July at Claude: From Smart Chat to an Employee You Showed Once

In July, Anthropic released multiple updates that collectively push Claude toward becoming an agent that learns from a single demonstration. The most underrated feature, Record a Skill, launched on July 21, allows users to record their screen and voice commentary while performing a task, and Claude builds a reusable skill from that demonstration. It is available on Pro, Max, and Team plans, only on desktop (Windows and macOS), and lives inside the Cowork interface. On July 24, Anthropic launched Opus 5, which includes an effort regulator with three levels (low, medium, high) that controls reasoning token volume and thus cost. Opus 5 pricing remains the same as Opus 4.8: $5 per million input tokens and $25 per million output tokens, with a context window of up to 1 million tokens and maximum output of 128K tokens. The effort regulator affects reasoning tokens, which are billed as output. The author advises using low for cheap errors (paraphrasing, JSON extraction) and high only for expensive errors (irreversible migrations, financial calculations). Other July updates include a voice mode that now supports Opus and Sonnet, 11 languages, and real actions in Gmail and Slack; MCP specification support for stateless serverless, interactive UI in responses, OAuth/OIDC authorization, and observability dashboards; published artifacts can now access live MCP connectors; managed agents get effort level settings and webhooks for memory and environment events. The author recommends adopting Record a Skill for repetitive manual tasks, the effort regulator set to low by default for cheap-error tasks, and logging which model actually responds to catch silent fallback. The overall theme of July is that Anthropic is no longer selling a smart model but assembling an employee you show once: Record a Skill teaches, effort regulator controls cost, and MCP and voice give it access to systems.