AI Briefing
KO

Anthropic Improves claude.ai and Desktop App Speed by 3x with Claude Tag

·2026.09.23 09:00

Key point

During a two-week sprint, over 3,000 changes were merged to optimize performance without any customer incidents.

1 / 3

Details

Anthropic used its internal research model Claude Tag to improve the core user experience speed of claude.ai and the desktop app by approximately 3x. During the two-week sprint from August 13 to 27, 2026, over 3,000 changes were merged, with 0 customer-facing incidents or rollbacks during the process.

Key Performance Metric Improvements

A geometric mean of 13 metrics showed a 3.1x performance improvement. Key metric changes are as follows:

  • claude.ai web new load: 3,085ms → 550ms (5.6x improvement)
  • Desktop app cold start: 6,310ms → 3,328ms (1.9x improvement)
  • Claude Cowork cloud message send: 928ms → 48ms (19x improvement)
  • Claude Code desktop message send: 250ms → 52ms (4.8x improvement)

Overall, it is estimated that tens of thousands of hours of user wait time are saved daily.

Optimization Methodology and Tools

Datadog MCP server and deterministic metrics (Valgrind, node --predictable) were utilized for performance bottleneck exploration and benchmark construction. To remove noise from wall-clock time, metrics such as JS instruction count and React commit count were measured.

  • Static Composer Introduction: Static HTML was baked in before React initialization, reducing the time to interactive from 4.00 seconds to 0.33 seconds.
  • Streaming Renderer Optimization: A budget of 8.33ms per frame was set to achieve 120fps, applying memoization of completed blocks and worker offloading. Main thread blocking time during long answer streaming decreased from approximately 750ms to 200ms.
  • Automated Guardrails: PRs fail if benchmark numbers rise, with gradual rollouts via Feature Flags (employees → 1% → full) and an automated review system in operation.

Human Oversight and Future Plans

PRs generated by Claude require approval from at least one human, while human team members provided direction through goal setting and trade-off decisions. Additional improvements for 95th percentile performance and very long conversation scenarios are currently planned, and follow-up posts regarding upstream contributions to Electron, Chromium, etc., are being prepared.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.