Gemini 4 Argon

Google launched Gemini 4 Argon on September 30, 2026, describing it as a new frontier model built to sustain deep reasoning across long-horizon professional workflows. Authored by Koray Kavukcuoglu of Google DeepMind, the announcement targets software engineering, legal and finance knowledge work, and cybersecurity defense. Argon features a 1 million-token limit for multi-step problem solving and is initially rolling out to vetted cyber defenders via Google's Fairwind Program under a phased safety release. Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% off input token price. Google is engaged with the U.S. government's voluntary pre-release model access process and is gathering feedback from early testers before broader release. The model is already powering internal Google workflows, with teams citing gains in specialized coding, deeper research, and writing quality across thousands of employees.

Analytics Vidhya's hands-on review from October 14, 2026 puts Gemini 4 Argon through four practical tasks covering video generation, website building, data analysis, and reconciling conflicting documents. Reviewer Vasu Deo Sankrityayan reports that Argon delivered generally useful results while fabricating at least one unsupported claim in a launch summary. The piece examines capabilities, limitations, access paths, and pricing for prospective enterprise and developer users. The review frames Argon as stronger on creative, analytical, and document-heavy tasks relative to prior Gemini releases, with a feature set aimed at long-context professional work. It catalogs practical strengths and failure modes rather than formal benchmark numbers, complementing official Google announcements with practitioner observations. Readers are guided through Argon's access options and pricing before concluding on real-world usability for serious deployments.

Google announced Gemini 4 Argon availability on Google Cloud on October 13, 2026, positioning it as the intelligence layer underpinning the new Gemini Enterprise agent for work. Written by Michael Gerstenhaber, VP of Product Management for Gemini at Google Cloud, the post says Argon was trained for depth and precision to maintain context across long, multi-step tasks with strong domain expertise in software builds, finance, legal, and cybersecurity, including tools like CodeMender. The post highlights Argon's ability to read long mixed-format documents combining text, tables, charts, and images, then produce a single grounded output, while reporting negative findings honestly and pausing before irreversible steps. Early customers described the model as strong at turning huge, unstructured inputs into cohesive outputs for multi-stage enterprise work. Google is emphasizing enterprise-ready behaviors, including honoring system boundaries and clearly flagging when information cannot be found.

Reuters reported on October 13, 2026 that Google released Gemini 4 Argon to developers and businesses worldwide, framing it as the company's flagship AI model in coding, finance, law, cybersecurity, and science. The release follows Google's decision not to ship Gemini 3.5 Pro, a DeepMind leadership shakeup that moved Demis Hassabis aside, and the departure of several Gemini leaders. Early testers include Deloitte, LexisNexis, Ambience Healthcare, and Isomorphic Labs. API launch pricing was set at $2 per million input tokens and $10 per million output tokens, set to double after November 30. Teams using Argon in Google's Antigravity coding tool increased agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. Google is also expanding its Fairwind Program to hundreds of additional vetted cybersecurity partners who receive a version of Argon without standard cyber guardrails.

Vals AI lists Gemini 4 Argon with a release date of September 30, 2026 and ranks it first of 48 models on its Vals Index with a 68.90% accuracy score. The model posts category-leading scores including first on Finance Agent v2 at 65.40%, second on Code Migration at 68.17%, and third on CyberBench v1.1 at 77.86%. Other ranked results include Legal Research Bench, MedCode, EMB, and ProofBench v1.1, giving a broad technical benchmark profile. Reported configuration includes a 1 million-token context window, 262k max output tokens, and token costs of $4.00 input and $20.00 output per million on Vals AI's measurement basis. Gemini 4 Argon also shows top-half placements on the Vals RSI Index, CUA-bench, and Time Horizon Index KSP, alongside lower-cost-per-test relative to its accuracy tier. These independent benchmark numbers offer a third-party view of Argon's relative standing among recent frontier models.

Google Cloud's developer guide introduces Gemini 4 Argon as the first model in the Gemini 4 family and Google's frontier model for deep work in coding, finance, legal, cybersecurity, and science. The guide documents API conventions, a migration path from Gemini 3.1 Pro to 4 Argon, and quickstart instructions for Agent Studio integration. It is positioned as the central reference for developers building on the Gemini Enterprise Agent Platform. Key technical changes include a unified 1,048,576-token limit covering both input and output tokens, long decodes up to 1,048,576 tokens beyond the 65,536-token per-request output cap, configurable LOW, MEDIUM, and HIGH thinking levels, and implicit caching with durable caching. The Google Gen AI SDK handles continuation tokens automatically across GenerateContent requests, or developers can manage continuation in their own loop. Mandatory API rules, behavioral conventions, and known issues are catalogued for teams adopting 4 Argon.

The Gemini 4 Argon model card published by Google DeepMind on October 13, 2026 documents the model as a new large-scale sparse mixture-of-experts transformer with native multimodal support for text, vision, and audio inputs. It is explicitly described as not a modification or fine-tune of a prior model, marking Gemini 4 as the next generation in the Gemini series. Argon targets complex workflows in real-world software engineering, enterprise knowledge work, and cybersecurity defense. The model accepts text strings, images, audio, and video files within a combined 1 million-token input-output context window. Pre-training drew on publicly available web documents, licensed data, and Google user data across modalities, while post-training added instruction tuning, reinforcement learning, human-preference data, and multi-step reasoning, problem-solving, and theorem-proving samples. Model cards document known limitations, mitigation approaches, and safety performance and will be updated with new evaluations.