OpenAI Launches GPT-Rosalind for Life Sciences

OpenAI released the GPT-Rosalind model series, designed for enterprise-level life science research, integrating GPT-5.5's agent programming and tool use capabilities for drug discovery, analysis design, and experimental workflows.

OpenAI introduced GPT-Rosalind, a new model series purpose-built for life sciences research at enterprise scale. The system combines GPT-5.5's agentic coding and tool-use capabilities with specialized intelligence for drug discovery, molecular analysis, experimental design, and laboratory workflows. Industry observers note this represents a significant step toward domain-specialized frontier models that can operate with the precision and safety required in regulated scientific environments. The release underscores OpenAI's strategy of building vertical AI solutions for high-impact sectors beyond general-purpose chat interfaces. Early adopters in pharmaceutical companies have already begun evaluating the model for target identification and lead optimization pipelines, areas where traditional computational methods have historically required months of manual effort. The integrated agentic framework allows researchers to describe experimental goals in natural language and receive executable protocols along with predictive analysis, compressing timelines that previously spanned weeks into hours.

Anthropic Automates 95% of Business Analytics with Claude

Anthropic shared best practices for automating 95% of business analytics queries using Claude, achieving approximately 95% accuracy. The core method involves building a strong data foundation to reduce ambiguity, establishing knowledge sources that map user questions to controlled entities, and developing skill modules as on-demand markdown readouts. Without skills, accuracy drops to just 21%. Once integrated, accuracy stabilizes above 95%. The approach frees data science teams from repetitive queries, letting them focus on causal modeling, forecasting, and other strategic work that demands deeper analytical rigor rather than routine report generation.

NVIDIA Cosmos 3 Tops 7 Physical AI Benchmarks

NVIDIA Cosmos 3, the open omni-model for physical AI, now ranks first across seven leaderboards covering world generation, robot action policy, and industrial vision understanding. The benchmarks include Artificial Analysis, PAI-Bench, Physics-IQ, and R-Bench, confirming Cosmos 3's position as the leading open model for physical AI applications spanning manufacturing, logistics, and autonomous systems.

Ideogram 4.0: The World's Best Open Image Model

Ideogram 4.0 has been released with open weights, supporting download, fine-tuning, and deployment on user-owned hardware. The model is available across all Ideogram plans and through the API. With strong community reception and integration on Hugging Face as ideogram-4-nf4, the release signals growing momentum for open-weight image generation models that rival proprietary alternatives. Users can generate high-resolution images with precise text rendering and exceptional design capabilities, making it suitable for brand-level creative work without vendor lock-in.

NVIDIA and Microsoft Partner for Secure AI Agents on Windows

NVIDIA and Microsoft have announced a partnership to deliver secure, user-controllable AI agents on Windows. At the center of the collaboration is the NVIDIA OpenShell runtime, which provides governance tools, policy enforcement, and intelligent local-to-cloud query routing for agentic workloads. Enterprise software leaders including ServiceNow, SAP, and Siemens are building AI agents with NVIDIA technology, targeting engineering, healthcare, software development, and business operations. The partnership extends from Windows devices to AI factories at scale, with Jensen Huang and Satya Nadella jointly demonstrating the agentic AI vision at Microsoft Build in Taipei. Separately, NVIDIA announced local AI agent upgrades for DGX Spark and RTX PCs, including NVIDIA Broadcast 2.2 and upcoming RTX acceleration for Adobe applications and Blender.

Uber Caps Coding Agent Spending at $1,500 per Employee per Month

Uber has set a $1,500 monthly token cap per employee for each AI coding tool including Cursor and Claude Code. With engineers typically using two tools, the annual cap reaches approximately $36,000, roughly 11% of the median engineering salary of $330,000. The policy reflects enterprise efforts to manage AI tool budgets while still providing broad access. Industry analysts view the cap as a practical signal of how large organizations value coding agent productivity, moving beyond per-seat licensing toward consumption-based governance. The move follows a period of rapid adoption where uncontrolled token expenditure had begun outpacing anticipated returns.

Replit Launches ViBench: First End-to-End App Creation Benchmark

Replit CEO Amjad Masad unveiled ViBench, the first benchmark designed for evaluating end-to-end application creation based on real-world tasks. The benchmark reveals that while GPT-5.5 leads on standard SWE benchmarks, Anthropic's Opus 4.8 remains the king of vibe coding on both price and performance when measured by complete app generation. ViBench tests models across the full development lifecycle including UI design, backend logic, database integration, and deployment readiness, filling a gap in current evaluation frameworks that focus narrowly on isolated coding tasks rather than holistic application building capability.