Alphabet’s Google DeepMind launches Gemini 4 Argon with new pricing, multimodal features, and robotics claims

Google is giving developers, enterprises, researchers, and consumers access to its newest frontier AI model, Gemini 4 Argon, positioning it as a system built for “deep work” on complex tasks that require sustained reasoning. In a product and research blog post published October 13, Google DeepMind and Google’s chief AI architect said Argon is intended to improve performance on multi-step, real-world workflows, with emphasis on depth, precision, and clearer responses.

Google says Argon is strong across domains including coding, finance, legal work, cybersecurity, and scientific tasks, and it claims the model ranks first on the Vals Index. Google described the Vals Index as a measure of model performance across economic sectors such as finance, coding, legal, and tax-related work, and said Argon placed at the top of that benchmark.

On pricing, Google said Argon launches today at an introductory rate of $2 per million input tokens and $10 per million output tokens. It also said cached input tokens are priced at 95% off, and that after the introductory period ends on November 30, 2026, the price will move to $4 per 1 million input tokens and $20 per 1 million output tokens. Google added that the new rates apply to an Argon model intended for public access.

Google’s post also tied Argon to internal and partner results. It said teams at Google have used Argon on “hardest engineering problems” and that, over the past few weeks, software engineers were shipping higher-quality code faster. Google cited Antigravity as one of the uses and claimed Argon helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third.

For enterprise and startup users, Google said Argon is already generating “strong wins” through an early access program. It described Argon as combining frontier reasoning with deep visual perception, expanded domain knowledge, and improved writing capability, and it said Argon follows style instructions better than prior Gemini models. Google also highlighted multimodal strengths aimed at professional workflows, including deep multimodal document comprehension, longitudinal data analysis, and interactive analytical sandboxes.

In security and research, Google said researchers and security specialists are using Argon to test hypotheses in fields such as cyber defense, science, mathematics, and robotics. It described use cases including autonomously auditing critical codebases and formal proofs, and catching subtle zero-day vulnerabilities and logic flaws. Google also said it is collaborating with third parties on capability measurement and safeguards, and directed readers to a Gemini 4 Argon Model Card for additional details.

Google’s robotics claims were more specific and were paired with a named partner. It said the company tested whether Argon could learn an out-of-distribution physical skill by watching, describing an internal “video-to-sim-to-real” pipeline. In the example, Google said Argon watched a raw video of a human playing cornhole, mapped the environment’s spatial coordinates, calculated the required whole-body joint movements, and then built simulation code. Google said the approach was transferred onto Apptronik’s Apollo humanoid and that the robot sank seven consecutive shots in a real-world trial.

The company is also expanding parts of its rollout beyond immediate public availability. Google said it is beginning expansion to hundreds of additional Fairwind Program partners, and pointed to Wiz as an early user through Wiz’s “Scan for Good” initiative, which it described as protecting critical infrastructure for the public good and for free. Google also said it will offer 10,000 scientists the highest usage tier in Google Antigravity free of charge for six months, with onboarding on a rolling basis subject to eligibility, capacity, and selection.

Google’s post framed Argon as part of a broader “Gemini 4 family” roadmap and emphasized implementation details for developers. It said Argon supports a “1M unified input-output token limit” via the Gemini API, and referenced new developer-facing documentation including a continuation protocol for handling longer work. The post did not provide independent benchmarking methodology, confidence intervals, or additional numeric breakdowns behind its claims, and the examples it cited were largely described through Google’s own and partner testing.

For what to watch next, the most immediate indicates will be how widely Argon is adopted in production settings and whether third-party evaluations and Model Card documentation corroborate the performance and safety assertions. The rollout timing, the end date for introductory pricing, and ongoing partner expansion through the Fairwind Program could also shape early usage patterns as developers and enterprises decide how to structure longer-running, multimodal workflows on the Gemini API.