Agentic AI takes center stage; Anthropic limits access

Anthropic / Claude ecosystem

Anthropic tests removing Claude Code from Pro subscription amid capacity constraints

Anthropic is conducting a small A/B test with approximately 2% of new Pro subscribers, temporarily removing access to Claude Code from their $20/month plan. The company indicated this is due to 'usage has changed a lot and our current plans weren't built for this,' suggesting a re-evaluation of its subscription tiers and the compute costs associated with AI coding agents. Existing Pro and Max subscribers are unaffected.

Anthropic introduces Managed Agents to simplify AI agent deployment and operations

Anthropic has launched Managed Agents on its Claude platform, offering a managed execution layer for agent-based workflows. This capability allows developers to define agent behavior, tools, and constraints while offloading runtime responsibilities like orchestration, sandboxing, session state management, and credential handling to the platform. It aims to reduce the infrastructure complexity for deploying long-running, multi-step agent workflows in production.

Frontier model providers

OpenAI launches 'Workspace Agents' in ChatGPT for team-based task automation

OpenAI has introduced 'Workspace Agents' in ChatGPT, an evolution of GPTs designed for teams to automate complex tasks and long-running workflows within organizational contexts. Powered by Codex, these agents can prepare reports, write code, and respond to messages, gathering context from systems, following team processes, and asking for approval when needed. They are available in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans and are free until May 6, 2026.

OpenAI releases ChatGPT Images 2.0 with web search, enhanced text rendering, and advanced AI 'thinking'

OpenAI has launched ChatGPT Images 2.0, a major upgrade to its image generation capabilities, now available in ChatGPT and Codex. The new model, powered by GPT Image 2, features 'thinking capabilities' to interpret prompts better, analyze structure, and pull in real-time information from the web. It offers improved quality, native 2K output, flexible aspect ratios, multilingual text rendering, and the ability to generate up to eight consistent images from a single prompt. It also includes new watermarking and content filtering for safety.

Google DeepMind launches Deep Research and Deep Research Max autonomous AI research agents

Google DeepMind has released two new autonomous research agents, Deep Research and Deep Research Max, in public preview via the Gemini API. Built on Gemini 3.1 Pro, these agents can search the open web, user uploads, and connected data sources via Model Context Protocol (MCP) servers, generate charts natively, and consult over 100 sources per task. Deep Research is optimized for speed, while Deep Research Max is designed for exhaustive, asynchronous background workflows, conducting up to 160 search queries per task.

Ant Group unveils Ling-2.6-Flash, a new LLM prioritizing efficiency and agentic AI applications

Ant Group has officially released Ling-2.6-Flash, a new large language model (LLM) designed for efficiency and real-world AI agent applications. Leveraging a sparse Mixture-of-Experts (MoE) architecture with 104 billion total parameters (only 7.4 billion active), it delivers high intelligence at significantly lower cost and latency. Benchmarked by Artificial Analysis, it achieves an 86% reduction in inference cost and SOTA performance for its size on AI agent benchmarks like BFCL-V4, TAU2-bench, SWE-bench Verified, Claw-Eval, and PinchBench. It's available via API, OpenRouter, and Alipay Tbox, with a commercial version, LingDT, from Ant Digital Technologies.

AI developer tooling & infrastructure

No significant new developments.

Cloud & platform providers

Google Cloud unveils Gemini Enterprise Agent Platform and eighth-generation TPUs

Google Cloud has introduced the Gemini Enterprise Agent Platform, a new management hub for building, scaling, governing, and optimizing AI agents for enterprises. Concurrently, it unveiled its eighth-generation Tensor Processing Units (TPUs), specifically splitting them into TPU 8t for AI training and TPU 8i for AI inference. TPU 8t offers 2.8x more power than Ironwood, while TPU 8i improves inference performance by 80% per dollar, featuring more on-chip SRAM for agent workloads.

Cloudflare Sandboxes reach General Availability, offering persistent, isolated Linux environments for AI agents

Cloudflare has announced the general availability of Sandboxes and Cloudflare Containers, providing persistent, isolated Linux environments for AI agent workloads as part of its Agents Week. The GA release adds features like secure credential injection, PTY terminal support, persistent code interpreters, filesystem watching, snapshot-based session recovery, and active CPU pricing, charging only for used cycles. Figma is reportedly already running production agent workloads on this infrastructure.

Cloudflare achieves 93% internal R&D adoption of its AI engineering stack, built on its own platform

Cloudflare has achieved a 93% adoption rate for AI coding tools across its R&D organization (3,683 users) by building an internal AI engineering stack entirely on its own platform products. The infrastructure processed over 241 billion tokens through AI Gateway and 51 billion tokens through Workers AI monthly, demonstrating measurable impact with developer productivity nearly doubling (merge requests increasing from ~5,600 to over 8,700 weekly). This internal-only stack leverages AI Gateway, Workers AI, Zero Trust, Sandbox, and Code Mode.

