TL;DR: Google announced its Gemini 4 Argon frontier model on September 30, 2026, but limited first access to trusted cyber defenders while it completes safety checks.

Google moves to a new frontier tier

Google on September 30, 2026, announced Gemini 4 Argon as its new frontier model for complex, long-horizon work, according to Google's official Keyword blog post titled "Introducing Gemini 4 Argon."

The announcement was authored by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. He described Argon as built to sustain deep reasoning across complex workflows.

Google positioned the system for three priority areas: real-world software engineering, enterprise knowledge work such as legal and finance tasks, and cybersecurity defense. The company said the model delivers frontier performance across those workflows.

The name marks a clear generational step. Google had spent the summer releasing smaller Flash models after discussing a Gemini 3.5 Pro model expected in June, according to Ars Technica in its report titled "Google announces Gemini 4 Argon AI model, but you can't use it yet."

Google did not present Argon as broadly available on announcement day. It framed the debut as the start of a phased release focused first on defensive cybersecurity use.

Why most users cannot try it yet

Immediate access is restricted to what Google calls a set of trusted cyber defenders through its Fairwind Program.

Google said safely releasing capabilities at this level requires a phased approach. It said it is actively engaged in the U.S. government's voluntary process for pre-release model access while it gradually expands access.

The company said it will gather feedback from early testers as it iterates on guardrails before making Argon available to developers, enterprises, and consumers as soon as possible.

Reporting by The Verge in "Google announces Gemini 4 and says it's so capable that only 'trusted cyber defenders' can have it right now" likewise described a limited launch intended to allow checks for misalignment.

Google said it will strengthen what it calls critical frontier safeguards before broader rollout. Those safeguards include defenses against misuse and prompt injection attacks and monitoring for misalignment.

Ars Technica reported that Google has not offered a specific timeline for general availability. That same reporting said general availability is expected to begin with paid API users and Google AI Ultra subscribers after the cybersecurity testing phase.

In other words, the completed event on September 30 was the announcement and limited rollout. Wider release to developers, businesses, and consumers was described by Google as a future step.

Price and output capacity disclosed

Although general access remains future, Google disclosed introductory API pricing for Argon.

Google said Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens. Cached input tokens will be priced at 95% off the input token price.

Ars Technica noted that the pricing was described as available for a limited time. The figures establish the commercial structure Google intends to use when API access expands.

Google also disclosed a major capacity change. The company said it is expanding Argon's output token limit to 1 million tokens, up from the previous 64,000 tokens.

The company said the additional headroom is intended to let the model think deeply and generate hundreds of thousands of tokens in a single trajectory. Google said that depth can help solve difficult problems in one extended run.

The Verge reported that Google shared benchmark charts comparing Gemini 4 with competing models from OpenAI and Anthropic. Those comparisons were presented by Google as part of the launch materials.

Coding results lead Google's performance claims

Software engineering was central to Google's launch case.

Google said Argon sets a new state of the art on DeepSWE v1.1 at 77.9%. Google described DeepSWE v1.1 as a measure of performance in real-world long-horizon software engineering tasks.

Ars Technica reported that Google's claimed 77.9% score was higher than GPT-6 Astra, Fable 5.1, and Opus 5.5. Those rival names and the comparison were included in Google's launch presentation materials discussed in that coverage.

Google engineers have used Argon for daily tasks ranging from debugging to large-scale codebase migrations and algorithm design, according to Google. The company said thousands of Googlers have highlighted strengths in specialized coding tasks, deeper research, and writing quality.

Google also pointed to enterprise automation and visual understanding. It said Argon ranks first on AutomationBench, Zapier's benchmark for end-to-end execution across business functions, with a score of 51.3%.

For long video understanding, Google reported a state-of-the-art score of 91.7% on LVBench. The company said that capability supports chart analysis, identification of details from long videos, and actions based on document series.

Finance, legal, and economic-impact tests

Beyond code, Google emphasized knowledge work tied to measurable economic activity.

Google said Argon is the leading model on the Vals Index. Google described the Vals Index as measuring economic impact across finance, coding, legal, and tax work, with sectors weighted by their contribution to U.S. GDP.

