Anthropic Pairs a Physics Showcase With Tighter Enterprise Controls for Claude

A physics benchmark and a look at agentic trading

Anthropic's research team published "Yes, Claude can do Nine Loops," describing a nine-loop scattering-amplitude calculation in N=4 super-Yang-Mills theory, a class of computation from theoretical high-energy physics used as a stress test for advanced symbolic and mathematical reasoning. The post frames the result as a capability demonstration rather than a product feature, consistent with Anthropic's practice of using hard, well-defined technical problems to illustrate model reasoning depth.

In a second research post, "Project Swap," Anthropic poses the question of what happens when agents trade on people's behalf, an inquiry into the behavior and implications of autonomous agents making financial decisions for users. No results or specific findings from that inquiry were detailed in the material published.

Opus 5.5 moves into the workplace

Anthropic announced Opus 5.5 for Work, an enterprise-oriented release built on the Opus 5.5 model. It arrives the same day as a cluster of governance and compliance changes across Claude Code and the API, suggesting a coordinated push toward business and enterprise customers rather than an isolated model update.

Claude Code locks down which models run and how

Three consecutive Claude Code releases add administrative controls aimed at organizations managing fleets of developers. Version 2.1.283 introduces availableModelsMatch set to "exact," which restricts an allowed model entry to only the specific version named, so new model releases stay blocked until an administrator explicitly lists them, alongside a deniedModels setting that blocks specific models even when other rules would allow them, and an opt-in gateway hint header (x-claude-code-prompt-id) that lets LLM gateways group the requests belonging to a single user prompt.

Version 2.1.281 extends similar control to AWS deployments: Bedrock upstreams can now use assume_role, so the gateway calls Bedrock through an IAM role obtained via STS, potentially in another AWS account and with one session per developer, and can attach a named Bedrock guardrail that must be applied to all Bedrock upstreams or none. Version 2.1.282 adds a setting letting claude --chrome run alongside an exclusive managed-mcp.json, plus startup, /status, and claude doctor notices that surface telemetry variables in project settings that were silently ignored or that disabled telemetry, giving administrators visibility they previously lacked.

API changes: refusal billing resumes, compliance data narrows

Anthropic's API release notes state that billing is resuming for refusals that arrive before any output when stop_details.category is "bio," "frontier_llm," or "reasoning_extraction" — Anthropic specifies these are the categories where it measures low volumes of false positives, the reasoning it gives for reintroducing billing there; mid-stream refusals were already billed regardless of category.

Separately, the Compliance API's local session endpoints exited beta for Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook. The same release notes disclose that the Compliance API's Activity Feed no longer returns file names, project document names, or artifact titles for file, project-document, and artifact activities — the filename and title fields are now empty or omitted, and this applies even to activities recorded before the change took effect.