Anthropic AI Updates: May 7, 2026

1. Anthropic Buys 220K-GPU SpaceX Cluster, Doubles Claude Code Limits

Anthropic. Anthropic announced a deal for over 300 megawatts of capacity from SpaceX’s Colossus-1, totaling more than 220,000 NVIDIA GPUs and coming online within a month. The new compute lets Anthropic double Claude Code rate limits, drop peak-hour throttling for Pro and Max users, and raise Claude Opus API rate limits. Source

2. Anthropic Commits $200B to Google Cloud Over Five Years

Anthropic. Reports detail a $200 billion, five-year Google Cloud commitment from Anthropic, putting it among Google’s largest committed customers. The deal reinforces Google as a principal training and inference platform for Claude alongside the new SpaceX capacity and AWS Trainium use. Source

3. Claude Managed Agents Add Dreaming, Outcomes, and Multi-Agent Orchestration

Anthropic. Claude Managed Agents picked up three features: “dreaming” (a scheduled job that mines past sessions to curate memories), Outcomes (rubric-graded evaluation that lifted task success up to 10pp on Word and PowerPoint tasks), and multiagent orchestration where a lead agent delegates to subagents over a shared filesystem. Customers cited include Harvey (6× completion improvement on legal tasks), Netflix, and Spiral. Source

4. Claude Code v2.1.129–v2.1.132 Add Plugin URLs and Session ID Env Vars

Anthropic. Claude Code shipped three back-to-back patch releases on May 6 with new toggles aimed at integrators: a --plugin-url flag that fetches a plugin .zip for the current session, a CLAUDE_CODE_SESSION_ID env var exposed to Bash subprocesses, and a CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN opt-out for the fullscreen renderer. Added env vars also force-enable synchronized terminal output and run package-manager auto-upgrades on Homebrew/WinGet. Source