At the 2026 Sapphire conference, SAP unveiled its 'Autonomous Enterprise' strategy, announcing the integration of over 200 AI agents into core business applications such as finance, supply chain, and human resources. SAP has formed a strategic partnership with Anthropic, adopting Claude as the core reasoning engine for its Joule intelligent assistant. SAP also launched a €100 million partner fund to accelerate AI deployment and deepened collaborations with AWS, Google Cloud, Microsoft, NVIDIA, and Palantir. SAP claims this strategy can reduce ERP migration costs by more than 35% and shorten financial closing cycles from weeks to days. Notably, SAP's market cap has declined from $300 billion to around $200 billion, while partner Anthropic's valuation is now five times higher, highlighting the market's preference for AI model companies over traditional software firms.

Anthropic has officially launched Claude for Legal, marking a major expansion into legal technology. The platform includes 12 plugins tailored to specific legal domains (corporate, employment, privacy, etc.), integrates MCP connectors for leading legal tools such as DocuSign, LexisNexis, and Thomson Reuters, and fosters an open-source ecosystem contributed by Harvey, Legora, and others. After full deployment at law firm Freshfields, usage surged nearly 500% within six weeks. Leveraging strong document understanding capabilities, Claude has become the foundational LLM for most legal AI tools. Anthropic emphasizes human-in-the-loop decision-making and enhances legal workflow efficiency through customized plugins and a Word plugin integration.

Perceptron AI has launched its flagship video analysis model Mk1, delivering exceptional performance in spatial and video understanding tasks at just 10–20% of the cost of comparable models from Anthropic, OpenAI, and Google. Mk1 outperforms cutting-edge models like GPT-5 and Gemini 3.1 Pro on benchmarks such as EmbSpatialBench and VSI-Bench, priced at $0.15 per million input tokens and $1.50 per million output tokens. The model features causal reasoning about the physical world, supports continuous video processing at two frames per second, and can accurately detect object dynamics and read analog gauges. The company adopts a dual-track strategy: closed-source Mk1 is offered via API for enterprise use, while the open-source Isaac series targets edge deployment. Founders are from Meta FAIR.

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released its first in-house model, TML-Interaction-Small, ending over a year of product silence. The multimodal model processes audio, video, and text synchronously in 200-millisecond intervals, enabling truly real-time interaction. With 276 billion parameters (12 billion activated), it achieves a 0.40-second round-trip latency on FD-bench V1, outperforming GPT-Realtime-2.0 and Gemini-3.1-flash-live. The company previously raised $2 billion at a $12 billion valuation but faced challenges including co-founder departures and fundraising setbacks. The model is currently available only for research preview, with plans to release larger versions; benchmark results have not yet been independently verified.

At the Android Show, Google introduced a new laptop category called Googlebook, designed specifically for Gemini AI, signaling a shift from Chromebook to AI-first devices. The key feature, Magic Pointer—developed by DeepMind—triggers contextual AI suggestions when users hover the cursor, such as creating meetings from email dates or merging image previews. Built on the Android tech stack, the device natively runs Android apps and allows direct access to phone files. Initial models are crafted by Acer, ASUS, Dell, HP, and Lenovo, expected to launch in fall 2026; Samsung did not join the initial rollout. This release follows Apple's introduction of the $600 MacBook Neo by two months.

Isomorphic Labs, led by Google DeepMind co-founder Demis Hassabis, announced a $2.1 billion Series B funding round led by Thrive Capital, with participation from Alphabet, GV, MGX, Temasek, CapitalG, and the UK Sovereign AI Fund. The funds will advance the development of its AI drug design engine, IsoDDE, and push drug discovery into clinical stages. The company has established strategic partnerships with Novartis, Eli Lilly, and Johnson & Johnson. Hassabis stated the goal is to solve all diseases using AI, noting that the engine has demonstrated unprecedented speed and reproducibility in internal projects. The company had previously signaled plans to raise over $2 billion.

Microsoft has launched MDASH, a multi-model agent security system that leverages over 100 specialized AI agents working collaboratively to identify 16 new vulnerabilities in Windows networking and authentication stacks, including four critical remote code execution flaws. In private testing, MDASH detected all 21 known vulnerabilities with zero false positives and achieved a top score of 88.45% on the public CyberGym benchmark. The system uses a multi-stage pipeline architecture, improving accuracy through multi-model collaboration and specialized agents. Retrospective testing showed 96% and 100% recall rates for historical vulnerabilities in the clfs.sys and tcpip.sys components over the past five years. MDASH is currently in limited private preview.

The Zhejiang Provincial Health Information Center and Ant Health jointly released AntAngelMed, an open-source medical large language model with 103 billion parameters, using a 1/32 activation ratio MoE architecture. During inference, only 6.1 billion parameters are activated, achieving up to 7x higher computational efficiency compared to dense models of similar scale. It delivers over 200 tokens/second on H20 hardware and supports 128K context length. AntAngelMed ranks first on three major healthcare benchmarks—OpenAI HealthBench, MedAIBench, and MedBench—demonstrating exceptional performance in medical knowledge QA, ethical safety, and complex reasoning. The model underwent three-stage training: large-scale medical corpus pretraining, supervised fine-tuning, and GRPO reinforcement learning. Weights are open-sourced under the Apache 2.0 license.