The company also cited leading performance on Vals Finance Agent v2 for multi-step financial research. It separately cited Harvey's Legal Agent Benchmark for legal research and drafting.

Those three evaluations — Vals Index, finance-agent research, and legal-agent research — were presented together as evidence that the model can sustain multi-step professional tasks.

Ars Technica characterized the announcement as backed by a raft of benchmarks covering long-horizon tasks. The Verge similarly reported that Google presented many benchmarks where Gemini 4 compared favorably with OpenAI and Anthropic systems.

Google did not announce consumer-facing finance or legal products on September 30. The cited results were model evaluations relevant to future developer and enterprise use.

Internal Google projects cited as early proof

Google said Argon is already changing internal work and engineering.

One example involved quantum computing research. Google said Argon helped researchers optimize the spacetime resources, measured as qubits multiplied by gates, of subroutines that limit important applications. In one example, Google said the model beat the published baseline by 40% in a matter of minutes.

A second example involved data-center memory. Google said a team of Argon agents analyzed fleet-wide profiling telemetry to identify and apply memory optimizations across Google's data centers. The company said the work freed up over 300 TiB of memory once rolled out, with estimated total savings of 500 TiB to 1 PiB.

A third set of examples involved migration of C and C++ codebases to Rust. Google said Argon agents are working on efforts ranging from tens of thousands of lines in core libraries such as re2 and libgav1 up to more than 800,000 lines for the Fuchsia Zircon kernel.

Google said those rewrites are undergoing rigorous automated and manual auditing, emulation testing, and review before production rollout because many affected systems are critical.

For libgav1, Google's open-source software for decoding video, Google said Argon agents took an existing Rust port and replaced 32,000 lines of SIMD code through profile-guided experiments and study of compiler output. The result, according to Google, was memory-safe Rust code that the compiler could vectorize automatically.

Google said that libgav1 result runs 2.7 times faster than the prior Rust port, produces identical video output, and moves closer to optimized C++ performance. Ars Technica also highlighted the re2, libgav1, and 800,000-line Fuchsia Zircon figures as examples of internal use.

Cybersecurity defense is the first rollout

Cybersecurity is both the first deployment track and the reason Google gave for restricting access.

Google said it trained Argon to be highly capable at cybersecurity defense. It said the model can autonomously find, validate, and patch critical software vulnerabilities.

For trusted defenders and Google's own internal teams, Google said it will release Argon without cyber guardrails so those users can use its full defensive capabilities. That distinction applies only to the authorized defensive-testing group, not to general users.

Wiz is already using Argon for cybersecurity defense through its Scan for Good initiative, according to Google. Google described that program as dedicated to protecting critical public infrastructure without charge by finding and remediating high-risk exposures.

In an early demonstration, Google said the model uncovered a critical vulnerability that could expose personal information in a system used at hospitals around the world. Ars Technica reported that Google claimed other frontier models missed the flaw but did not provide specifics.

Google also linked the cautious rollout to concern about misalignment after high-profile hacking incidents over the summer, as discussed in Ars Technica's coverage. Google said Argon was designed with systems that monitor its chain-of-thought and can halt the model if it moves outside bounds.

Google said reasoning transparency is important in frontier-model development for that reason. The Verge reported that Google plans additional work on misuse defenses, prompt-injection defenses, and misalignment monitoring before expanding availability.

What happens next

The next stage, as described on September 30, is continued testing rather than immediate consumer launch.

Google said feedback from early testers will inform guardrail changes. Participation in the federal voluntary pre-release access process will continue alongside gradual expansion.

Paid API users and Google AI Ultra subscribers were identified by Ars Technica as the expected first groups for general availability after cybersecurity testing. Google itself said developer, enterprise, and consumer availability would follow as soon as possible, without naming a date.

The timing placed Google one day after OpenAI's DevDay conference, where OpenAI introduced its Dots AI agent and GPT-6.1 Sol model, according to The Verge. That same Verge report noted OpenAI had said it would not release a planned GPT-6.1 Astra model because of safety concerns.

The launch also followed Kavukcuoglu's appointment as head of DeepMind in August, as noted by The Verge. For now, consumers and most developers remain observers while defenders test whether Argon's coding, reasoning, and patching performance holds outside Google's internal systems.

Sources referenced