On June 9, Anthropic officially launched two new models: Claude Fable 5 for public use and Mythos 5, limited to trusted users. Both models share the same core and achieved 80.3% on SWE-bench Pro (GPT-5.5 scored 58.6%) and 88% on Terminal-Bench 2.1. Stripe used Fable 5 to complete a 50 million-line Ruby code migration in one day—equivalent to over two months of team effort. Internally, Mythos 5 improved protein design efficiency by approximately tenfold. API pricing is set at $10 per million input tokens and $50 per million output tokens, more than 50% lower than the preview version. Fable 5 automatically downgrades high-risk requests in cybersecurity and biochemistry to Opus 4.8, with a false trigger rate below 5%. Mythos 5 is available only to select institutions such as Project Glasswing.

On June 9, Cohere released North Mini Code 1.0, its first open-source agent-style programming model, using a MoE architecture with 30 billion total parameters and 3 billion active. It can run on a single H100 GPU (FP8 precision). The model supports a 256K context window and up to 64K output length, designed specifically for agent-based software engineering, enabling sub-agent coordination, architecture mapping, code review, and terminal tasks. It scored 33.4 on the Artificial Analysis Coding Index, with 2.8x higher throughput and 30% lower latency than Mistral Devstral Small 2. Released under the Apache 2.0 license on Hugging Face, it allows enterprises full freedom to modify and commercialize. This launch comes less than three weeks after its previous model Command A+, signaling Cohere’s accelerated iteration pace.

On June 9, Broadcom, Apollo, and Blackstone Group announced the formation of the AI XPV strategic platform, aiming to drive over 20 gigawatts of global AI compute deployment by 2028 through customized XPU and networking solutions. The platform's initial phase, led by Apollo and co-funded by Blackstone, commits $35 billion to support Anthropic’s previously announced plan to expand its compute infrastructure beyond 1 gigawatt, with deployments expected from mid-2026 on Fluidstack-based sites. Additionally, Bloomberg reported that Google is providing financial backing for Anthropic’s $35 billion data center lease, deepening their financial ties beyond prior equity investments and raising concerns about circular transactions and systemic risks in the AI industry.

On June 9, Google introduced Gemini 3.5 Live Translate, an audio model enabling near real-time speech-to-speech translation across over 70 languages. It automatically detects language while preserving speaker tone, rhythm, and pitch, with latency of just a few seconds. The model is now publicly available via the Gemini Live API and Google AI Studio for developers, accessible in private preview for enterprise users in Google Meet, and usable by general users through the Android and iOS Google Translate apps. Android adds a 「Listen Mode」 allowing users to hear translations directly through the speaker without headphones. All generated audio includes SynthID digital watermarking to prevent misuse. Partner Grab is testing it for real-time driver-passenger call translation, covering over 10 million calls monthly.

On June 9, Apostol Vassilev, senior scientist at the National Institute of Standards and Technology (NIST), published a mathematical proof in 「IEEE Security and Privacy」 drawing from Kurt Gödel’s 1931 incompleteness theorems. The proof demonstrates that no finite set of fixed safety rules can universally make AI systems robust against adaptive adversarial prompts. This implies AI systems can never be fully immune to 「jailbreak」 attacks, as attackers will always find ways to circumvent mechanisms. Vassilev proposes shifting AI security from a 「deploy once, protect forever」 model to one of 「continuous monitoring and updating,」 employing red teaming, dynamic defense updates, and rapid recovery capabilities. The goal is to raise attack costs beyond the resources of potential adversaries, achieving economic deterrence.

On June 9, Glean announced support for NVIDIA Nemotron 3 Ultra on its enterprise AI platform, bringing its total model offerings to over 30. Nemotron 3 Ultra delivers 91% of the capability of cutting-edge LLMs while maintaining the cost advantages of open models. Glean’s Waldo search model has been optimized using NVIDIA Nemotron 3 Nano, reducing latency by 50% and token usage by 25%. On the same day, Sedai launched the world’s first platform for autonomous AI agent optimization, AI Agent Optimization, which uses intelligent routing to automatically select among major LLMs including OpenAI, Anthropic, VertexAI, and Bedrock. It has already been adopted by enterprises such as GSK and KnowBe4. This reflects a broader shift in enterprise AI from 「one-size-fits-all」 models to on-demand selection to manage rising operational costs of generative AI.

NVIDIA CEO Jensen Huang’s recent Asia trip secured SK Hynix’s supply of HBM4 and HBM5 memory modules for the upcoming Vera Rubin platform, strengthening the next-generation AI hardware supply chain. Meanwhile, SK Telecom, Naver, Hyundai, and LG are increasing AI investments, reinforcing NVIDIA’s ecosystem dominance in Asia. However, due to U.S. export controls, NVIDIA expects zero revenue from China starting in Q2 of FY2027, a market that once accounted for nearly a quarter of its data center sales. Its market share in China has dropped from 95% in 2024 to 55% in 2025, with Huawei capturing 20% as a local alternative. Analysts remain positive about its 「sovereign AI」 strategy, projecting around $30 billion in revenue from such projects in FY2026,约占 total revenue by 14%.

On June 9, Brave security researchers released a report showing that indirect prompt injection is a widespread security vulnerability affecting all large language model agents, regardless of whether they are deployed in the cloud or locally. The study demonstrated through real-world cases that Mozilla Tabstack (cloud-deployed) and Cotypist (macOS local deployment) could both be manipulated by hidden instructions embedded in malicious web pages or local documents, leading to data leaks and content tampering. The root cause lies in current LLM architectures’ inability to reliably distinguish between developer instructions and external data. A Futurum Group survey found that 53% of enterprises cite privacy and security as primary barriers to adopting generative AI. The report debunks the myth of 「local AI being safer」 and emphasizes that deployment mode alone cannot resolve this structural risk.

Beacon, an 「AI-native」 holding company headquartered in Toronto and San Francisco, announced a $225 million Series C round led by General Catalyst and HarbourVest, with participation from Lightspeed, bringing its total funding to over $500 million in under two years. Beacon employs a unique 「anti-private equity」 model, acquiring small profitable software companies with annual recurring revenue under $20 million, particularly in underserved verticals like youth sports leagues and campgrounds. Its internal 「acceleration team」 rebuilds these businesses on a shared AI-native platform, automating back-office functions such as accounting and payroll. Over the past year, this approach has driven portfolio EBITDA growth of over 50%. Founders argue that AI has drastically reduced coding costs, creating a historic opportunity to modernize industries that account for over 55% of U.S. GDP but have long been overlooked.

On June 9, NVIDIA unveiled Cosmos 3, touted as the world’s first fully open AI omnimodel, designed specifically for physical AI applications. Built on a hybrid Transformer architecture, it integrates visual reasoning, world generation, and action prediction, capable of processing text, images, video, ambient sound, and motion, with high-precision physics simulation. Cosmos 3 enables robots, autonomous vehicles, and visual agents to generalize in real-world environments with less training data, suitable for tasks like grasping and dexterous manipulation. NVIDIA also announced the formation of the Cosmos Consortium, with founding members including Agile Robots, Doosan Robotics, LG, Samsung, Skild AI, and Li Auto. Deloitte forecasts that cumulative global industrial robot installations could reach 5.5 million units by 2026.