There is a number that reframes everything else in this issue. Dario Amodei told Fortune on Friday that Anthropic saw 80-fold growth in annualized revenue in Q1 2026, crossing a $30 billion run rate from roughly $9 billion at the end of 2025 (fortune.com). The company had planned for 10-fold. The gap between plan and result explains why, in the same week they ran their first developer conference, they signed two compute contracts worth a combined $1.8 billion and 300 megawatts.
Two deals, three days. SpaceX's Colossus 1 data center in Memphis: over 220,000 NVIDIA GPUs drawing more than 300 megawatts (anthropic.com). Akamai Technologies: $1.8 billion over seven years, consumption-based, ramping as capacity comes online later in 2026, largest deal in Akamai's history (bloomberg.com). Akamai shares closed May 8 up 27 percent at $148.38.
The conference landed Wednesday. The SpaceX announcement landed the same morning. By Friday, the compute story had lapped the product story twice over. No model shipped this week. The shortage is the story.
CPO Ami Vora opened "Code with Claude 2026" at 09:00 PT on Wednesday in San Francisco with a line that should have cued the crowd: "This event is about how we are making our products work better for you." No keynote model reveal. What the stage delivered instead was a set of agentic features that Anthropic has apparently been running internally long enough to build confidence in.
The marquee feature is Code Review (claude.com). It dispatches a coordinated team of parallel agents against a pull request. Each agent specializes in a different error category, leaves comments, and proposes fixes. Anthropic said it is already used by every internal team at the company. That is a reasonable thing to claim about a tool you built for yourself; it is also a convenient endorsement.
CI Auto-Fix ships alongside it: Claude files automatic code fixes against failing CI checks on open pull requests. Remote Agents extends the surface further, letting you control your laptop from your phone via Claude. Managed Agents received two structural additions. Multi-agent Orchestration: create coordinated fleets of agents. Outcomes: define a success criterion, then Claude iterates toward it. The SDK surface for both landed the same day in anthropic-sdk-python v0.100.0 and anthropic-sdk-typescript v0.95.0.
The thread running through all of it is delegation. Not just assistance. The review, the fix, the orchestration: these are things a team member does, not a tool. Whether the reliability is there to back that framing is what you find out when you ship something that depends on it.
Two compute announcements in three days. First: SpaceX Colossus 1 in Memphis, more than 220,000 NVIDIA H100, H200, and GB200 GPUs drawing over 300 megawatts (anthropic.com). The visible effect for users: doubled Claude Code rate limits for paid plans, peak-hour caps removed for Pro and Max subscribers, higher Opus API throughput. Anthropic also flagged interest in multi-gigawatt orbital compute capacity as part of the SpaceX arrangement. Second: Akamai, $1.8 billion over seven years (bloomberg.com). Both closed in the same week Amodei disclosed 80-fold Q1 growth and a $30 billion run rate. The company needed 10-fold capacity. It got eight times that.
The week before Code with Claude, Anthropic launched ten pre-built agent templates for financial services: pitchbook creation, KYC file screening, month-end close (fortune.com). The release deepened Microsoft 365 integration across Excel, PowerPoint, Word, and Outlook, and added a data partnership with Moody's. Reported by Fortune and The Register. The financial services push and the developer conference ran in the same week without appearing to be coordinated. They probably are.
On Thursday, the Anthropic Institute published a formal research agenda (anthropic.com). Led by co-founder Jack Clark, now carrying the title Head of Public Benefit, the institute consolidates three existing teams: Frontier Red Team, Societal Impacts, and Economic Research. Four named focus areas: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. The stated mission is to understand and shape the consequences of powerful AI systems before those consequences do the shaping first. The gap between that research agenda and the shipping velocity it is studying is not specified.
Five Claude Code versions between May 6 and May 9: v2.1.131 through v2.1.137 (with gaps at 134 and 135, github.com). The notable additions: a worktree.baseRef setting for controlling branch points in new worktrees (v2.1.133), hook event streaming and deferred hook decisions in claude-agent-sdk-python v0.1.74, and API key redaction in TypeScript SDK debug logs (v0.95.1). The fix list covered MCP routing drops after /clear, OAuth credential races, Windows VSCode activation failures, plan mode incorrectly not blocking file writes, and --resume failing on paths containing underscores.
Mythos was not on the Code with Claude stage. Before the conference, speculation had pointed toward a model reveal. Vora's framing ("this event is about how we are making our products work better for you") was the answer. Code with Claude continues in London on May 19 and in Tokyo on June 10. No model deprecations announced this week.
A 1:1 record of every Anthropic release in the window, 2026-05-05 through 2026-05-09.
claude update or reinstall.CLAUDE_CODE_SESSION_ID to the Bash tool subprocess environment. Fixes fullscreen rendering after sleep/wake cycles, cursor positioning with Indic scripts and emoji, and pasting text that starts with /.claude update or reinstall.Adds worktree.baseRef setting (fresh or head) controlling whether new worktrees branch from origin/<default> or local HEAD. New sandbox.bwrapPath and sandbox.socatPath managed settings for Linux/WSL. A parentSettingsBehavior admin-tier key for SDK managedSettings policy merge control. Hooks now receive the active effort level via effort.level in event JSON and $CLAUDE_EFFORT env var. Memory management improved under pressure.
Fixes: parallel session credential failures after refresh-token race; Edit/Write permission rules incorrectly matching drive roots; Remote Control stop not fully canceling sessions; /effort changing level across concurrent sessions; subagents failing to discover project and plugin skills.
claude update. Set worktree.baseRef in .claude/settings.json.New: settings.autoMode.hard_deny for unconditional classifier blocks in auto mode. CLAUDE_CODE_ENABLE_FEEDBACK_SURVEY_FOR_OTEL re-enables session quality surveys for enterprises using OpenTelemetry.
Fixes: MCP servers, plugins, and claude.ai connectors silently dropping after /clear; login loop from concurrent credential writes overwriting freshly-rotated OAuth tokens; MCP OAuth refresh token races across multiple servers; 400 API error when extended thinking emitted a redacted block after a tool call; --resume and --continue failing on project paths with underscores; plan mode failing to block file writes when a matching Edit(...) allow rule exists. WSL2: image paste from Windows clipboard now works via PowerShell fallback.
claude update. Add settings.autoMode.hard_deny rules to .claude/settings.json.claude update or reinstall.pip install anthropic==0.99.0npm install @anthropic-ai/sdk@0.94.0pip install anthropic==0.100.0npm install @anthropic-ai/sdk@0.95.0include_hook_events option. Deferred hook decisions ("defer" in PreToolUseHookSpecificOutput). Strict MCP config option. Permission context enrichment: adds decision_reason, blocked_path, title, display_name, description fields. updatedToolOutput for post-tool hooks. xhigh effort level for Opus 4.7. Subprocess cleanup on parent exit via atexit handler.pip install claude-agent-sdk==0.1.74pip install claude-agent-sdk==0.1.75api_error_status: int | None to ResultMessage, surfacing HTTP status codes (429, 500, 529) from API failures. Fixes ToolPermissionContext.suggestions deserialization. Bundles Claude CLI 2.1.132.pip install claude-agent-sdk==0.1.76npm install @anthropic-ai/sdk@0.95.1Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.