Google Launches Gemini 4 Argon for Cyber Defenders First
Google on 30 September 2026 launched Gemini 4 Argon for trusted cyber defenders first, citing broad enterprise benchmark leads and introductory API pricing of $2/$10 per million tokens.
PromptCrates Editorial
Staff Writer
Google on Wednesday, 30 September 2026, announced Gemini 4 Argon, a frontier model the company says leads or ties rivals on a majority of disclosed enterprise benchmarks while starting life in a tightly limited rollout. In a Google blog post by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, Argon is framed for long-horizon software engineering, legal and finance knowledge work, and cybersecurity defense. Access begins with trusted cyber defenders through Google’s Fairwind Program, with broader developer, enterprise, and consumer availability planned after U.S. voluntary pre-release reviews.
How Gemini 4 Argon is rolling out
Google is not treating Argon as a same-day consumer chatbot drop. The company says it is engaged in the U.S. government’s voluntary process for pre-release model access and will expand gradually while it strengthens safeguards. TechCrunch and VentureBeat both emphasize that Fairwind partners are the first public cohort. Google’s blog, as VentureBeat also reports, names paid API customers and Google AI Ultra subscribers as the opening wave when wider access arrives.
Pricing is aggressive on the way in. Argon launches at an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input tokens at a 95 percent discount to the input price. After the introductory period, Google says pricing moves to $4 per million input tokens and $20 per million output tokens. VentureBeat notes that the introductory list rate is one-fifth of GPT-6 Astra’s listed $10 / $50 per million tokens and half of Claude Opus 5.5’s $4 / $20 list, though buyers still cannot validate production cost until API access opens.
Google also raised Argon’s maximum output to one million tokens, up from a prior 64,000-token ceiling. That ceiling matters for agentic migrations, audits, and legal review chains that previously hit generation walls mid-task. The phased access story sits beside other Google AI packaging moves covered on PromptCrates, including the company’s Gemini educator UK roadshow and research on Diffusion Controller image steering.
Where Argon leads and where it trails
Google’s disclosed comparison table, summarized by VentureBeat, covers 18 benchmarks. Argon leads outright on 12 and ties for first on one; GPT-6 Astra leads three and ties one; Claude Opus 5.5 leads two. Strong Argon numbers include Harvey’s Legal Agent Benchmark at 19.6 percent versus 5.4 percent for GPT-6 Astra and 3.8 percent for Claude Opus 5.5; AutomationBench at 51.3 percent versus 41.4 percent for GPT-6 Astra and 42.5 percent for Claude Opus 5.5; GraphWalks at 84.2 percent versus 71.8 percent and 66.8 percent; Vals Finance Agent v2 at 65.4 percent; DeepSWE v1.1 at 77.9 percent; and LVBench long-video understanding at 91.7 percent. On CWE-bench v1 vulnerability remediation, Argon ties GPT-6 Astra at 68 percent.
The race is not a clean sweep. VentureBeat reports Astra still leads FrontierSWE v2 and Terminal-Bench Science 0.1 by 10.5 points each, while Claude Opus 5.5 leads Terminal-bench 4.0 and PostTrainBench. Buyers should treat “most top scores” as a breadth claim, not proof that Argon wins every coding or terminal workload. Google separately cites the Vals Index leadership claim for economic-impact knowledge work and Gray Swan indirect prompt-injection robustness, where Argon’s attack success rate is listed at 0.7 percent versus higher rates for several peer models.
Internal Google examples are concrete if still self-reported. Argon helped quantum researchers beat a published spacetime baseline by 40 percent in minutes; Argon agents identified memory optimizations expected to free more than 300 TiB once rolled out, with 500 TiB to 1 PiB estimated total savings; and Argon agents are migrating C/C++ codebases to Rust, including work on libgav1 where a SIMD rewrite produced a memory-safe decoder Google says runs 2.7 times faster than the prior Rust port with identical video output. Those stories matter for CIOs evaluating whether frontier agents can touch production-critical systems under audit.
Cyber defense focus and safety gates
Cybersecurity is the launch wedge. Google says Argon can autonomously find, validate, and patch critical vulnerabilities, and that trusted defenders and internal Google teams will receive Argon without cyber guardrails so they can use full defensive capability. Wiz is already using Argon through its Scan for Good initiative; Google says the model found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide. That defensive framing lands in the same news cycle as OpenAI’s safety delays and industry agent-security work such as Nvidia’s open agent safety platform and PromptCrates coverage of OpenAI’s GPT-6.1 Astra safety delay.
Before broad availability, Google says it is strengthening defenses against cyber and CBRN misuse, indirect prompt injection, misalignment during long tool-using trajectories, and insecure agent sandboxes. Monitoring of chain-of-thought and actions, sealed evaluation environments, and refusal behavior for harmful dual-use requests are all called out in the blog. Enterprise buyers should ask how Fairwind feedback will change guardrails before Ultra and API GA, and whether “without cyber guardrails” for defenders implies different product SKUs or policy exceptions.
Primary reporting for this article: Google’s 30 September 2026 Gemini 4 Argon blog by Koray Kavukcuoglu; TechCrunch by Lucas Ropek; and VentureBeat by Carl Franzen. Anchored facts include Fairwind-first access, U.S. voluntary pre-release process, introductory and post-introductory token prices, one-million-token output limit, disclosed benchmark leads and deficits, internal productivity examples, Wiz Scan for Good use, and the planned path to paid API customers and Google AI Ultra subscribers, which Google’s blog and VentureBeat state.
