Anthropic AI Updates: September 26, 2026

1. Claude Computed a Nine-Loop Amplitude That Humans Had Only Reached at Eight

Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used Fable 5.1 inside Claude Science to compute a nine-loop scattering amplitude in N=4 super Yang-Mills theory, a calculation previously carried only to eight loops. Claude worked largely autonomously in Python and SymPy, following the established “bootstrap” constraint-satisfaction recipe; the bootstrap run cost about $100, equivalent to 96 CPUs for a week, and both approaches together came to one or two thousand dollars. The result was independently checked by Lance Dixon at SLAC and computed concurrently by a Chinese team, and the authors stress that Claude used known methods with more compute rather than new physics. Source

2. Anthropic Opened a Submission Portal for the Claude Plugin Directory

Anthropic launched a portal that lets developers on paid Claude plans submit plugins to the Claude directory, either as a single remote MCP connector or as a GitHub-hosted bundle of MCP servers and Agent Skills. Submissions are auto-validated and safety-scanned on arrival, developers can track review status and recommended changes, choose when to publish, and see post-launch analytics covering installs by product surface, versions, listing views, and search discovery. The portal supports MCP Apps for interactive UI and Enterprise Managed Auth, and a unified discovery experience across Claude and Claude Code is due in the coming weeks. Source

3. Claude Code 2.1.283 Gives Admins Exact-Match Model Allowlists and a Denylist

Claude Code 2.1.283 added an availableModelsMatch managed setting that, set to "exact", makes an availableModels entry allow only the version it names so new model releases stay blocked until explicitly listed. A new deniedModels managed setting blocks specific models even when availableModels would permit them. The release also added /doctor prompt-audit (also /checkup prompt-audit), which scans CLAUDE.md files, skills, agents, and commands for prompting patterns written for older models. Source

4. Gateway Prompt IDs, Load-Test Mode, and Richer OpenTelemetry Spans

The same Claude Code release added an x-claude-code-prompt-id hint header, opt-in via CLAUDE_CODE_GATEWAY_HINT_HEADERS=1, so LLM gateways can group every request serving one user prompt. The Claude apps gateway gained an opt-in load_test_mode block that builds and signs requests without sending them upstream and returns a canned reply, and MCP tool, WebFetch, and WebSearch outputs now appear in the tool.output OpenTelemetry span event when OTEL_LOG_TOOL_CONTENT=1. On the Claude Tag side, admins get a “Channels Claude can search” setting that limits Slack search to public channels Claude has been added to. Source