Anthropic Tops OpenAI at $30B ARR; Claude Managed Agents Goes Beta

Anthropic Overtakes OpenAI in Revenue; Claude Managed Agents Launches

• Anthropic surpassed OpenAI in annualized revenue, now running at $30B ARR vs. OpenAI's $24B — up from $9B in late 2025. Enterprise large accounts (>$1M/year) doubled from 500 to 1,000+ in under two months.
• Anthropic launched Claude Managed Agents in public beta — a fully managed infrastructure layer for deploying AI agent fleets without custom setup. Pricing: standard API token costs plus $0.08/active session-hour. Features: sandboxed code execution, scoped permissions, MCP server connectivity, end-to-end tracing, and checkpointing. Early users include Notion, Rakuten, Asana, and Sentry.
• The New Stack reports internal benchmarks show 10-point improvements in file generation success rates vs. standard prompting; Anthropic claims it cuts time-to-market for AI agents from months to days. Advanced memory, multi-agent orchestration, and self-evaluating agents are in limited preview.
• Accenture invested in Replit and formed a strategic partnership to bring Replit Agent 4 to enterprise clients globally. Replit has 50M+ users and is used at 85% of Fortune 500 companies.

---

New Developer Tools: VS Code Agents App, Zenflow, Colab MCP, Citrix AI Gateway

• Microsoft shipped VS Code 1.115 with a new "VS Code Agents" companion app — a preview app for agent-native development enabling parallel agent sessions across multiple repos, inline diff review, PR creation, and agent feedback loops. A new send_to_terminal tool gives agents interactive access to background terminals. VS Code has moved to a weekly release cadence.
• Zencoder launched Zenflow Work (April 9), extending its AI orchestration platform to automate "the other 75% of enterprise work" — planning, communication, and coordination across Jira, Linear, Notion, Gmail, and Google Docs. Model-agnostic (Claude, ChatGPT, Gemini); VS Code and JetBrains plugins included.
• Google open-sourced the Colab MCP Server, letting MCP-compatible agents (Gemini CLI, Claude Code) delegate compute-intensive or risky code execution to Google Colab remotely, create and manage notebooks, and access GPU infrastructure from local agent setups.
• Citrix launched NetScaler AI Gateway (April 9), an enterprise governance layer for AI application delivery. Capabilities: token-based rate limiting for LLM cost control, token-latency load balancing, spillover routing between models, prompt management/redaction, PII protection, and a NetScaler Console MCP Server to extend governance to agentic workflows.
• OutSystems shipped four product updates alongside its Agentic Systems Engineering framework (previously reported): ODC self-hosted deployment for private/public cloud and on-premises Kubernetes, ODC Semantic Search beta for meaning-based retrieval in apps and agents, an A2A Connector for inter-agent communication with non-OutSystems agents, and the new Mentor conversational development experience in ODC Studio.

---

A2A Turns One: 150+ Orgs, Cloud Integration, and an Agent Payments Protocol

• The A2A Protocol reached its one-year milestone with 150+ supporting organizations — up from 50 at launch — including AWS, Google, IBM, Microsoft, Salesforce, SAP, and ServiceNow. (Previously: AAIF formally released v1.0 at the MCP Dev Summit — now the ecosystem has tripled in backing.) Version 1.0 adds enterprise multi-tenancy, signed Agent Cards for identity verification, and a migration path for early adopters.
• Microsoft integrated A2A into Azure AI Foundry and Copilot Studio; AWS added it to Bedrock AgentCore Runtime — marking the first major cloud-native A2A integrations.
• A new complementary Agent Payments Protocol (AP2) — enabling secure, agent-driven financial transactions — launched with 60+ supporting organizations alongside the A2A milestone announcement.

---

Security: AI Framework Attack Wave, 76% NHI Surge, EU Deadline Looms

• A coordinated attack week hit AI infrastructure on March 24: three critical CVEs targeted LangChain/LangGraph (60M+ weekly downloads), including "LangGrinch" CVE-2025-68664 (CVSS 9.3) for secret extraction and RCE; a Langflow RCE was weaponized within 20 hours and added to CISA's KEV; the LiteLLM PyPI supply chain compromise harvested AWS/GCP/Azure tokens from ~40,000 downloads. Security researchers warn the vulnerability-to-exploitation window has collapsed to hours for AI infrastructure.
• SANS Institute reports AI agents have driven a 76% surge in non-human identities in enterprises — 74% of organizations are already running agents that require credentials, and NHI counts have "quietly doubled or tripled." 92% fail to rotate machine credentials on a 90-day cycle; 5% don't know if they're running agentic AI.
• A new open-source Python SDK called Asqav signs every AI agent action with a quantum-safe ML-DSA-65 (FIPS 204) signature and chains them into a tamper-evident audit trail. Supports LangChain, CrewAI, LiteLLM, Haystack, and the OpenAI Agents SDK. Includes policy enforcement, m-of-n multi-party approval for critical actions, and EU AI Act article mapping.
• EU AI Act enforcement begins August 2026, requiring unique agent identities, cryptographically-signed immutable logs, rapid revocation (within seconds), and Article 13-compliant vendor documentation. Multi-agent systems that chain decisions across organizational boundaries are flagged as the hardest to audit.

---

OpenAI Codex at 3M Weekly Users; "Spud" Cybersecurity Model in Development

• OpenAI Codex hit 3 million weekly users — gaining 1 million users in under a month. Sam Altman reset all usage quotas to celebrate and pledged to repeat this at every million-user increment up to 10M. Older Codex models (GPT 5.2-codex, GPT 5.1 codex mini) will be retired starting April 14.
• OpenAI is developing a new cybersecurity model codenamed "Spud" to rival Anthropic's Mythos (previously: Anthropic briefed government officials that Mythos is "far ahead of any other AI model in cyber capabilities"). Greg Brockman described Spud as rooted in two years of research and a meaningful AGI step.