Google DeepMind released Gemini 4 Argon globally on October 13, 2026, making its newest multimodal reasoning model available to developers, enterprises and consumers through multiple platforms including the Gemini API, AI Studio, Google Antigravity, the Gemini app and AI Mode in Search, according to a blog post by Koray Kavukcuoglu published October 13, 2026. The model had been announced on September 30, 2026, with initial access limited to trusted cybersecurity defenders.
The launch targets users working on complex, multi-step tasks across coding, finance, legal, cybersecurity and scientific domains. Google reported that Argon ranks first on the Vals Index, which measures model performance across critical economic sectors including finance, coding, legal and tax work, as of the October 13 announcement.
Pricing and Technical Specifications
Gemini 4 Argon launched with introductory API pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95%, according to the October 13 blog post. Google Cloud documentation dated October 13, 2026, states the introductory rate applies through November 30, 2026, after which standard pricing will be $4 per million input tokens and $20 per million output tokens.
The model supports a unified 1,048,576-token context window shared by input and output, according to the October 13 announcement. Google’s API documentation specifies that the default synchronous output is capped at 65,536 tokens, with a continuation protocol enabling longer generation within the shared million-token budget. The model accepts text, images, video, audio and PDFs as input but produces only text output; it does not support image generation, audio generation or the Live API, according to Google’s developer documentation dated October 13, 2026.
Internal Engineering Deployments at Google
Google reported internal use cases demonstrating the model’s application to production engineering problems. The Chrome team used Argon in Google Antigravity to separate core webpage handling code from a legacy folder within Chromium’s 40-million-line codebase, completing an architectural refactor across hundreds of files that the team said would have taken months, according to the October 13 blog post. The changes will allow live webpages to move between tabs, side panels and floating windows without reloading.
Google’s Agent Development Kit team reported that Argon autonomously optimized their Python codebase overnight, cutting in-memory session preparation time by 40% in long sessions and speeding up cold imports by 21%, with changes now live in the google/adk-python repository, according to the October 13 announcement and a developers blog post published the same day. Google researchers also reported using Argon to accelerate the vLLM inference engine for Gemma 4 E2B by 2.85 times on an A100 GPU, increasing throughput from approximately 37,900 to over 108,000 tokens per second, according to the October 13 blog post.
Enterprise and Research Applications
Wiz, a cybersecurity firm participating in Google’s Fairwind Program, used early access to Argon to identify vulnerabilities in government and healthcare infrastructure, according to the October 13 announcement. In collaboration with an unnamed federal agency, Argon helped identify a gap that risked exposing sensitive data and government infrastructure. In healthcare, the model helped discover a public file download bypass that exposed a master key and granted administrator access, potentially affecting hundreds of thousands of private patient records, according to the blog post.
In scientific research, Google reported that Argon helped accelerate quantum computing work by translating an internal version of AlphaQubit, Google’s machine-learning quantum error correction decoder, into a single optimized hardware kernel, nearly doubling inference throughput. A team at Google working with Professor Matt Landreman of the University of Maryland discovered a broader class of counterexamples to Grad’s conjecture, a 1967 hypothesis about stellarator magnetic fields, using Argon, according to an arXiv preprint dated October 2026 cited in the announcement.
Gamow Labs, a rare disease research organization, used Argon to analyze de-identified newborn genomes and identified a chromosomal rearrangement that routine software had missed for years, according to the October 13 blog post. On Gamow Labs’ private evaluations derived from clinical and lab data, Argon scored 57% on RareBench for rare-disease diagnosis assistance and 52.2% on LabBench for wet-lab reasoning, according to a Gamow Labs blog post referenced in the announcement.
Robotics and Physical AI
Google demonstrated Argon’s application to robotics through a “Video-to-Sim-to-Real” pipeline in which the model watched a video of a human playing cornhole, calculated required joint movements, and autonomously built simulation code. When transferred to Apptronik’s Apollo humanoid robot, the robot completed seven consecutive successful throws in a real-world trial, according to the October 13 announcement and a developers blog post published the same day. Argon scored 87% on the Embodied Reasoning (ERQA) benchmark, which tests a model’s ability to understand and solve physical, spatial problems in 3D environments, according to the October 13 blog post.
Safety and Access Programs
Google stated that Argon includes defenses against misuse, jailbreak attempts and prompt injection. External biological red teaming could unlock less than 1% of biological sub-tasks tested, and violations on an adversarial set designed to stress test the model reduced nearly twofold compared to previous Gemini models, according to the October 13 announcement and the Gemini 4 Argon Model Card published by Google DeepMind. On coding benchmarks, shortcut attempts—instances where the model might take unauthorized actions to complete a task—dropped to 2%, and Argon achieved a 0.7% attack success rate across 15 adaptive attempts on the Gray Swan Indirect Prompt Injection benchmark, according to the October 13 blog post.
Google announced it will offer 10,000 scientists the highest usage tier in Google Antigravity free of charge for six months, onboarded on a rolling basis and subject to eligibility, capacity and selection. Applications opened October 13, 2026. The company is also expanding its Fairwind Program, which provides access to Argon without cyber guardrails for trusted cybersecurity defenders, to hundreds of additional partners, according to the October 13 announcement.
Limitations and Model Context
The Gemini 4 Argon Model Card published by Google DeepMind acknowledges hallucinations, possible slowness or timeouts, and a March 2026 knowledge cutoff. The model card describes Argon as a sparse mixture-of-experts transformer and states it is not a modification or fine-tune of a previous model. Google’s October 2026 evaluations show stronger results on some knowledge-work, long-context and multimodal benchmarks compared with competing AI models, but weaker results on some software-engineering and terminal benchmarks, according to the model card dated October 13, 2026. These are vendor-reported evaluations.
Source: blog.google.
Be the first to comment