Claude Enterprise Adds Model-Level Entitlements, Spend Alerts, and Richer Admin Analytics
Anthropic has rolled out richer admin analytics, model-level entitlements, and spend alerts for Claude Enterprise. Admins get deeper visibility into how teams use Claude, the ability to gate access model by model, and proactive warnings before costs run past budget.
Model-level entitlements are the quiet standout. As the lineup fans out across Opus 5, Sonnet 5, and Haiku tiers — each with its own price and capability profile — procurement and platform teams want to say “this group gets Opus, that group gets Sonnet” without policing it by hand. Pair that with spend alerts and you get the cost-governance layer big buyers ask for before they scale seats.
Claude for Government Opens Beta Access Requests
Following this week’s launch of Claude Code and Cowork inside a FedRAMP High environment, Claude for Government is now in beta with new customers able to request access. Agencies that want Claude running against local files, memos, and RFPs on managed devices can get in the queue, and Anthropic stays the contracted and billing party — no separate cloud-provider relationship required to start.
It’s the natural next step after the platform work landed: build the authorized, audit-logged environment first, then open the door to onboard agencies. Expect the government track to keep pulling security and compliance features forward for everyone else.
Claude Code 2.1.226 Ships Spend Limits, Workspace Trust, and Remote Control by Name
Claude Code 2.1.226 (August 8) adds gateway spend-limit support so you get usage warnings as you approach a cap, workspace trust prompts for claude agents, the ability to start Remote Control sessions by name, and improved photo handling. Under the hood it fixes a stack of OAuth, session, history, UI, and headless-session bugs.
The theme is daily-driver polish: fewer surprise bills, clearer prompts before an agent touches an untrusted workspace, and steadier long-running and headless sessions. None of it is flashy, all of it is the stuff you feel when you live in the terminal.
Managed Agents Can Now Load Skills Straight From a GitHub Repo
Claude Managed Agents sessions can now pull skills from a GitHub repository: point a session at a repo and any skills in the root .claude/skills directory are discovered automatically at session start. No manual upload, no per-session copying — your skills live in version control and travel with the codebase.
This is a meaningful workflow unlock. Skills become a reviewable, versioned artifact in the same repo as the code they operate on, so a team can PR a new skill, roll it back, and know every agent session is running the same definition. It nudges skills from ad-hoc snippets toward proper software you ship and maintain.
Cowork on Mobile and Web: Approve an Agent From Your Phone
Cowork, Claude’s agent for non-developers, now spans mobile and web as well as desktop. Sessions and files follow you across devices, with background work, scheduled tasks, shared chats and projects, and mobile approvals so you can green-light an agent action from your phone.
The approval-on-mobile piece is the one to watch. It keeps a human in the loop even when the agent runs while you’re away from your desk — kick off a research or file task at your machine, approve the next step on the walk to lunch, check the result on the web later. Cowork is quietly graduating from a desktop utility into an ambient assistant.
Claude’s Connectors Directory Passes 950 MCP Servers
Claude now lists more than 950 MCP servers in its connectors directory, used by millions of people every day. The catalog keeps swelling as vendors ship official connectors on the back of the new 2026-07-28 spec and its stateless, serverless-friendly core.
Nearing four figures, the directory is less a novelty and more the default integration surface for Claude. For teams, the practical question has flipped from “is there a connector” to “which of these do we trust” — which is exactly why the enterprise-managed auth and plugin-scanning features from earlier this week showed up when they did.
Claude Is Being Rebuilt for the Org Chart
Read today’s items together and a pattern falls out: model-level entitlements, spend alerts, gateway limits in Claude Code, government beta onboarding, versioned skills in GitHub. Every one of them answers a question a buyer asks, not a question a user asks. Who gets which model? What will this cost, and can I be warned first? How do I audit and roll back what my agents can do? That’s the vocabulary of procurement, platform teams, and compliance — the people who decide whether a tool becomes standard-issue across ten thousand seats.
The Sonnet 5 price change on September 1 ($2/$10 rising to $3/$15) is the pressure that makes all of this land: as usage scales and costs climb, the org needs governance, not just capability. Anthropic is answering with the boring, essential plumbing — entitlements, alerts, audit, versioning. Reportedly steering toward a public listing at a massive valuation, the company doesn’t win the largest accounts on model quality alone; it wins them by being governable. The flashy launches get the headlines, but weeks like this — when the news is mostly controls — are what turn a clever tool into infrastructure.