Anthropic announced on May 28 the completion of a $65 billion Series H funding round, achieving a post-money valuation of $965 billion, surpassing OpenAI ($852 billion) to become the world's most valuable AI startup. The round was co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with strategic participation from Samsung, SK Hynix, Micron, and Amazon’s previously committed $5 billion. The company’s annualized revenue has exceeded $47 billion, with projected revenue growth of 130% and potential first-time operational profitability. Since its founding in 2021, it has raised nearly $144 billion cumulatively, making this likely its final private financing before IPO. Funds will be allocated to safety research, compute expansion, and scaling the Claude product line.

On May 28, Anthropic released Claude Opus 4.8, delivering significant improvements in programming, Agent capabilities, and complex reasoning. A new user-controllable 'thinking investment' mechanism enables a 2.5x speed increase in fast mode, reducing cost to one-third of previous levels (input $10/output $50 per million tokens), while standard pricing remains unchanged. Honesty is greatly improved, with probability of unsupported conclusions reduced by approximately fourfold. The 'Dynamic Workflows' feature allows coordination of hundreds of sub-Agents to process large-scale tasks in parallel, such as migrating codebases with hundreds of thousands of lines. Users have already leveraged it to port Bun from Zig to Rust (750,000 lines of code) within 11 days, achieving a test suite pass rate of 99.8%. It scores 84% on Online-Mind2Web and 88.6% on SWE-bench Verified, outperforming GPT-5.5. The model is now available via the Claude API and platforms including AWS and Google Cloud.

On the third anniversary of ChatGPT, DeepSeek launched two new models: V3.2 emphasizes cost-performance efficiency, supports 128K context length, and achieves Agent capability with 'thinking while using tools' for the first time, matching GPT-5 in performance. The Speciale edition focuses on extreme reasoning, achieving gold medal-level results in international competitions such as IMO, IOI, and ICPC. The models employ DSA sparse attention to enhance efficiency, leverage an 'expert distillation + hybrid reinforcement learning' framework to boost capabilities, and utilize a large-scale agent pipeline enabling AI self-evolution. Both models are open-sourced, with weights published on HuggingFace and ModelScope. This release marks that open-source models have approached or even surpassed top closed-source models in reasoning and Agent capabilities.

At Google I/O 2026, Google DeepMind CEO Demis Hassabis stated that AGI could emerge as early as 2029, one year earlier than prior predictions. He said humanity is at the 'foothills of the singularity,' with AI already exhibiting preliminary 'soft self-improvement' abilities, such as coding assistants accelerating R&D. Hassabis warned that society is unprepared for the rapid pace of AI development, noting policy actions lag far behind technological progress, and urged governments to accelerate AI safety regulation. He revealed major AI companies are discussing future safety frameworks, but the window for preparation is rapidly closing. Regarding U.S. efforts to mandate government testing of AI models before release, he called it 'a step in the right direction.'

Under Elon Musk’s leadership, SpaceX has deployed Version 1.0 of a new AI training software stack written entirely in C, optimized for its cluster of 220,000 NVIDIA GB300 GPUs and 800G networking. By precisely mapping hardware topology, implementing deep pipeline parallelism, and running close to bare metal, the system eliminates overhead from Python interpreters and general-purpose frameworks, claiming performance over a full order of magnitude faster than Google JAX on large training jobs. Model FLOPS utilization (MFU) is expected to exceed 80%, significantly higher than the industry average of 50%-67%. This reduces pre-training that previously took 2–3 months to about one week, or enables training models ten times larger in the same timeframe. Musk stated the next step is to write the inference stack in C to support large-scale reinforcement learning.

Cisco’s AI threat intelligence team tested 15 frontier models from OpenAI, Anthropic, Google, Amazon, and xAI, finding multi-turn prompt injection attacks succeed far more often than single-turn ones. Grok 4.1 Fast by xAI achieved an 88.3% multi-turn attack success rate (ASR) versus 34.1% in single-turn; GPT-5.4 rose from 2.7% to 24.7%; Gemini 3 Pro increased from 18% to 73%. The study notes that reliance on single-turn evaluations in mainstream benchmarks may lead to 'security washing,' failing to reflect real-world attack scenarios. Enabling reasoning mode reduced ASR by over 40 percentage points. The report recommends vendors disclose multi-turn ASR data and mandates human safety reviews for models with gaps exceeding 15 percentage points.

At its inaugural event, Mistral AI expanded into a full-stack enterprise AI provider: launching Mistral for Industrial Engineering, integrating LLMs with physical simulation for aerospace and automotive industries, already deployed at Airbus, BMW, and EDF; announcing Mistral Compute’s €4 billion investment to build 1GW of data center capacity in Europe; rebranding Le Chat as Vibe, a unified Agent platform covering enterprise workflows and coding Agents; and integrating specialized models like Pixtral into its flagship Mistral Medium 3.5, shifting toward native multimodal reasoning. Its core positioning is to serve enterprises and governments unwilling to entrust data to U.S. hyperscalers, through open-weight models and proprietary infrastructure.

The Illinois legislature passed SB315, becoming the third U.S. state after New York and California to regulate frontier AI technology. The law requires AI companies with annual revenue over $500 million and substantial compute capacity to establish and publish transparency frameworks, hire third-party auditors for compliance, and report critical security incidents within 72 hours (24 hours in emergencies). Violations carry fines up to $3 million per incident, with no private right of action. The law takes effect in 2028 and has been submitted to the governor for signature. OpenAI and Anthropic support the bill, viewing it as setting an industry benchmark.

Liquid AI released LFM2.5-8B-A1B, a sparse Mixture-of-Experts (MoE) model designed for on-device deployment, with 8.3 billion total parameters but only 1.5 billion activated per token. It supports a 128K context window and covers nine languages. Pretraining data expanded from 12T to 38T tokens, with vocabulary doubled to 128K. Through two-stage reinforcement learning optimization, non-hallucination rate improved from 7.46 to 63.47. On an M5 Max CPU, decoding speed reaches 253 tokens/s, around 30 tokens/s on mobile devices, and up to 18.5K tokens/s on a single H100. It supports frameworks including llama.cpp, MLX, and vLLM, with open weights released under the LFM1.0 license.

China’s State Administration for Market Regulation and National Development and Reform Commission jointly issued the 'Guidelines for Artificial Intelligence Metrology System and Capacity Building (2026 Edition),' systematically laying out six areas including foundational support, general technologies, and core technologies. The guidance aims to address issues such as algorithmic 'black boxes' and poor decision explainability, deploying key technical攻关 on monitoring and characterizing internal states of AI systems, and establishing reliable, secure, and trustworthy AI metrology standards. It supports building national metrology R&D centers, constructing full-chain metrology capabilities covering algorithms, compute, and data, and establishing mechanisms for sharing high-quality datasets. This marks China’s AI development shifting from 'scaling up' to 'strengthening foundations,' integrating metrology technologies into 14 key sectors.