Anthropic AI Updates: April 23, 2026

1. Anthropic’s Survey of 81,000 Claude Users Maps Productivity Gains Against Displacement Fear

Anthropic. Anthropic’s Economic Research team published results from a survey of roughly 81,000 Claude.ai users that links stated usage patterns to how people feel about AI’s impact on their work. Average self-reported productivity came in at 5.1 out of 7 (“substantially more productive”), with 48% of respondents saying AI expanded their scope of work versus 40% citing raw speedups. Displacement anxiety tracks exposure: the more a worker’s tasks overlap with what models can do, the more they worry about losing their job, with early-career respondents most likely to voice concern even when they report the largest speedups. Source

2. Anthropic Launches a Monthly Economic Index Survey Run Through Claude

Anthropic. The company announced the Anthropic Economic Index Survey, a recurring study conducted via the Anthropic Interviewer that asks Claude users about how they use AI, what they expect, and what they fear. The goal is to build a rolling qualitative corpus that captures AI-driven economic shifts before they show up in conventional labor statistics. The 81,000-user study above is the first cut of this data. Source

3. Anthropic Reframes MCP as the Protocol Layer for Production Agents

Anthropic. A new post on the Claude blog positions Model Context Protocol as the standard layer for agents operating in the cloud — replacing direct API calls and bespoke CLIs with a single transport that handles authentication, discovery, and semantics across clients and deployment environments. The piece codifies server-side design guidance the team has been giving customers: group tools by intent rather than endpoints, support code-style orchestration when the API surface is large, and ship interactive surfaces alongside primitives. On the client side, it recommends progressive disclosure to keep context windows workable when agents pull in many MCP servers. Source

4. Claude Code 2.1.118 Adds Vim Visual Mode, Custom Themes, and MCP Tool Hooks

Anthropic. Releases 2.1.117 and 2.1.118 shipped overnight. 2.1.118 adds vim visual mode (v) and visual-line mode (V) with selection and operator support, a named-themes system driven by JSON files in ~/.claude/themes/ that plugins can contribute to, and a new type: "mcp_tool" option that lets hooks call MCP tools directly. Other additions include a DISABLE_UPDATES env var that kills every update path, WSL inheritance of Windows-side managed settings via wslInheritsWindowsSettings, "$defaults" tokens that merge custom rules into auto-mode lists, and claude plugin tag for cutting release tags with version validation. 2.1.117 adds forked-subagent support for external builds via CLAUDE_CODE_FORK_SUBAGENT=1 and loads agent-frontmatter mcpServers for main-thread sessions launched with --agent. Source