Gemini 4 Argon: our next era of frontier intelligence

What it claims

Gemini 4 Argon is not generally available yet. Access is limited to cyber defenders and trusted testers in the Fairwind Program. The page says the next release goes to paid API customers and Google AI Ultra subscribers, and then to developers, enterprises and consumers more broadly. It gives no date. For vetted defenders and Google's internal teams, Argon ships without cyber guardrails, so the public release will behave differently on security tasks.

Evidence

All results are the vendor's own. Most give a score but no competitor scores.

• DeepSWE v1.1 (long-horizon software engineering): 77.9%, described as state of the art.
• AutomationBench (Zapier, end-to-end business functions): 51.3%, ranked #1.
• LVBench (long video understanding): 91.7%, state of the art.
• CWE-bench v1 (vulnerability remediation): 68%, which ties for first rather than leading.
• The page claims a lead on the Vals Index, Vals Finance Agent v2, Harvey's Legal Agent Benchmark and Gray Swan's Indirect Prompt Injection benchmark, with no numbers for any of them.
• Vulnerability discovery is compared only with 3.8 Flash Cyber. Both benchmarks are internal: Google's covers 20 languages, and Wiz's is a black-box penetration test of live web systems without source code. No figures are published for either.

Internal case studies:

• libgav1 (video decoder):
• Starting point was an existing Rust port.
• Agents replaced 32K lines of SIMD with safe Rust that the compiler auto-vectorizes, working through profile-guided experiments and inspection of compiler output.
• Result: 2.7× faster than the Rust port, with identical output.
• It is described as "closer to" the optimized C++, not equal to it.
• C/C++ to Rust migrations: from tens of thousands of lines (re2, libgav1) up to 800K+ lines (Fuchsia Zircon kernel). These rewrites are still going through audit, emulation testing and review before production.
• Fleet memory optimization: over 300 TiB freed so far, with an estimated 500 TiB to 1 PiB in total.
• Quantum subroutine optimization: beat the published baseline on spacetime cost (qubits × gates) by 40% in one example.

How to use it

• Pricing: $2 per million input tokens and $10 per million output tokens. Cached input is 95% off the input price. This is labelled an introductory price, so expect it to change.
• Output limit: 1M tokens, up from 64K. This is aimed at single trajectories that produce hundreds of thousands of tokens.
• The page gives no model ID, API shape or context-window size.

Limits

• Monitoring may stop runs: the deployment monitors chain-of-thought and actions for misalignment and halts execution when it judges that necessary. Agentic jobs can be stopped mid-run.
• Misuse filtering: refusals target cyber and CBRN misuse and are backed by activation monitoring. Guardrails are still being iterated with early-tester feedback, so refusal behavior at general availability is not final.
• Weak benchmark comparisons: competitor numbers are absent for nearly every benchmark. The cyber comparisons are against Google's own prior model on internal benchmarks.
• Migrations unproven in production: the large-scale Rust migrations are explicitly not yet in production.
• Government review: Google is taking part in the U.S. government's voluntary pre-release access process, which could affect timing.