Inference Hooks Arrive in Beta for Enterprise Compliance Teams
Anthropic has rolled out inference hooks in beta for Enterprise, giving compliance and security teams a real-time control point for data-loss prevention. The hooks inspect prompts and tool calls before they reach the model — and they span the whole surface: chat, Claude Code, Cowork, and more.
It’s a meaningful shift in where governance lives. Instead of bolting policy checks onto outputs after the fact, regulated buyers can now enforce DLP rules at the inference boundary itself, blocking or redacting sensitive data mid-flight. For industries like aerospace, defense, and finance — where what leaves the building matters as much as what the model says — that’s the kind of control that unblocks deployments.
Claude Code’s 50% Usage Boost Extended Through August 19
Anthropic has extended the temporary 50% weekly usage boost for Claude Code subscribers through August 19, 2026. The extra headroom applies on top of standard weekly limits, giving heavy agentic-coding users more runway before caps bite.
The extension lands the same week Claude Code made Opus 5 its default Opus model — a nudge, essentially, to spend the boosted budget on the newer, cheaper-per-task flagship. If your team runs long agent sessions, it’s free rope for two more weeks.
Claude Code v2.1.223 Adds Org-Wide Marketplace Controls and Subagent Guardrails
Today’s v2.1.223 release is aimed squarely at teams. Admins can now add owner-wildcard entries (owner/*) to marketplace settings to allow or block every repo under a GitHub org in one line — a big quality-of-life win for locking down which plugins and skills a fleet can install.
There’s also a new warning when a workflow agent requests a subagent model that’s restricted, so misconfigured multi-agent runs fail loud instead of silent. And cloud sessions now surface a /teleport hint showing how to continue the same session locally. It builds on Monday’s v2.1.221, which shipped the Focus view (Ctrl+Alt+F) that hides tool activity behind an expandable per-turn summary.
Claude.ai Connectors Now Auto-Available in Claude Code — and M365 Gets Write Tools
Any connector you’ve configured for Claude.ai — Gmail, Calendar, Slack, and the rest — is now automatically available inside Claude Code, including within Artifacts, with no separate setup. One OAuth grant, every surface. It’s the kind of plumbing change that quietly removes a whole category of “why can’t Code see my tools?” friction.
Alongside it, the Microsoft 365 connector gained write tools that run server-side on a schedule with no device online. That rides the broader MCP 2026-07-28 spec — stateless core, stronger OAuth/OIDC, and versioned extensions for Apps and Tasks — which Claude now supports across the board.
Cognizant Becomes a Global Premier Partner, Crosses 30,000 Claude-Trained Staff
Cognizant has expanded its partnership with Anthropic to become one of a handful of Global Premier Partners in the Claude Partner Network. As part of it, the firm says more than 30,000 associates have completed Claude training under a new “Frontier Certified” workforce model.
The proof points are the interesting part: Cognizant is embedding Claude across manufacturing, life-sciences, and insurance client systems, citing an AI contract-intelligence deployment that cut review time up to 40% while lifting extraction accuracy above 88%. It’s another data point in Anthropic’s systems-integrator flywheel — distribution through the big consultancies rather than direct sales alone.
Opus 5’s ARC-AGI-3 Score Keeps the HN Debate Going
Two weeks after launch, Claude Opus 5 is still driving discussion — the original Hacker News thread hit 1,378 points and 746 comments. The benchmark that won’t die down is ARC-AGI-3, where Opus 5 posted 30.2%, roughly three times the next-best model.
That eye-popping gap is exactly why some commenters are pushing back, raising the usual question of possible training-data contamination on a benchmark that’s meant to resist memorization. Skepticism aside, the broader read holds: Opus 5 tops Frontier-Bench and trails Fable 5 on CursorBench 3.2 by half a percent — at roughly half the cost. The portfolio strategy is working.
The Enterprise Moat Is Being Built at the Inference Boundary
Look at today’s releases together and a theme snaps into focus. Inference hooks for DLP, org-wide marketplace wildcards, subagent-model guardrails, connectors that flow through a single grant — none of these make the model smarter. They make it governable. Anthropic keeps shipping the controls that let a security team say yes.
That matters more than a benchmark right now, with the company reportedly on track for its first profitable quarter against a roughly $47B run-rate and a confidential IPO filing on the table. Revenue at that scale doesn’t come from hobbyists; it comes from regulated enterprises that won’t deploy without a policy layer between their data and the model. The frontier labs are converging on capability. The durable edge is increasingly about who makes the agent safe to point at real systems — and today, that’s where Anthropic spent its shipping budget.