[ITmedia PC USER] General availability of “Gemini 4 Argon” begins / Meta announces “Muse Wonder” series

This is a series where you can "binge-read" news on weekends that you may have missed but are a little curious about. This time, let's check out the main news released mainly during the week of October 11th!

General availability of “Gemini 4 Argon” begins

On October 13, Google began general availability of its AI model "Gemini 4 Argon." When it was announced on September 30th, it was only available in advance for cybersecurity personnel, and there was no indication of when it would be available to the general public.

For developers and businesses, Gemini API can be used with paid plans from AI Studio etc. For general users, Google AI Ultra and Google AI Pro subscribers can use the Gemini app and Google Search's "AI mode." Google's agent-based development environment, Google Antigravity, is available to Ultra subscribers.

API pricing is $2 for input and $10 for output per 1 million tokens during the implementation period until November 30th, with a 95% discount on cached input tokens. After the introductory period, the input will be $4 and the output will be $20. Gemini API can handle up to 1 million tokens in total for input and output, and has also prepared a new "continuation protocol" for long-term inference.

According to the company, it ranked first in the Vals Index, which measures performance in economic fields such as finance, coding, legal affairs, and taxation, and scored 87% in the physical and spatial reasoning ERQA, outperforming other Frontier models. It also ranked first in Gray Swan's benchmark for resistance to indirect prompt injection, with an attack success rate of 0.7% over 15 adaptive attacks. Internally, the company says agents are submitting more than 35% more code through Antigravity, and code rollbacks have decreased by nearly a third.

In addition, applications are being accepted for a program that will provide 10,000 scientists with the highest level of use of Antigravity for free for six months. Applicants will be selected based on screening and availability. The Fairwind Program for cybersecurity professionals will also add hundreds of participating partners.

In addition, Google plans to offer other models of the Gemini 4 family in the future.

Meta announces “Muse Wonder” series

On October 12th, Meta announced the AI model ``Muse Wonder'' and its higher-level version ``Muse Wonder Full.'' From the same day, Muse Wonder became available for developers with the Meta Model API and the coding tool Muse Code, and Muse Wonder Full was made available to Muse Code subscribers. It will soon be installed in the personal AI agent "Muse" announced in September, and will be rolled out to subscription users first.

The company is launching two capabilities for individual and business agent work. One is Active Memory, which maintains important context about the user across conversations and recalls task-relevant information when needed. Active Focus is the ability to continue working toward a difficult goal by trying different approaches and overcoming obstacles. When necessary, users will be asked for confirmation before particularly important operations.

The company's gigawatt-class data center ``Prometheus'' was used for learning. By adopting a base model that is larger and more powerful than the previous Muse Spark, it is said that it will demonstrate better performance than expected from the trends of Muse Spark 1.3, even on unknown tasks.

According to the comparison table published by the company (each model has the maximum inference setting), Muse Wonder Full scored 13.1 on the legal agent's "Harvey LAB-AA," 79.0 on the "DeepSWE 1.1" that measures long-term coding, and 45.4 on the business workflow "AutomationBench-AA." It outperforms the highest scores of the compared Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5 (8.6, 74.2, and 41.6 for each benchmark, in order). On the other hand, Opus 5.5 is higher in Terminal-Bench 4.0-AA, ProgramBench, Artificial Analysis Index, etc.

In terms of safety, the company says it has taken the strongest measures to date to protect against misuse of chemical, biological, radioactive, and nuclear (CBRN) and cyber attacks. Cyber ​​capabilities for defensive purposes remain available. A report summarizing additional safety evaluations will be published soon.