This is a series where you can "binge-read" the news you've accidentally missed, but are a little curious about, over the weekend. This time, let's check out the main news released mainly during the week of September 27th!
Google announces new frontier model "Gemini 4 Argon"
Google announced its new frontier model "Gemini 4 Argon" on September 30th. It is a model designed to sustain deep inference in complex and long-term workflows, and is said to demonstrate performance in software development, knowledge work such as legal and finance, and cybersecurity.
First, it will be gradually expanded to trusted cybersecurity personnel through the "Fairwind Program," which was also used in September's "Gemini 3.8 Flash Cyber." The company also participates in the U.S. government's voluntary framework for pre-publication model access. The introduction price is $2 for input and $10 for output per 1 million tokens, and cached input tokens will be discounted by 95% of the input price.
The limit for output tokens has been increased from the previous 64,000 tokens to 1 million tokens. It is said that hundreds of thousands of tokens can be generated in a single process.
According to Google, it scored 77.9% on DeepSWE v1.1, which measures long-term software development tasks, 51.3% (first place) on Zapier's business benchmark AutomationBench, and 91.7% on LVBench, which measures long-form video understanding. In CWE-bench v1, which evaluates the ability to fix vulnerabilities, it tied for first place with a score of 68%.
Thousands of employees are already using it internally, and the company has expanded its work to migrate C/C++ code to Rust to include Fuchsia's Zircon kernel (over 800,000 lines).
Please note that the service will be provided to some cybersecurity personnel and internal teams without any cyber-related guardrails. In preparation for general availability, the company has strengthened safety measures in four areas: abuse prevention, prompt injection countermeasures, misalignment monitoring, and system robustness.
OpenAI releases AI model “GPT-6.1 Sol”
On September 29th, OpenAI announced the AI model ``GPT-6.1 Sol'' in conjunction with the developer event ``DevDay 2026.'' It is an upgraded version of ``GPT-6 Sol'' released on September 22, and is said to have excellent performance in agent-type coding, as well as computer operations and professional work.
API fees are $2 for input and $10 for output per million tokens, the same as GPT-6 Sol. Cached inputs cost $0.10, a 95% discount off standard input rates.
The company explains that this input/output charge is one-fifth that of the top model "GPT-6 Astra", and that the performance in difficult tasks is close to that of Astra. While Astra remains the company's top model in terms of overall capabilities, GPT-6.1 Sol offers a new balance of power and cost for users who want to perform critical tasks more frequently and API users who build and operate applications at scale.
In a benchmark published by the company, DeepSWE v1.1, which evaluates development tasks on an actual code base, achieved a score equivalent to GPT-6 Astra at about one-fifth the cost. It also beat the highest score of GPT-6 Sol by 6.4 points.
In the business workflow AutomationBench, it outperformed Anthropic's "Claude Opus 5.5" by 2.2 points at about one-third the cost when the amount of thinking was set to "medium". On the other hand, GPT-6 Astra has the highest score (68.1%) in Terminal-Bench Science 0.1, which recommends Astra for the most difficult scientific research tasks.
Availability begins on the same day for ChatGPT Work and Codex, and is available for Plus, Pro, Business, Enterprise, and Edu users. However, it is not yet available in "Chat". The API provides it as "gpt-6.1-sol".
In addition, we plan to provide a high-speed version of ``GPT-6.1 Sol Ultrafast'' in the coming days, and Codex will be able to generate tokens at up to 8 times the standard speed.
