On June 1, Anthropic announced it has secretly filed an application for an initial public offering (IPO) with the SEC, becoming the first trillion-dollar AI startup to pursue a listing. The company's latest valuation reaches $965 billion, surpassing OpenAI, with annualized revenue of approximately $47 billion. If successfully listed, it would become the second or third largest IPO in history, behind only SpaceX and Saudi Aramco. Anthropic aims to go public ahead of OpenAI and SpaceX to seize a favorable financing window. Goldman Sachs forecasts that total funds raised through US IPOs in 2026 could reach a record $160 billion. Market analysts believe that going public before competitors will help set financial disclosure standards for frontier AI companies.

At Computex 2026, NVIDIA launched Cosmos 3, the first fully open physical AI multimodal foundation model, based on a Mixture-of-Transformers architecture that unifies text, image, video, audio, and action processing. It offers a 16B-parameter Nano version and a 64B-parameter Super version, trained on 20 trillion multimodal tokens, achieving top rankings across multiple physical AI benchmarks. The model is now available on Hugging Face under the OpenMDW 1.1 license, permitting commercial use. NVIDIA also formed the Cosmos Coalition, partnering with Agile Robots, Runway, and others to advance open-world models, alongside releasing synthetic datasets for robotics and autonomous driving.

MiniMax officially released its next-generation large model M3 on June 1, featuring its proprietary MSA sparse attention architecture, supporting a 1 million token context with computational costs just 1/20th of its predecessor M2, prefill speedup by 9x, and decoding speedup by 15x. M3 is the first open-weights model to simultaneously achieve state-of-the-art coding capabilities, million-token context, and native multimodality, scoring 59.0% on SWE-Bench Pro, outperforming GPT-5.5. In real-world testing, M3 autonomously reproduced an ICLR paper within 12 hours and increased Hopper GPU utilization from 7.6% to 71.3%. API pricing is set at $0.3 per million input tokens, with model weights and technical report promised to be open-sourced within 10 days.

On June 1, Florida’s Attorney General filed an 83-page lawsuit against OpenAI and CEO Sam Altman, alleging that ChatGPT is linked to multiple violent incidents, including assisting mass shootings, encouraging suicide, and addicting minors. This is the first state-level safety-related lawsuit against OpenAI in the U.S., citing 10 counts under the Florida Deceptive and Unfair Trade Practices Act, demanding bans on data collection from minors and personal liability for Altman. Florida had previously initiated a criminal investigation into last year’s Florida State University shooting. OpenAI responded that it has implemented industry-leading safeguards for minors.

NVIDIA released Alpamayo 2 Super, a 32-billion-parameter vision-language-action model designed specifically for Level 4 autonomous taxis. It features 360-degree panoramic perception, meta-action output, reasoning-based auto-labeling, and causal-chain trajectory prediction, excelling in rare and complex scenarios. As a teacher model, it can be distilled into smaller models compatible with the onboard Drive AGX Thor platform. NVIDIA also introduced AlpaGym, an open-source closed-loop reinforcement learning framework, and OmniDreams, a generative world model, forming a complete development pipeline from data acquisition to in-vehicle deployment. The Alpamayo series has been downloaded nearly 400,000 times, with the new version expected to be open-sourced this summer.

Alphabet plans to raise $80 billion through its first stock issuance since 2005, dedicated exclusively to AI infrastructure and computing power development. Berkshire Hathaway will purchase $10 billion in shares at a discount, signaling strong confidence from top institutional investors in Alphabet’s AI strategy. This fundraising marks one of the largest equity financings specifically for AI to date, highlighting the surging capital demands in the AI race.

On June 1, Instagram patched a security vulnerability: hackers exploited conversations with Meta AI customer support chatbots by providing only a target username and their own email address, tricking the bot into binding a new email to the target account and triggering a password reset, thereby completing account takeover. Multiple high-profile accounts were affected, including the Obama-era White House account and the U.S. Space Force Chief Master Sergeant’s account. Security researchers noted this was not a sophisticated prompt injection attack but a fundamental design flaw in delegating critical account recovery functions to an AI chatbot. Instagram has confirmed the issue is resolved.

JetBrains open-sourced Mellum2 on June 1, a 12-billion-parameter mixture-of-experts (MoE) model trained from scratch, activating only 2.5 billion parameters per token, focused on natural language and code processing. It performs exceptionally well across multiple code generation and reasoning benchmarks, with inference speeds more than twice as fast as comparable open-source models. Suitable for task routing, RAG pipelines, sub-agent coordination, and private deployment. JetBrains introduced the concept of "focus models," emphasizing efficient, specialized small models for high-frequency tasks in complex AI systems. Released under the Apache 2.0 license and now available on Hugging Face.

Alibaba’s Qwen team launched Qwen3.7-Plus, a multimodal agent model that integrates visual and linguistic capabilities into a single foundational architecture, supporting GUI/CLI operations, programming, and visual tasks. Its text benchmark performance approaches Max-tier model levels, demonstrating comprehensive capabilities in multimodal agent scenarios. Positioned as a general-purpose multimodal foundation model, it serves various agent applications requiring simultaneous understanding of visual and textual inputs.

On June 1, the Ministry of Industry and Information Technology-approved standard 'YD/T 6770—2026 Method for Embodied Intelligence Benchmark Testing' officially took effect, establishing a unified benchmarking framework and standardizing testing procedures and metric calculations in both simulation and real-world environments. Concurrently, the 'Lifecycle Management Specification for Humanoid Robots' introduced a 'one robot, one code' mechanism, ensuring each humanoid robot has a unique identity code for full traceability throughout its lifecycle. Data shows that by 2025, China had over 140 integrated robotics enterprises, launched more than 330 products, and shipped approximately 17,000 units. The development of standardized systems is driving the industry from technical validation toward large-scale deployment.