Anthropic advances math, watermarking; robotics scales

Anthropic / Claude ecosystem

Anthropic’s Unreleased Claude Model Improves Riemann Zeta Function Lower Bound

Anthropic reports that an unreleased Claude model improved the proven lower bound for zeros on the critical line of the Riemann zeta function from 41.6% to 67.2%, using a multi-phase workflow with Lean formalization and external review. This marks a significant step in AI-assisted mathematical research.

Anthropic to Embed Invisible Watermarks in Claude Outputs for EU AI Act Compliance

Anthropic announced it will add undetectable watermarks to all Claude-generated text and files, along with signed provenance metadata, to meet EU AI Act transparency requirements. The watermarks persist through copy-paste and are not visibly perceivable.

Anthropic, Macquarie, and GIC Launch Theseus Infrastructure for AI Data Centers

Anthropic, Macquarie Asset Management, and GIC formed Theseus Infrastructure, a platform to finance, develop, and lease large-scale data centers to Anthropic, aiming to secure scalable AI compute capacity in the US. The deal includes commitments to cover grid-upgrade costs and electricity-price increases.

Frontier model providers

Gemini 2.5 Pro Gets 30% Price Cut and Doubled Rate Limits

Google DeepMind has reduced the price of Gemini 2.5 Pro by 30% and doubled paid-tier rate limits, effective immediately across Google AI Studio and Vertex AI, making the advanced model more cost-accessible.

Mistral AI Aims to Build 1 GW of European AI Compute by 2030

Mistral AI announced plans to build 1 gigawatt of dedicated AI compute in Europe by 2030, offering region-locked inference through Mistral Regional Endpoints and a new Priority Tier, aiming to lock in enterprise customers with sovereign AI infrastructure.

AI developer tooling & infrastructure

Leaked Details Suggest Cursor Composer 3 (Vega) – Faster, Cheaper AI Coding Model

Leaks around Cursor's next iteration, Composer 3 codenamed 'Vega', indicate significant performance and cost improvements, though no official release date has been confirmed.

Centaur 2.0: Open-Source AI Agent with Permissioned Access and MCP Interoperability

Paradigm released Centaur 2.0, an open-source AI agent rebuilt with permissioned access controls, per-channel context, and MCP interoperability, enabling consistent tool and data usage across models.

SearchUnify MCP Server Provides Governed Grounding Layer for Enterprise LLMs

SearchUnify launched an MCP server that acts as a governed grounding layer, connecting MCP-compatible assistants to authorized enterprise data and tools, with permissioned, auditable answers.

Cloud & platform providers

No significant new developments.

AI policy, regulation & governance

Australian AI Safety Institute Publishes Governance Framework for AI Agents

The Australian AI Safety Institute released its first publication, a report proposing a three-tier governance framework for AI agents interacting across organisations, identifying governance gaps and recommending controls.

Industry & market moves

OpenAI Completes $7 Billion Employee Tender Offer

OpenAI reportedly concluded a $7 billion tender offer to provide liquidity to employees, reflecting its high valuation and ongoing private status, while fueling speculation about a potential future IPO.

Hassabis Addresses Departure Speculation, Outlines AGI Timeline in LinkedIn Post

Following reports of his potential departure alongside Jeff Dean, Google DeepMind co-founder Demis Hassabis shared a lengthy LinkedIn post clarifying his stance and hinting at AGI timelines, adding new context to the leadership reshuffle.

Meta Unveils $1B Community Fund for Data Center Areas, Pushes Open-Weight AI Vision

As part of its global AI strategy, Meta announced a $1 billion fund to support communities near its data centers, alongside its promotion of open-weight AI models like Muse Glimmer and Spark. The move aims to foster goodwill and compete with rivals.

Paradium.AI Acquires InfoSentience to Scale AI-Driven Content Generation

The Arena Group's AI division, Paradium.AI, acquired InfoSentience to enhance its data-driven content generation capabilities, aiming to expand B2B revenue through AI-powered content services.

Fei-Fei Li's World Labs Acquires SceniX for Robotic Training Data Pipeline

World Labs, led by Fei-Fei Li, acquired SceniX to build a joint pipeline for real-to-sim-to-real robotic training, addressing bottlenecks in robotic simulation and data generation.

AI product & feature launches

The Browser Company to Add Agentic Capabilities to Dia in Fall 2026

The Browser Company confirmed that its Dia browser will gain agentic browsing features this fall, enabling AI-driven task completion directly within the browser, potentially reducing reliance on external AI assistants.

Upwork Expands AI Integration with 'Talent Is Now Everywhere AI Works'

Upwork is deepening its AI integration, enabling AI-driven talent matching and workflow automation across its platform, signaling a strategic shift toward AI-enabled freelancer marketplaces.

Alibaba to Release 2.4 Trillion Parameter Open-Weight Model

Alibaba announced it will release an open-weight model with 2.4 trillion parameters this week, significantly expanding the availability of large-scale open models for research and commercial use.

NVIDIA Magpie TTS: Open-Weight, Low-Latency Multilingual Voice Agent Builder

NVIDIA released Magpie TTS, an open-weight text-to-speech model optimized for low-latency multilingual voice agents, offering full deployment control on Hugging Face and NVIDIA platforms.

Inception Labs released Mercury 2, a new AI model tailored for search applications, promising faster and more accurate results for enterprise search use cases.

Dyna Robotics Unveils DYNA-2 World-Action Model Trained on 1M Hours of Human Video

Dyna Robotics released DYNA-2, a foundation model for robotics trained entirely on over 1 million hours of human egocentric video, establishing a new scaling law for physical AI and advancing robot manipulation capabilities.

Figure AI Demonstrates Home Automation with Robots Tidying Bedrooms

Figure AI showcased its humanoid robots performing home tasks such as tidying bedrooms, indicating progress toward practical domestic automation and broader home robotics applications.

NVIDIA Launches Nemotron 3.5 Lightning and NeMo Switchyard for Efficient Agentic AI

NVIDIA introduced Nemotron 3.5 Lightning, an optimized model for agentic AI workloads, along with NeMo Switchyard, an open-source routing library to smartly direct queries among models for better performance and cost.

IBM and CHRIST University Launch AI Innovation Center in India

IBM and CHRIST University inaugurated a new AI Innovation Center in Bangalore to advance AI research, education, and talent development, strengthening the AI ecosystem in India.

Research with immediate practical relevance

No significant new developments.