on Google's official blog on September 30, Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, described Argon as a model built to sustain deep reasoning across complex, long-horizon workflows: real-world software engineering, enterprise knowledge work in fields like legal and finance, and cybersecurity defense.
Why it matters: the headline is not a faster chatbot. It is a model designed to carry a long, multi-step job through to the end in a single run. Google is pairing that with a much larger output budget, strong results on long-horizon benchmarks, and an unusually cautious rollout that starts with cyber defenders rather than the general public.
What Google announced
Gemini 4 Argon is Google's new flagship model. According to Google, it delivers frontier performance on complex workflows in three areas: software engineering, enterprise knowledge work, and cybersecurity defense. The company also highlights creative writing as a strength.
Access is limited for now. Argon is rolling out to a set of trusted cyber defenders through Google's Fairwind Program, alongside a group of trusted testers. Google says safely releasing capabilities at this level requires a phased approach, and that it is participating in the U.S. government's voluntary process for pre-release model access while it gradually expands availability.
The next audiences are developers, enterprises and consumers, "as soon as possible," starting with paid API customers and Google AI Ultra subscribers. Google has not published a date for that wider release.
Pricing is already public. Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% below the standard input price. After the introductory period, the price moves to $4 per million input tokens and $20 per million output tokens.
One million output tokens
The most concrete technical change is output length. Google is raising Argon's output token limit to 1 million tokens, up from the previous 64,000. Google calls this industry-leading.
The reasoning is straightforward. When a model has room to think deeply and generate hundreds of thousands of tokens in a single trajectory, it can work through a difficult problem in one go instead of stopping, summarizing and restarting. For long coding tasks, large document work or extended analysis, that headroom changes what can be attempted in a single request.
It also changes how teams should think about cost. A single long run can now produce far more output than before, so output pricing, not just input pricing, becomes a planning question for anyone building on the model.
Benchmarks: long-horizon work and knowledge work
Google reports that Argon sets a new state of the art on DeepSWE v1.1 with a score of 77.9%. DeepSWE measures performance on real-world, long-horizon software engineering tasks, which is exactly the kind of work the model is built for.
Beyond coding, Google says Argon is the leading model on the Vals Index, which measures economic impact across finance, coding, legal and tax work, with each sector weighted by its contribution to U.S. GDP. Google also reports leading results on Vals Finance Agent v2, which covers multi-step financial research, and on Harvey's Legal Agent Benchmark, which covers legal research and drafting.
On AutomationBench, Zapier's benchmark for end-to-end execution across core business functions, Argon ranks first with 51.3%. On LVBench, which measures long video understanding, it posts a state-of-the-art 91.7%. Google adds that the model is strong when knowledge work depends on visual understanding, such as professional chart analysis or acting on a series of documents.
As with any vendor-published results, these numbers come from Google's own announcement. Independent evaluations will matter once developers get access.
How Google already uses it
Google says thousands of its employees already use Argon internally, and it shared three examples.
- Quantum computing: Argon helped researchers optimize the qubit and gate resources of subroutines that bottleneck important applications. In one case, it beat the published baseline by 40% in minutes.
- Memory efficiency: a team of Argon agents analyzed fleet-wide profiling telemetry and applied memory optimizations across Google's data centers, freeing more than 300 TiB of memory, with an estimated 500 TiB to 1 PiB in total savings.
- Codebase migrations: Argon agents are migrating C and C++ code to Rust, from tens of thousands of lines in core libraries up to more than 800,000 lines for the Fuchsia Zircon kernel. For libgav1, Google's open source video decoder, agents produced a memory-safe decoder that runs 2.7 times faster than the earlier Rust port, with identical video output.
Google notes that these large rewrites go through automated and manual auditing, emulation testing and review before reaching production.
A model trained for cyber defense
Cybersecurity is the most distinctive part of the launch. Google says it trained Argon to be highly capable at defense: the model can autonomously find, validate and patch critical software vulnerabilities. For trusted defenders and Google's internal teams, Argon is being released without cyber guardrails so they can use its full defensive capabilities.
Wiz is already using Argon through its Scan for Good initiative, which protects critical public infrastructure at no cost. According to Google, the model uncovered a critical vulnerability in healthcare software used by hospitals worldwide, a risk that previous frontier models had missed.
On CWE-bench v1, which evaluates vulnerability remediation, Argon ties for first place with 68%. Google also reports gains over Gemini 3.8 Flash Cyber in vulnerability discovery, including on Wiz's black-box penetration testing benchmark for live web systems.
Safeguards before broad release
Google lists four areas it is strengthening before a wide rollout. First, misuse: the model is designed to refuse harmful cyber and CBRN requests while preserving legitimate dual-use research, under Google's Frontier Safety Framework, with monitoring of internal activations to spot misuse. Second, prompt injection: Google calls Argon its most resilient model yet against indirect prompt injection and says it leads Gray Swan's benchmark. Third, misalignment: monitors watch the model's chain of thought and actions and stop execution when it goes beyond what the user intended. Fourth, hardened, sealed sandboxes for high-risk training and evaluation.
What to do
- If you work in security, check whether your organization qualifies for early access through Google's Fairwind Program, and prepare a well-scoped defensive use case.
- If you build on the Gemini API, model your costs with both prices in mind: $2 and $10 per million tokens at launch, then $4 and $20. Long outputs make output pricing the key variable.
- Identify one long, multi-step workflow in your team, such as a migration, a research report or an audit, and define how you would test Argon on it and verify the results once access opens.
Sources
- Google, "Gemini 4 Argon: our next era of frontier intelligence": https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- 9to5Google, "Google says Gemini 4 release is coming 'as soon as possible'": https://9to5google.com/2026/09/24/google-says-gemini-4-release-is-coming-as-soon-as-possible/
