🧭 Claude Sets a World Record in Particle Physics: Nine-Loop N=4 SYM Hexagon Amplitude

Anthropic researchers gave Claude a single instruction — compute the planar hexagon amplitude in N=4 super-Yang-Mills (SYM) theory at nine loops — and it did it. That breaks the previous eight-loop record established by Lance Dixon and Andy Liu in 2023, and it is the most complex symbolic computation of this kind ever completed. The calculation ran for roughly one week across 96 processors at a cost of $1,000–$2,000 per computational method.

What is an amplitude and why does nine loops matter?

Scattering amplitudes describe the probability of quantum particles interacting. N=4 SYM is a maximally supersymmetric quantum field theory that physicists use as an idealised "laboratory" for exploring quantum gravity and string theory. Computing the amplitude at L loops means summing over all Feynman diagrams with L independent momentum integrations — a combinatorial explosion that grows super-exponentially with each additional loop. At nine loops, the raw symbol space (the algebraic object encoding the amplitude's structure) contains millions of terms, making manual computation essentially impossible.

How Claude approached it

• Autonomous loop strategy: Claude was not given a step-by-step algorithm. It selected its own sequence of integrability constraints, bootstrap equations, and consistency checks — pausing to request additional compute when its intermediate representations became too large for working memory.
• Parallelism at scale: Multiple candidate symbol-letter bases were explored simultaneously across the 96-processor cluster, with Claude synthesising results across branches.
• Self-verification: The output passes all known consistency checks for this amplitude class, including collinear limits and multi-Regge kinematics — the standard peer-review criteria for amplitude computations.
• Cost profile: $1,000–$2,000 per method at current API rates. For context, assembling a team of postdocs to attempt this by hand would cost more and likely not converge.

🧭 Claude Plugin Directory Now Accepting Developer Submissions — MCP Connectors, Agent Skills, and Both

Following the Claude Marketplace launch on September 24, Anthropic has opened the Plugin Directory submission portal, letting developers on paid plans submit their own plugins for listing. A plugin can be an MCP connector, an Agent Skill, or a combined package exposing both. Submissions flow through a review queue with a status tracker, and approved plugins surface in the Marketplace's browsable directory with post-launch analytics visible to the publisher.

Submission requirements and review criteria

• Eligibility: Paid plan required (Starter, Team, or Enterprise). Free-tier accounts cannot submit.
• Plugin types accepted: MCP connectors (tool definitions, server manifest), Agent Skills (prompt + tool bundles with an invocation schema), or combined packages where an MCP server also exposes an Agent Skill surface.
• Review criteria: Functionality testing, security scan, policy compliance check (the same operator guidelines that govern API usage), and UX review for clarity of the description and parameter names. Anthropic targets a 5-business-day review SLA.
• Post-approval analytics: Install counts, weekly active users, and error-rate telemetry are available in the Marketplace publisher dashboard — useful for deciding whether to prioritise a connector's maintenance.

🧭 Claude Enterprise Gains Smart Reports Beta — See Which Tasks Your Team Runs Most, and What They Cost

Claude Enterprise has a new analytics dashboard in beta: Smart Reports. It surfaces aggregated, anonymised data on how a team uses Claude — the task types that run most often, token cost by department or project, and repeated prompt patterns that might be worth packaging as shared Skills. The dashboard is opt-in, off by default, and must be enabled by an Enterprise Owner in the Console.

What Smart Reports shows

• Task frequency breakdown: Top task categories (code generation, document drafting, data extraction, Q&A, etc.) ranked by invocation count across the workspace over a rolling 30-day window.
• Cost by group: Token spend broken down by team, project tag, or API key — useful for chargeback reporting or identifying unexpectedly expensive workflows.
• Friction signals: Sessions where users abandoned mid-conversation, retried prompts three or more times, or gave a thumbs-down rating are surfaced as potential quality gaps — actionable inputs for Skills or system-prompt tuning.
• Skill recommendations: If Smart Reports detects a high-frequency prompt pattern that consistently succeeds (low abandonment, high thumbs-up rate), it flags it as a candidate for a shared Skill — with a one-click export to the Skill builder.

Availability constraints

Smart Reports is not available for HIPAA-configured workspaces, zero-data-retention setups, or Team-plan accounts. Because the dashboard aggregates usage data server-side, it cannot function under retention policies that prohibit Anthropic from storing conversation metadata. Enterprise Owners can enable it per-workspace; Admins do not have access to the toggle.