Gemini 4 Argon lands on Google Cloud: hand off hours-long multi-step work and get finished output back

Enterprise users can now hand off deep, multi-step work that used to take hours or days and get back finished work they can use, not just answers and drafts. Teams previously used Gemini mainly for answering questions and drafting content, while deep multi-step work stayed with people. Google Cloud made Gemini 4 Argon available in Gemini for business on web, desktop and mobile, and to developers via API under model ID gemini-4-argon, in Google AI Studio and in Google Antigravity.

Argon is trained to keep going until a long task is done while respecting system boundaries, reporting negative findings honestly, and pausing for confirmation before irreversible steps. Such long tasks previously tended to lose context or cross boundaries partway through. Google says Argon reads long documents mixing text, tables, charts and images, pulls findings from many sources into one structured answer, and tells the user when it could not find something or finish a task.

On Google's own tools, autonomous Argon runs produced measurable engineering gains. Those optimizations previously required manual investigation of framework overhead and code performance. On the open-source Agent Development Kit, Argon ran autonomously overnight within automated testing guardrails; its optimizations are now live in google/adk-python, cutting in-memory session prep time by 40% in long sessions and speeding cold imports by 21%. Google engineering teams using Argon in Antigravity increased agentic code submissions by more than 35% while cutting code rollbacks by nearly a third.

For security and regulated work, Argon emphasizes verifiable precise output and resilience to prompt injection. Finance, legal and tax work previously demanded calculations and data extraction that are precise, repeatable and verifiable. Google says Argon leads on the Vals Index, which measures economic impact across finance, coding, legal and tax work, and on Harvey's Legal Agent Benchmark; it ties for first on CWE-Bench v1 with a top score of 68%; and on the Gray Swan Indirect Prompt Injection benchmark it cut attack success across 15 adaptive attempts to 0.7%.

This report covers Google Cloud's release description of Gemini 4 Argon, including availability channels, pricing, self-reported engineering results and partner evaluations. It is for researchers and engineers evaluating enterprise agent models, especially teams handling long documents, codebase refactors or security audits. Google's stated next steps are the model card and developer guide, plus the Security Summit for cyber defenders.

Google calls Argon its most resilient model yet against indirect prompt injections and reports cutting attack success across 15 adaptive attempts to 0.7%, but that is a result on a single benchmark. HubX reports identifying 100% of top-priority vulnerabilities with 0% fabricated vulnerabilities in its evaluations, Harvey reports substantial improvements over Gemini 3.1 Pro on its legal agent evaluations, and ServiceNow reports roughly 20% improved task success over Gemini 3.8 Flash in early access on NowAI-Bench; all three are partner self-reports. Argon has cyber mitigations blocking offensive tasks, and an unguarded version is released only to vetted trusted defenders through the Fairwind Program.