Anthropic, DeepMind push agents; OpenAI ships GPT-5.5

Anthropic / Claude ecosystem

Anthropic launches 'Project Deal,' an experiment in AI agents facilitating commercial transactions for humans

Anthropic conducted 'Project Deal,' a week-long experiment in its San Francisco office where Claude AI agents represented employees in a classified marketplace, negotiating and completing 186 deals for physical goods. The agents autonomously identified matches, proposed prices, and counter-offered without human intervention. The study found that 'smarter' agents achieved objectively better outcomes, though participants with 'weaker' agents did not perceive a disadvantage.

Anthropic confirms three bugs caused Claude Code performance degradation, resets usage limits

Anthropic has acknowledged that Claude Code experienced significant performance degradation over the past two months due to three separate bugs: a misjudged reasoning effort default, a flawed caching optimization that caused memory loss, and a system prompt change that reduced coding quality. The company has resolved these issues, rolled back changes by April 20 (v2.1.116), and is resetting usage limits for all subscribers as of April 23. The API was unaffected.

Anthropic expands Claude Connectors to include 15 new consumer-focused apps

Anthropic has launched 15 new consumer-focused integrations, called 'Connectors,' for its Claude AI assistant. These new integrations include services like Spotify, Uber, Uber Eats, Instacart, Intuit TurboTax, and Booking.com, expanding Claude's app-like functionality beyond business tools. Claude suggests the most relevant app based on conversational context without sponsored recommendations, and requires user confirmation before any transactions.

Frontier model providers

OpenAI releases GPT-5.5, its 'smartest and most intuitive' model yet, towards an AI 'superapp'

OpenAI has launched GPT-5.5, an upgraded AI model touted as its 'smartest and most intuitive to use model' yet, with enhanced capabilities in agentic coding, knowledge work, and scientific research. The release is a step towards OpenAI’s vision of a unified AI 'superapp' integrating ChatGPT, Codex, and an AI browser into a single service. GPT-5.5 is available to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, with API access coming soon.

OpenAI launches GPT-5.5 Bio Bug Bounty to strengthen safeguards against biological threats

OpenAI has launched the GPT-5.5 Bio Bug Bounty program, inviting cybersecurity researchers, biosecurity experts, and AI red teamers to identify vulnerabilities that could enable malicious actors to exploit AI for harmful biological research. The program's central challenge is to develop a 'universal jailbreak' prompt that consistently forces GPT-5.5 to bypass safety filters and answer a five-question biosafety challenge, with a top prize of $25,000. Testing is limited to GPT-5.5 within the Codex Desktop environment.

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.

DeepSeek AI releases V4 series with 1M-token context and efficient sparse attention architecture

DeepSeek AI has released preview versions of its DeepSeek V4 series, consisting of two Mixture-of-Experts (MoE) models: V4-Pro (1.6T total params, 49B active) and V4-Flash (284B total params, 13B active). Both natively support a 1 million-token context window, leveraging a hybrid architecture with Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for substantial efficiency gains at inference. DeepSeek V4-Pro-Max leads open-source models in coding and mathematics benchmarks and is reportedly fully compatible with Huawei Ascend chips.

AI developer tooling & infrastructure

Runloop launches benchmark orchestration platform with Weights & Biases integration for trusted AI agent deployment

Runloop has launched its Benchmark Job Orchestration platform, designed for continuous evaluation and scalable deployment of AI agents. It integrates with Weights & Biases Weave, providing full traceability and deep visibility into agent behavior beyond high-level metrics. The platform executes thousands of benchmark scenarios in parallel across real-world environments, detecting regressions and enabling data-driven decisions on agent and model selection. It aims to build trust in production AI systems by ensuring reliable performance within defined boundaries.

Cloud & platform providers

Google Cloud commits $750 million to accelerate partners' agentic AI development

Google Cloud has announced a $750 million fund to provide resources and incentives to its 120,000-member partner ecosystem, accelerating joint customers’ transformations with agentic AI. The fund supports AI value identification, prototyping, agent building and deployment, upskilling, and embedded Google forward-deployed engineers (FDEs) for global consulting firms, systems integrators, software partners, and channel partners.

Google rebuilds Workspace for agents with 'Workspace Intelligence,' enabling context-rich AI workflows

Google has introduced 'Workspace Intelligence,' a significant upgrade to Google Workspace, designed to understand real-time semantic relationships among apps like Gmail, Docs, and Sheets. This system supercharges context for agentic workflows, allowing Gemini to gather information across apps, understand what's important, and tailor outputs to user communication patterns. New features include 'Ask Gemini in Chat' for daily briefings and task completion, natural language spreadsheet editing, AI Overviews in Drive, and an AI Inbox for summarization.

AI policy, regulation & governance

Discord group gains unauthorized access to Anthropic's Mythos AI model, raising cybersecurity concerns

A Discord group has reportedly gained unauthorized access to Anthropic's powerful Mythos AI model, raising serious questions about the guardrails surrounding advanced AI cybersecurity tools. This incident highlights the potential for misuse of models capable of identifying and exploiting software vulnerabilities.

