Claude leads 26% of Anthropic's own R&D, Financial Advisors plugin launches, Smart Reports in beta

Anthropic disclosed on September 18 that Claude now leads 26% of the company's model research and development — meaning Claude can complete most of a given task end-to-end from a high-level prompt while remaining under human supervision. As recently as February, that figure was zero. About 90% of all Anthropic R&D now involves collaboration with Claude, meaning the model does large chunks of work under close human direction. Anthropic said roughly 30,000 agents are doing research and engineering work internally as of August. The company published the figures explicitly to warn about the pace of AI self-acceleration: 'Models accelerating their own development could make it more challenging for humans to understand or control these systems.' Anthropic urged other AI developers to share equivalent metrics using a public methodology so the numbers can be compared across labs over time. Anthropic also announced it is embedding external third-party evaluators within the company to independently monitor safety efforts.

Anthropic published a detailed engineering blog post on how its internal infrastructure team scaled continuous integration to handle the load generated by Claude Code agents. With Claude now leading or collaborating on most R&D, CI pipelines are seeing far more commits per day than any human team could generate — the post covers the test impact analysis system Anthropic built to selectively run only the tests most likely to be affected by a given change, reducing CI wall time by over 60% without meaningful coverage loss. The techniques are directly applicable to any team running agentic coding at scale.

Anthropic unveiled Claude for Financial Advisors at the Future Proof Festival in Huntington Beach on September 15. The product connects to custodians (BlackRock, Schwab, Vanguard), portfolio platforms, CRMs, and financial planning tools through a unified Claude Cowork plugin, streamlining advisor workflows across research, proposals, compliance documentation, and client communication. Dynasty Financial Partners simultaneously unveiled Dynasty AI 3.0, powered by Claude. The plugin is available today in the Cowork plugin browser. Enterprise plans are recommended for registered investment advisers because they include the audit logs required for regulatory recordkeeping. Firms requesting a new Enterprise license before the end of September receive a one-time usage credit.

Claude for Small Business expanded significantly on September 15 with 43 new workflows and 27 new connectors covering commerce (Shopify, Square, Stripe), CRM (Salesforce, HubSpot), HR and payroll (Gusto, Rippling), accounting (Xero, QuickBooks), communications (Zoom, Slack), and automation (Zapier). The update pushes Claude beyond back-office tasks into lead generation, proposals, client reporting, and pipeline management for SMBs. Training programs for small business owners are also new — structured onboarding covering how to deploy workflows across the expanded connector set.

Smart Reports entered beta for Claude Enterprise, giving admins AI-generated analysis of how their team uses Claude: what work is getting done, what it costs, where sessions run into friction, and which prompts or workflows are candidates for shared skills. Reports are generated automatically on a weekly cadence and surfaced in the admin analytics dashboard. Separately, the same underlying usage analytics features are rolling out to Pro and Max plan users on web, desktop, and mobile with no opt-in required — Team and Free plans will follow, with Enterprise admins notified at least 30 days before changes affect their orgs.

Claude Code's September mid-cycle release adds fast mode in remote sessions (cloud and self-hosted runners) — the host's fast-mode setting or typing /fast in the session applies where the org allows it. New gateway hint request headers (opt in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1) send x-claude-code-request-class, x-claude-code-agent-type, x-claude-code-prev-tool-durations, x-claude-code-compaction, and x-claude-code-context-compacted to LLM gateways for routing and observability. The /config panel gains mouse support in fullscreen mode — scroll with the wheel, click a setting's value to change it. Per-command allowed_domains can now be set in Bash, PowerShell, and Monitor in auto mode with sandboxing.

Turn-scoped system messages are now in beta on the Claude API (mid-conversation-system-clear-at-2026-08-21 header). Set clear_at: 'next_user_message' on a mid-conversation role: 'system' message and it renders for the current turn only, then stays in history at no additional token cost on subsequent turns. Per-turn reminders no longer accumulate tokens across a long conversation, do not invalidate the prompt cache, and do not interfere with thinking blocks. Particularly useful for agentic workflows that need to inject instructions at specific points without polluting the running context.