OpenAI has released LifeSciBench, a life sciences benchmark developed with 173 scientists, featuring 750 expert tasks across seven workflows including reasoning, evidence handling, and real-world constraints. The specialized model GPT-Rosalind outperforms GPT-5.5 across the board. Concurrently, OpenAI revealed that GPT-5.4, in collaboration with Molecule.one's Maria AI, completed an end-to-end drug discovery pipeline—from literature review to experimental validation in approximately 2.5 months (an additional half-month would be required for manual writing). In testing across 10,080 reactions, yield improvements reached 88% for boronic acid reactions and 83% for sulfonamide reactions, with 8 reactions showing more than a two-fold increase in yield. This marks a milestone where cutting-edge models can support full-cycle scientific research.

DeepSeek has secured over $7.4 billion in funding, pushing its valuation past $50 billion. The capital will be allocated toward R&D and expanding computing infrastructure. This round further solidifies its competitive position in the open-source large model arena. Meanwhile, Z.ai launched GLM-5.2, supporting a 1 million token context window and focused on long-duration coding tasks, with plans to release API access and open-source weights in one week.

AI coding tool Cursor has rolled out continuous cloud capability updates: introducing the /in-cloud command to launch sub-agents within isolated cloud virtual machines for long-running or parallel tasks; it also supports migrating local AI agents to the cloud for concurrent execution, enabling users to issue commands remotely via mobile devices. These enhancements strengthen Cursor’s capabilities in enterprise-grade AI programming scenarios—particularly parallel processing and remote collaboration—and align with SpaceX’s previously reported $60 billion acquisition of its parent company Anysphere.

Anthropic has launched Claude Design, a new design tool now available in beta for all paid plans, accessible via web and desktop. The tool maintains brand design system consistency, supports direct canvas editing, and enables bidirectional synchronization with Claude Code—designs can be handed off seamlessly to development, and design projects can be synced back from the terminal. Anthropic also hosted Claude Build Day in San Francisco, attended by 300 developers. Winning projects included a real-time engine that converts a single smartphone photo into an editable 3D scene.

Meta and Google have integrated new AI features into everyday devices. On June 15, Meta launched AI Mode search for Facebook, powered by the Muse Spark model, allowing users to receive public content recommendations from Facebook, Instagram, and Threads via Meta AI. It also introduced creative photo and video editing tools, with advanced features requiring a $3.99/month subscription. On June 16, Google released a Pixel update featuring Gemini Omni video creation and Lyria 3 music generation, enabling natural language-driven video creation and AI avatar generation. Music generation is free, while video features require a $20/month subscription. Tripo AI, a 3D world modeling company, also raised nearly $200 million in funding.

Allen AI has introduced MolmoMotion, a 3D motion prediction model capable of forecasting future 3D point trajectories of objects from just a single video frame and language instruction. The model uses object-attached 3D points in world space as a category-agnostic, viewpoint-stable motion representation, outperforming existing approaches—including pixel-space video generators and parametric 3D methods—on the new benchmark PointMotionBench. Motion knowledge transfers effectively to downstream tasks: robotic pick-and-place achieves a 76.3% success rate with faster learning, and video generation improves across all five metrics. The team has also open-sourced the million-scale dataset MolmoMotion-1M and benchmark suite.

Vercel has open-sourced eve, a file-system-first backend agent framework built on the principle 「directory as contract」—each agent corresponds to a file directory, where folders like tools/, skills/, and channels/ directly map to functional capabilities, eliminating boilerplate registration code. The framework includes six production-grade features: persistent execution, sandboxed computation, human approval workflows, secure connections, multi-channel support, and built-in tracking and evaluation. Vercel currently runs over 100 eve agents internally, including a data analyst agent handling 30,000 queries monthly and an autonomous sales development agent delivering a 32x return on a $5,000 annual cost.

AI researchers Gary Marcus and contributors from LessWrong argue that the Trump administration’s request for Anthropic to build a completely unjailbreakable large language model is technically impossible, due to the inherent insecurity of next-token prediction architecture—a fundamental issue shared across generative AI, not a failure of any single company. LessWrong further reveals that the government’s shutdown of Anthropic’s Fable and Mythos models was based on fabricated jailbreak claims—the so-called jailbreak merely involved asking the model to 「fix code,」 a standard defensive function. The article criticizes ad hoc export control policies for undermining U.S. AI leadership, forcing companies to shift toward non-U.S. partners.

OpenAI is testing an upgraded version of its GPT-Bidi speech model to make ChatGPT voice interactions more human-like. Unlike current voice assistants that require pausing between turns, GPT-Bidi supports simultaneous listening and speaking, handles interruptions, and adjusts responses mid-sentence. OpenAI has also added Codex support for the browser DevTools Protocol. Developer experiments show that LLMs embedded in robotic bodies can reason and learn, but still lack understanding of the physical world—next breakthroughs will require predictive 「world models」 capable of simulating physical outcomes.

Radical AI has built an autonomous lab to accelerate materials discovery, forming a closed-loop system where AI scientists generate hypotheses and automated robots synthesize and characterize materials in parallel. Within six months, the system produced and characterized 1,200 alloys—10 times faster than DARPA/GE MACH project’s goal of 500 per year—with potential to scale to 100 alloys per day. The AI proposed and tested 300 novel materials, identifying 10 with unprecedented top-tier performance, including discoveries in alloy families never previously studied in academic literature, potentially alleviating supply chain bottlenecks. The team has also open-sourced the TorchSim molecular dynamics framework and MATRIX benchmark dataset.