Gemini 4 Argon: our next era of frontier intelligence

What it claims

Gemini 4 Argon is Google's new frontier model, aimed at long-horizon coding, enterprise knowledge work (legal, finance, tax) and defensive cybersecurity. It is not generally available yet. Access is limited to trusted cyber defenders and testers through the Fairwind Program. For those defenders and for Google's internal teams, the model ships without cyber guardrails. Wider release will start with paid API customers and Google AI Ultra subscribers, then developers, enterprises and consumers. The page gives no date for that. Google is also taking part in the U.S. government's voluntary pre-release model access process during the phased rollout.

Evidence

Beyond the headline scores in the summary, the page gives these details:

• Comparisons are mostly unquantified. DeepSWE, LVBench and AutomationBench are reported as state of the art or #1 without competitor scores. The Vals Index, Vals Finance Agent v2, Harvey's Legal Agent Benchmark and Gray Swan IPI are reported as "leading" with no numbers at all.
• CWE-bench v1. The 68% is a tie, and the page does not name the other model. Google's prior CWE-bench result was 3.8 Flash Cyber on v0.
• Vulnerability discovery against 3.8 Flash Cyber. Two benchmarks are qualitative only:
• Google's internal benchmark spans 20 languages.
• Wiz's internal black-box web pentest benchmark covers attack-surface discovery, vulnerability identification and proof-of-concept generation, all without source code.
• Wiz Scan for Good. Argon found a critical PII-exposing vulnerability in healthcare software used by hospitals worldwide, which earlier frontier models had missed. No CVE or vendor is named.
• Quantum optimisation. Argon beat a published baseline on spacetime cost (qubits × gates) of a bottleneck subroutine by 40%, in minutes. This is a single example.
• Fleet memory. Agents working from fleet-wide profiling telemetry freed over 300 TiB once rolled out. Google estimates 500 TiB to 1 PiB in total savings, but that range is a projection, not a measurement.
• libgav1. The starting point was an existing Rust port, not the C++ original. Agents replaced 32K lines of SIMD code with safe Rust written so the compiler auto-vectorises it. They worked through repeated profile-guided experiments and by reading the compiler's output. The result:
• 2.7x faster than the Rust port
• identical video output
• "closer to," but not matching, the optimised C++

How to use it

• Pricing (introductory): $2 per million input tokens and $10 per million output tokens. Cached input tokens are 95% off the input price.
• Output limit: 1M tokens. The page presents this as room for single trajectories of hundreds of thousands of tokens.
• Not stated: API model identifier, context window, endpoints and general-availability date. You cannot adopt the model today unless you are in Fairwind.

Limits

• Pricing may change. It is explicitly introductory, and the page does not say what follows.
• Migrations are not in production. The large C/C++ to Rust work spans re2 and libgav1 (tens of thousands of lines) up to Fuchsia Zircon (800K+ lines). All of it is still undergoing automated and manual auditing, emulation testing and review before rollout. Treat it as unproven in production.
• Guardrails will apply on general release. General users will hit refusals for cyber and CBRN requests, enforced in part by monitoring the model's internal activations. Only trusted defenders get the unguarded model.
• Agent runs can be halted. A misalignment monitor watches Argon's chain-of-thought and actions and "stop[s] execution when necessary." If you are building long agentic workflows, plan for runs being interrupted by this monitor. The page does not describe the stopping criteria.
• Evidence is internal or partner-run. Most supporting results come from internal benchmarks, partner benchmarks (Wiz, Zapier, Harvey) or Google case studies, with no published methodology.