
Tech • AI • Robotics • Game
Google rolled out Gemini 4 Neon across Gemini and AI Studio as a faster, cheaper reasoning model aimed at coding, office work, and agent workloads. Early benchmark reports placed it near the frontier, including third on the Artificial Analysis Intelligence Index and strong showings on enterprise and agent tests. Pricing was cited at $3 per million input tokens and $15 per million output tokens, with major speed gains versus Gemini 4 Carbon. The launch positions Neon squarely between Hydrogen and Carbon, but in some hands-on tests it matched or beat the larger model.
OpenAI introduced GPT-6.2 Astra for professional reasoning, science, coding, and long-horizon tasks, alongside lower-cost GPT-6.2 Soul for general use and agents. Astra was reported at $8 per million input tokens and $40 per million output tokens, while Soul held at $2 in and $10 out. OpenAI highlighted gains on Agents Last Exam, where Astra scored 78.3%, and emphasized lower task costs through better planning and reduced token use. But early field reports were mixed, with some developers praising efficiency while others questioned real-world reliability outside benchmark conditions.
At Microsoft Ignite in San Francisco, Microsoft recast Copilot from an app feature into the primary interface for workplace computing. The centerpiece was Copilot IQ, built on Microsoft Fabric, designed to turn fragmented enterprise data into persistent operational context for agents and employees. Satya Nadella framed the shift as moving from isolated assistants in Word, Excel, and Teams to a governed AI layer spanning records, rules, meetings, suppliers, and customer data. The pitch squarely targeted business hesitation over security, compliance, and shadow IT by pairing autonomy with auditable controls.
A previously unseen Anthropic model labeled Silicon 2 surfaced in tooling, fueling speculation that a stronger Fable checkpoint is close. Placeholder pricing mirrored Fable 5.1 at $10 per million input tokens and $50 per million output tokens, while separate benchmark traces showed unidentified Anthropic checkpoints climbing toward the frontier. At the same time, Anthropic entered public markets with an indicated range of $70 to $78 a share, becoming the first major frontier lab to list. The combination of model chatter and an IPO sharpened scrutiny of Dario Amodei’s safety-first stance and the company’s competitive timing.
Nvidia faced a deepening governance and political storm after the forced removal of Jensen Huang from board leadership. Reports tied the uproar to longstanding concerns that billions of dollars in Nvidia chips may have been routed through a tiny Singapore company despite U.S. export restrictions touching China. Authorities were said to have known of anomalies since 2024, raising questions about the delay between awareness and public escalation. The episode has rattled confidence in the company just as the AI infrastructure race makes Nvidia central to global compute supply.
Executives from Blackstone, Micron, and Morgan Stanley argued that AI has shifted from software excitement to a capital-heavy buildout of data centers, memory, chips, and power. Blackstone President John Gray said the cycle now resembles an infrastructure era requiring trillions of dollars in real-world investment, not a lightweight internet-style scaling story. Demand for compute and electricity was described as outrunning supply, with constraints spanning permits, chip output, and grid capacity. The broader takeaway is that a large share of AI economics may accrue first to infrastructure owners rather than model makers.
Google also launched Gemini Task, a free personal AI agent in the United States, Canada, Australia, India, Japan, and South Korea. The product is built for delegation rather than chat, handling inbox triage, scheduling, shopping, reports, and direct messaging through native iOS and Android text apps. Its tight integration with Google services gives it an advantage in email, calendars, and documents, while the SMS-style thread aims to make the agent feel more persistent and assistant-like. The move adds a consumer-agent flank to Google's simultaneous push with Gemini 4 Neon.
GPT-6.2 Astra was presented less as a benchmark winner than as a system for real work, from guided visual exploration to partially automated administrative tasks. A live multimodal sequence moved through Paris landmarks including Trocadéro, the Eiffel Tower, Invalides, the Louvre, and Notre-Dame, while changing viewpoints, switching to aerial perspective, and altering lighting to 10 PM. Inside a modeled Louvre, it identified works such as the Mona Lisa, The Wedding Feast at Cana, and Liberty Leading the People, while tracking viewer orientation. The broader claim was that multimodal control and workflow execution, not raw leaderboard gains, will decide enterprise value.