Cloudflare outlines MCP architecture for enterprise AI agent governance to counter security risks

Cloudflare has outlined a reference architecture for scaling Model Context Protocol (MCP) deployments across enterprises, emphasizing centralized governance, remote server infrastructure, and cost controls. This comes amid research highlighting risks like prompt injection and supply chain attacks in MCP-based systems. Cloudflare advocates deploying MCP servers remotely on its platform, managing authentication via Cloudflare Access, and using an 'AI Gateway' for cost control and model routing. It also introduced 'Code Mode' to reduce token usage by collapsing tool interfaces.

AI policy, regulation & governance

No significant new developments.

Industry & market moves

Google DeepMind partners with global consultancies to accelerate enterprise AI adoption

Google DeepMind is partnering with Accenture, Bain & Company, BCG, Deloitte, and McKinsey to accelerate AI-driven transformation for global organizations. This initiative aims to bring frontier AI to businesses by enabling scaled, industry-specific AI capabilities, providing early access to Gemini models, and connecting DeepMind leadership with customer CEOs and boards to navigate AI R&D. The goal is to move businesses past the AI 'pilot phase' to scaled agentic adoption with measurable impact.

Google Cloud and Vista Equity Partners form partnership to accelerate enterprise agentic AI adoption

Google Cloud and Vista Equity Partners have announced a new partnership to accelerate the development, deployment, and distribution of agentic AI solutions across Vista's portfolio of 90+ enterprise software companies. The collaboration provides Vista firms with streamlined access to Google Cloud's AI stack, including Gemini models, AI Hypercomputer, and the Gemini Enterprise platform for building and deploying AI agents. It also creates go-to-market opportunities for Vista's portfolio companies through Google Cloud's Marketplace and co-sell programs.

Meta deploys employee tracking software to train AI agents on human workflows

Meta Platforms has started deploying internal tracking software on U.S.-based employee computers to capture mouse movements, clicks, keystrokes, and occasional screenshots. This data, collected via the 'Model Capability Initiative' (MCI), is being used to train Meta's AI models, specifically for building AI agents that can autonomously perform workplace tasks. Meta states the data is solely for model training and not for performance assessment, with safeguards in place for sensitive content.

xAI held discussions with Mistral and Cursor about a potential three-way partnership

Elon Musk's xAI has held recent discussions with French AI startup Mistral and AI coding startup Cursor about a potential three-way partnership. This follows SpaceX's (which owns xAI) announced deal to potentially acquire Cursor for $60 billion or pay $10 billion for collaboration. The move aims to accelerate xAI's competitiveness against AI rivals like Anthropic and OpenAI in AI coding services and AI agents, by leveraging Mistral's independent model development and Cursor's coding platform and user base.

DeepSeek reportedly seeks $20 billion valuation in first external funding round, attracting Tencent and Alibaba

Chinese AI startup DeepSeek is reportedly in talks with Tencent and Alibaba to raise its first external funding round, targeting a valuation exceeding $20 billion. This follows earlier reports of seeking $300 million at a $10 billion+ valuation. DeepSeek, previously self-funded by High-Flyer Capital Management, has focused on open-source technology, leading to debate about its revenue-light valuation compared to competitors like Moonshot AI, MiniMax, and Zhipu, which are also seeking or have achieved high valuations.

AI product & feature launches

Zscaler joins Anthropic's Project Glasswing to integrate Claude Mythos Preview for cyber defense

Zscaler has joined Project Glasswing, gaining access to Anthropic's Claude Mythos Preview model. Zscaler plans to integrate Mythos into its secure software development lifecycle to identify vulnerabilities and will share findings with other Project Glasswing participants. The company positions this as a strategic shift towards zero-trust network design as AI automates reconnaissance and vulnerability discovery.

Citi Wealth unveils 'Citi Sky,' an AI-powered financial advisor built with Google Cloud and Google DeepMind

Citi Wealth has launched 'Citi Sky,' an always-on AI-powered member of its wealth team, developed using Google Cloud and Google DeepMind technologies. Citi Sky aims to transform client experience by providing actionable insights and anticipating financial needs through advanced real-time avatar technology and Gemini's live audio/video models. It will be integrated into Citi Wealth platforms to work alongside financial advisors, offering guidance, market insights, and conversational interaction in English and Spanish, with a phased rollout starting this summer for Citigold clients.

PolyAI launches Agent Development Kit (ADK) for AI-native development in enterprise CX

PolyAI has introduced its Agent Development Kit (ADK), a new developer-first approach for building, deploying, and improving agentic AI for customer experience. The ADK integrates AI coding assistants like Cursor and Claude Code into the core development process, allowing teams to work in their preferred environments with full control, manage agents like enterprise software with version control, and build from various inputs in minutes. PolyAI reports that over 60% of its internal engineering work is now done autonomously through ADK-powered workflows.

Research with immediate practical relevance

MIT researchers develop RLCR method to teach AI models to express calibrated confidence

Researchers from MIT's CSAIL have developed RLCR (Reinforcement Learning with Calibration Rewards), a method that trains language models to produce calibrated confidence estimates alongside their answers. This technique, which reduces calibration error by up to 90% while maintaining or improving accuracy, addresses the problem of AI models exhibiting overconfidence regardless of their actual certainty. RLCR penalizes models for confidently wrong answers and unnecessarily uncertain correct ones, making confidence estimates practically useful for decision-making in fields like finance and medicine.