On June 9, Anthropic launched its first publicly available Mythos-class model, Fable 5, along with a restricted version of Mythos 5, both sharing the same weights. Fable 5 achieved 80.3% on SWE-Bench Pro, significantly surpassing Opus 4.8's 69.2% and GPT-5.5's 58.6%, and ranked first on the 「Artificial Analysis Intelligence Index」 with 64.9 points. The model supports a 1 million token context window and is priced at $10 per million input tokens and $50 for output, half the cost of the Mythos preview version. It includes built-in classifiers for cybersecurity, biochemistry, and reasoning distillation; when triggered, it falls back to Opus 4.8, with a trigger rate of approximately 5%. Pro/Max users can access it for free until June 22. Mythos 5 is only available to around 200 trusted institutions under Project Glasswing.

According to TechCrunch, multiple cybersecurity researchers have expressed dissatisfaction with the safety restrictions in Anthropic's new model, Fable 5. The model uses keyword-based detection to flag requests related to cybersecurity and biology, frequently blocking harmless tasks such as reading blogs, code reviews, or writing 「secure code」, which then triggers a fallback to Opus 4.8. IBM X-Force researcher Valentina Palmiotti and security expert Matt Suiche noted that the mechanism is overly broad. Anthropic offers a 「Cybersecurity Verification Program」, granting approved professionals less restrictive access, similar to OpenAI's Trusted Access Program. Experts acknowledge that initial over-restriction is understandable and expect gradual optimization moving forward.

On June 10, Google released DiffusionGemma, an experimental open-source model based on Gemma 4 architecture and Gemini Diffusion technology, featuring a 26-billion-parameter Mixture-of-Experts (MoE) structure with only 3.8 billion parameters activated during inference. The model generates up to 256-token text blocks in parallel via text diffusion, rather than autoregressive token-by-token generation, achieving speeds up to four times faster: over 1,000 tokens per second on NVIDIA H100 and up to 2,000 tokens/sec on DGX Station. It can run on an RTX 5090 with 18GB VRAM after quantization. With bidirectional attention and iterative self-correction capabilities, it is well-suited for local interactive applications like code completion and inline editing. Released under Apache 2.0 license on Hugging Face, it supports toolchains including MLX, vLLM, and Transformers.

On June 10, OpenAI and Visa announced an expanded partnership integrating Visa payment services into the OpenAI platform, allowing AI agents like ChatGPT to autonomously complete online shopping and payments upon user authorization. Users need only state a request (e.g., 「Order me a box of tissues to be delivered home next week」), and the AI handles the entire process—from search and price comparison to checkout and payment. This marks a shift from AI as an information assistant to an execution agent, potentially reshaping e-commerce entry points. Both companies will also develop enterprise-grade AI payment applications to advance AI use in business settlements and financial management. Mastercard and PayPal are also building similar ecosystems. Visa views 2026 as a pivotal year for 「AI-assisted shopping」, though widespread adoption will take time.

Full-stack robotics company NEURA Robotics announced on June 10 the closure of a massive $1.4 billion Series C funding round—the largest ever for a full-stack robotics firm. Led by Tether and joined by Qualcomm, Amazon, NVIDIA, Bosch, Schaeffler, and the European Investment Bank, the funds will accelerate development of its 「Physical AI」 platform, Neuraverse. NEURA plans to mass-produce humanoid robots targeting millions of units by 2030 and globally deploy NEURA Gyms—real-world training environments for cognitive robots. The company currently holds over $10 billion in orders, partnering with industrial and tech giants, and emphasizes an open, decentralized AI architecture combining edge intelligence with machine-native economic systems.

Zhipu Huazhang Technology Co., Ltd. (Stock Code: 02513) listed on the Hong Kong Stock Exchange, issuing 37.4195 million shares at HK$116.2 per share, raising a total of HK$4.3 billion, with an opening market cap exceeding HK$50 billion, making it the world's first publicly traded large model company. Cornerstone investors include 11 institutions such as Beijing Financial Holding Group, Taikang Life Insurance, and GF Fund, subscribing HK$2.98 billion—nearly 70% of the total offering. Zhipu focuses on general artificial intelligence, with its core GLM series launching versions from GLM-4.5 to 4.7, visual model GLM-4.5V, and agent AutoGLM. In the first half of 2025, revenue reached RMB 191 million, more than tripling year-on-year, while R&D investment hit RMB 1.595 billion, keeping the company in loss. GLM-4.6 has been adapted to domestic chips including Cambricon and Moore Threads. On the same day, MiniMax also listed in Hong Kong with a market cap surpassing HK$70 billion.

NVIDIA introduced Nemotron 3 Super, an open-source large language model designed specifically for AI agents, featuring 120 billion parameters and a 1 million token context length. Utilizing a hybrid Mamba-MoE architecture, LatentMoE, and multi-token prediction, it achieves three times faster inference and five times higher throughput. The model excels in Pinchbench and OpenClaw tasks, achieving an 85.6% success rate, approaching Claude Opus 4.6. Trained on 25 trillion tokens and post-trained across 21 reinforcement learning environments, it shows strong performance in software engineering, long-context understanding, and mathematical reasoning. It has already been integrated by companies including Perplexity, CodeRabbit, and Greptile. NVIDIA also plans to launch an enterprise-grade open platform, NemoClaw.

On June 10, the European Commission published the final version of its 「Code of Practice on the Transparency of Artificial Intelligence Generated Content」, aimed at supporting compliance with transparency obligations under Article 50 of the AI Act, which becomes enforceable on August 2, 2026. Drafted by six independent experts and informed by over 180 stakeholders, the Code consists of two parts: Part One requires generative AI system providers to apply machine-readable labels to AI-generated or manipulated content; Part Two obliges deployers to clearly disclose deepfakes and AI-generated texts involving public interest. Signing the Code is voluntary, but the transparency requirements themselves are legally binding. The EU has also introduced a set of icons for labeling AI-generated content.

Canadian AI company Cohere announced the acquisition of German AI firm Aleph Alpha, forming a transatlantic strategic alliance aimed at building indigenous European AI capacity. The combined entity will deliver secure, customized AI solutions, focusing on highly regulated sectors such as public administration, finance, and healthcare. This collaboration is seen as a significant counterbalance to the dominance of a few global tech giants in AI, helping Europe develop sovereign AI within controlled and compliant environments. By integrating their technological strengths and customer bases, the new organization aims to strengthen its position in the European market and promote a balanced approach to AI innovation and regulation.

Japanese electronics manufacturer TDK announced it will acquire U.S. startup Fabric8Labs for up to $400 million to enhance its technological competitiveness in AI data centers. Fabric8Labs specializes in manufacturing high-efficiency heat dissipation components using 3D printing, effectively cooling GPUs and meeting the growing thermal demands of AI data centers. Meanwhile, Google is reportedly planning to produce over 3 million of its custom TPUs through Intel’s foundry by 2028, and SK Hynix is testing Intel’s EMIB packaging technology for HBM integration. Due to LPDDR shortages, Nvidia has reduced the capacity of its next-generation AI server memory module SOCAMM2 from 192GB to 96GB. These developments reflect concurrent diversification and expansion in the AI chip supply chain.