Canada launches 'LIFT' program with $500M to accelerate AI adoption for over 1,000 SMEs

The Business Development Bank of Canada (BDC) has launched its new LIFT (Lead with Innovation and Focus on Technology) initiative, committing $500 million to help over 1,000 Canadian small- and medium-sized enterprises (SMEs) adopt AI. The program pairs business owners with expert AI advisors and offers flexible financing. Early users have shown 24% higher productivity. LIFT prioritizes Canadian-developed AI tools and equipment, aiming to boost national economic sovereignty and innovation.

Industry & market moves

Google announces up to $40 billion investment in Anthropic, expands cloud and compute partnership

Google's parent company, Alphabet, will invest up to $40 billion in AI startup Anthropic, with an initial $10 billion cash injection at a $350 billion valuation and another $30 billion contingent on performance targets. This deepens Google's partnership with Anthropic, providing a fresh 5 gigawatts of compute capacity from Google Cloud over the next five years. This follows Amazon's recent commitment of up to $25 billion to Anthropic.

Salesforce and Google Cloud deepen AI alliance for enterprise workflow agents with zero-copy data access

Salesforce and Google Cloud have announced a deep integration connecting AI agents across Slack, Google Workspace, Agentforce, and Gemini Enterprise. This partnership introduces zero-copy data access, Gemini-powered reasoning, and bidirectional workflow automation across core enterprise systems. It aims to address long-standing issues like manual context switching, fragmented data silos, and complex custom integrations in large organizations.

Meta signs multi-billion-dollar deal with AWS to deploy Graviton chips for agentic AI workloads

Meta has signed a multi-year, multi-billion-dollar agreement with Amazon Web Services (AWS) to deploy tens of millions of AWS Graviton processors for its AI workloads, making it one of the largest Graviton customers globally. This deal marks a significant expansion of their long-standing partnership, with Graviton5 chips specifically powering CPU-intensive agentic AI workloads such as real-time reasoning, code generation, search, and orchestrating multi-step tasks. This diversifies Meta's compute sources beyond GPUs and addresses the demand for efficient inference.

Cohere and Aleph Alpha merge to form transatlantic 'sovereign AI' powerhouse, backed by Schwarz Group

Canadian AI firm Cohere and German AI startup Aleph Alpha have announced plans to merge, creating a transatlantic 'sovereign AI' powerhouse valued at approximately $20 billion. The combined entity will maintain dual headquarters in Canada and Germany, focusing on providing secure, customized AI solutions for governments and regulated industries globally. As part of the merger, Germany's Schwarz Group (parent of Lidl) will invest $600 million in Cohere's upcoming Series E funding round and host Cohere's AI systems on its STACKIT cloud platform.

AI product & feature launches

OpenAI launches 'ChatGPT for Clinicians,' a free version for verified U.S. healthcare professionals

OpenAI has introduced 'ChatGPT for Clinicians,' a free, specialized version of ChatGPT for verified physicians, nurse practitioners, physician assistants, and pharmacists in the U.S. Designed to support clinical tasks like documentation, medical research, and evidence review, it leverages advanced AI models for complex clinical questions, offers reusable skills for workflows, and provides real-time, cited answers from medical sources. It also supports earning CME credits and offers optional HIPAA compliance.

Citi Wealth launches '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.

LexisNexis has launched Protégé General AI in Hong Kong, available within Lexis+ Hong Kong. This expands its personalized agentic AI capabilities, offering legal professionals secure and integrated access to general-purpose AI for research, communications, and enriching legal work with real-world context, all within a single, private, and encrypted platform. Users can seamlessly switch between legal-specific and general AI without compromising data security or privacy.

Grow Therapy introduces AI Coach with clinician oversight to support clients between therapy sessions

Grow Therapy has launched an AI-powered coach ('Coach') as a chat feature within its app, designed to support clients between therapy sessions with proprietary safety features developed and monitored by licensed clinicians. The AI coach draws from evidence-based mental health frameworks (CBT, DBT, ACT, BA), providing a nonjudgmental space for practicing skills and processing emotions. Since its testing phase began in December 2025, over 800,000 messages have been sent, with 50% of active providers having at least one client using it. Safety is central, with automated quality scoring and continuous review by licensed clinicians.

ThinkAhead Corporation launches Diluta, an AI productivity platform designed around human energy patterns

ThinkAhead Corporation has launched Diluta, an intelligent productivity platform that uses AI to align work with human energy patterns. Instead of rigid schedules, Diluta helps users discover their chronotype, understand their productivity archetype, automate routines, and optimize focus based on real human behavior. It aims to reduce burnout and constant context switching by scheduling demanding tasks during peak energy windows, integrating behavioral science with modern task management.

Research with immediate practical relevance

Google DeepMind unveils Decoupled DiLoCo for resilient, distributed AI training across global data centers

Google DeepMind has introduced Decoupled DiLoCo (Distributed Low-Communication), a new distributed architecture for training frontier AI models. This approach divides large training runs across decoupled 'islands' of compute with asynchronous data flow, isolating local disruptions and allowing other parts of the system to continue learning efficiently. It's more resilient and flexible than traditional methods, avoids communication delays, and successfully trained a 12 billion parameter model across four U.S. regions 20 times faster than conventional methods.