Claude Code's Enterprise Gateway Grows Up

Claude Code's Gateway Pricing Control Is the Signal Enterprise Admins Have Been Waiting For

If you're running Claude Code across an engineering org through a self-hosted gateway, your spend meter and your developers' /cost readouts have been lying to each other. v2.1.268 fixes that. Admins can now set pricing in gateway.yaml, and signed-in Claude Code clients receive those same rates through managed settings — so what a developer sees in /cost matches what the finance team sees in the spend dashboard.

That alone is worth the upgrade, but the release adds two controls that matter for any org nervous about public exposure of a self-hosted gateway: a startup warning when access_control.allow_cidrs is empty, and a runtime warning the first time a request arrives from a public address. If you deployed a gateway without locking down the CIDR allowlist — and more orgs than will admit it did exactly that — this release is now telling you at startup. Pair that with the new gatewayInternalNetworks managed setting, which lets you restrict /login to your org's own public IPv4 block, and you have a defensible perimeter story for the first time.

What to do: Pull the release, set access_control.allow_cidrs explicitly, and push gatewayInternalNetworks through managed settings before the next all-hands demo where someone on an airport Wi-Fi network tries to log in.

Effort Caps and System Prompt Snapshots: v2.1.267 Tightens Operator Control

v2.1.267 ships maxEffortLevel — a top-level or per-model setting that caps how hard Claude Code can push on any provider, including Bedrock, Vertex, and Foundry. Developers can still dial down from the cap, but they can't exceed it. For tenants where token spend is tracked against project budgets, this is the first native lever that doesn't require a sidecar proxy to enforce.

The --system-prompt-snapshot off flag is less flashy but tactically useful during active prompt engineering: it forces Claude Code to re-read the system prompt on every request instead of replaying the conversation's recorded version. Any team iterating on prompt text in a long session knows the pain of stale-snapshot drift. This makes that iteration loop reliable without having to abandon the session.

What to do: Add maxEffortLevel to your org's managed settings baseline now, before individual teams start configuring per-model overrides in ways that are hard to audit later.

Plugin Eval and Bash Diffs: v2.1.269 Changes the Testing Conversation

v2.1.269 introduces claude plugin eval — a command that runs a plugin's eval suite against Claude Code and returns scored, reproducible results in both JSON and HTML. This matters because "does this plugin work" has until now been a vibe-check: run it, look at the output, shrug. Scored evals with a consistent format means plugin behavior can be gated in CI the same way unit tests are.

The release also adds a diff of files changed by a Bash command to the Bash tool result when bashEditDiffEnabled is set. That sounds minor until you're debugging a multi-step agentic run where a Bash tool silently rewrote three config files. Now the trace tells you exactly what changed, which dramatically shortens the audit loop for any engineering team running Claude Code in an automated pipeline.

The OTEL addition — OTEL_METRICS_INCLUDE_REPOSITORY tagging metrics and events with vcs.* repository attributes — means you can now correlate Claude Code activity with specific repos in your observability stack. For orgs that have asked "which repos is Claude Code touching most?" this is the first native answer.

What to do: Enable bashEditDiffEnabled in your org's managed settings today. Start designing plugin eval gates for any MCP plugins you're shipping to internal users.

Boris Cherny's Take on Code Quality Is a Governance Checklist

Simon Willison quoted Boris Cherny — Claude Code's creator — on what Anthropic actually runs around Claude-written production code: lint rules, tests, Claude-driven end-to-end tests, Claude-powered fuzzers running daily, automated code reviews and security reviews, automated refactoring. Cherny's point is that AI-written code should clear a higher bar than human-written code, not a lower one.

That's a governance posture, not just a quality observation. If Anthropic is running daily fuzzers against their own Claude-generated code, an enterprise org shipping Claude Code to 500 engineers without equivalent gates is accumulating technical and security debt faster than their traditional review cycles can catch it. The teams we've deployed Claude Code with that are ahead of this problem all share one trait: they treated Claude Code rollout as a CI/CD architecture project, not a developer productivity rollout.

What to do: Audit your current Claude Code rollout against Cherny's list. If you're missing fuzzing and automated security review, those are the gaps to close before you scale seat count further.

Purview DLP for Cowork: The Policy Gap Closes in October

The Microsoft 365 Roadmap added Purview Data Loss Prevention for Microsoft Cowork, entering preview in September and targeting GA in October. The integration lets admins apply a single DLP policy across Microsoft 365 Copilot and Cowork — blocking prompts containing sensitive information types, preventing labeled knowledge sources from being used in Cowork, and restricting sensitive prompts from hitting Bing for web search.

If you've been holding Cowork rollout pending a governance story, this is the unlock. The more urgent action is for orgs that already deployed Cowork without DLP coverage: you currently have a policy gap between your M365 Copilot DLP posture and Cowork sessions. That gap closes when this hits GA, but only if you've built the policies in Purview to apply to it.

What to do: Pull your existing Copilot DLP policies in Purview and confirm they're scoped to extend to Cowork. If they aren't, draft the updated scope now so you can push on GA day rather than two weeks after.

---

Watch next week for the Purview DLP for Cowork preview to light up in tenant admin centers — that's when the policy configuration questions become real and urgent for every Cowork-adjacent deployment.