Today's AI news: Google DeepMind launches Gemini 4 Argon, OpenAI DevDay reveals new tools, Cogentic proves open research problems, and Meta's tax story breaks.
Google DeepMind Unveils Gemini 4 Argon
Google DeepMind has officially announced Gemini 4 Argon, which it describes as the beginning of a new era of frontier intelligence. The model was the subject of detailed coverage from Latent Space, which notes it is positioned as DeepMind's answer to Astra and Fable — and comes with a striking 1 million token output capacity. At launch, however, access is heavily restricted: only government users and trusted cyber defenders enrolled in the Fairwind Program can use the model. There is no word yet on a broader public rollout. The announcement signals a significant step up in DeepMind's frontier model ambitions.
OpenAI DevDay 2026: Computer Use and a New Jev Competitor
OpenAI held its DevDay event, and the Latent Space team published detailed coverage of the conference, interviewing the leaders of OpenAI's Computer Use Agent (CUA) team and API platform. A notable headline from the event is that OpenAI shipped a competitor to Jev — a computer use product — in just one week. The coverage also digs into the broader debate around computer use as an AI capability, pushing back on commentary from Dwarkesh Patel on the subject. DevDay underscores OpenAI's continued push to make AI agents practical tools for developers and enterprises alike.
OpenAI on AI and the "Eternal Complement" — Execution Over Ideas
OpenAI published a thought piece titled The Eternal Complement, arguing that advanced AI may matter most for the routine execution work behind breakthrough ideas — not just for generating the ideas themselves. The piece makes the case that AI's greatest economic impact could be in dramatically accelerating the unglamorous but essential work of turning concepts into reality, and that execution, not ideation, could define the shape of the next economy and the pace of scientific progress.
Cogentic: Multi-Agent System Solves Open Research Problems in Maths and Economics
A research team introduced Cogentic, a multi-agent orchestration framework designed for automated proof discovery on genuinely open research problems. The system uses an iterative prove-and-verify loop: an orchestrator directs a population of independent provers across competing proof directions, with adversarial verification components checking their outputs, and confirmed intermediate results are logged in a persistent verified ledger that later rounds can build upon. Using Gemini as its base model, Cogentic has produced novel, independently verified results on five open problems spanning online learning, auction theory, and mechanism design. Each result has been confirmed by domain experts and is being written up in companion papers. The system represents one of the more striking demonstrations of AI tackling real, unsolved mathematical and theoretical computer science questions.
Meta Using AI Data Centres to Avoid Billions in Federal Taxes
A New York Times report — which ranked among the most-discussed items on Hacker News today — details how Meta has structured its AI data centre investments in ways that allow the company to avoid billions of dollars in federal taxes. The story adds a significant policy and regulatory dimension to the ongoing expansion of AI infrastructure, raising questions about how governments will respond to the fiscal implications of the AI infrastructure boom.
Allen AI Releases OLMo-Core 3: Open Training Infrastructure for Large MoEs
The Allen Institute for AI, via Hugging Face, announced OLMo-Core 3 — an open, scalable training infrastructure toolkit specifically designed for large Mixture-of-Experts (MoE) models. The release continues Allen AI's commitment to open-source AI development and is aimed at researchers and engineers who need robust, reproducible infrastructure for training frontier-scale models without relying on proprietary systems. The toolkit is available on Hugging Face.
Frequently Asked Questions
What is Gemini 4 Argon and who can use it?
Gemini 4 Argon is Google DeepMind's latest frontier AI model, announced on 1 October 2026. It features a 1 million token output capacity. At launch, access is restricted to government users and trusted cyber defenders in the Fairwind Program, with no public rollout announced yet.
What did OpenAI announce at DevDay 2026?
At DevDay 2026, OpenAI showcased advances in its Computer Use Agent capabilities and revealed it had shipped a competitor to Jev — a computer use product — in just one week. The event featured discussions with the leaders of OpenAI's CUA team and API platform.
What is Cogentic and what has it achieved?
Cogentic is a multi-agent AI framework for automated proof discovery on open research problems. Using Gemini as its base model, it has produced novel, independently verified results on five open problems in online learning, auction theory, and mechanism design.
How is Meta using AI data centres to reduce its tax bill?
According to a New York Times report covered on Hacker News, Meta has structured its AI data centre investments in a way that allows the company to avoid billions of dollars in US federal taxes, raising significant policy questions about the fiscal impact of the AI infrastructure expansion.