› Protime Labs / Signals / Recap
Weekly signals · Anthropic enterprise + Microsoft 365 Copilot
Claude Sonnet 5.5, Admin API Governance, and Copilot Studio Agent Identity
Monday, October 5, 2026
Claude Sonnet 5.5 Is Now the Default — What That Means for Your API Contracts
Claude Sonnet 5.5 landed on September 28 and Anthropic immediately set it as the default Sonnet model on the API. The model carries a 1M-token context window at the same $2/$10 per million token price point as Sonnet 5, with cache reads dropping to $0.20/Mtok. Simon Willison's early benchmarking confirms it beats Sonnet 5 across the board at equivalent or lower effective cost.
Why it matters: If your Claude Code deployment, agentic workflow, or API integration pins to claude-sonnet-* without a full version string, it already flipped. The Claude Code v2.1.284 release confirms claude-sonnet-5-5 is the new default — that change hit developer workstations the same day. The 1M context window also opens up full-repo ingestion patterns that were previously only viable on Opus.
What to do: Audit every service account and workflow that calls a Sonnet alias and verify the model string is pinned to the exact version you validated. Run your eval suite against claude-sonnet-5-5 before assuming the upgrade is free. For teams we've deployed Claude Code to, check that .claude/config files on managed machines specify the model explicitly — the gateway spend limit display added in v2.1.284 will at least surface any cost drift immediately.
Anthropic Admin API Now Has RBAC, Spend Limits, and Per-User Cost Reports
The Anthropic TypeScript SDK v0.130.0 and Python SDK v1.10.0, both shipping September 30, added a substantive block of Admin API surface: Claude Enterprise analytics, workspace-level spend limits, RBAC groups and roles, per-user usage and cost reports, and the Plugins/Plugin Marketplaces admin endpoints.
Why it matters: This is the governance layer enterprise customers have been waiting for. Until now, controlling who could consume what within a Claude Enterprise org required either external rate-limiting infrastructure or accepting a flat account-level view. Per-user cost reports and spend limits mean you can now enforce budget boundaries at the team or project level through the API — not just through dashboard intervention after the fact. The RBAC groups and roles addition is the building block for least-privilege access to Claude capabilities across business units.
What to do: If you're running Claude Enterprise, get the SDK updated and map your current access model to the new RBAC primitives before your next billing cycle. Per-user cost attribution is the single most-requested feature from clients running multi-team deployments — this is the time to define your group taxonomy, not after you've already provisioned 200 users under a flat org.
Claude Code Gets Mods — and a Security Regression Worth Checking
Claude Code v2.1.287 introduced Claude Mods: a plugin system that allows deeper behavioral modification. The built-in "You Should Know" mod deploys a side agent that watches sessions and flags things the user or Claude might miss. It's opt-in and requires first-party sessions with telemetry enabled.
That same release week also shipped v2.1.289, which fixed a security regression from v2.1.288: deny or ask rules on nested parts of compound shell commands were not holding over user-installed mod approvals on managed machines. Read deny rules were also failing to apply to files @-mentioned or opened through symlinks in the IDE.
Why it matters: For managed deployments — particularly on GCC-adjacent or high-sensitivity engineering environments — a window existed where deny rules could be bypassed via symlinked paths or compound shell commands modified by a user-installed mod. v2.1.289 closes both. If you're running a pinned version of Claude Code on managed machines, v2.1.289 is a mandatory update, not an optional one.
What to do: Force-update Claude Code to v2.1.289 on all managed machines now. If you have a Claude Code policy that blocks user-installed mods, verify the allowedProviders managed setting introduced in v2.1.285 is configured correctly — it lets you restrict which API providers and, by extension, which plugin sources are permitted at the org level.
Copilot Studio Agents Get First-Class M365 Identities — and Self-Learning
Two Microsoft 365 Roadmap items published September 30 reframe how Copilot Studio agents participate in the tenant.
Persistent identity for Copilot Studio Agents (GA November 2026) gives each agent its own M365 account: a mailbox, Teams presence, and Office access. Admins govern it through the same policy surface as any user account. This is a material change — an agent is no longer a feature of a human user's session; it's a provisioned principal with its own audit trail.
Self-learning for Copilot Studio Agents adds a feedback loop where the agent analyzes completed runs, identifies repeated tool sequences that could be deterministic workflows, and surfaces recommendations to makers — with expected impact and supporting evidence — before any change is published.
Why it matters: Persistent identity resolves the most awkward governance gap in Copilot Studio today: shared agents running under human user credentials. When the agent has its own identity, Conditional Access policies, Purview audit logs, and eDiscovery all apply cleanly. The self-learning loop is interesting for cost control — collapsing repeated agentic steps into deterministic workflows reduces model calls and the failure surface that comes with them.
What to do: Start drafting your agent identity provisioning policy now, before GA in November. Decide whether agents get assigned licenses through the same IDAM process as human accounts, or through a separate pool. The self-learning feature requires makers to review and approve every recommendation — confirm your governance model assigns that responsibility to a named owner, not a shared mailbox.
Purview Auto-Labeling Capacity Jumps to 500K Files Per Day
Microsoft Purview's auto-labeling scale increase for SharePoint and OneDrive raises the per-tenant-per-day ceiling from 100,000 files to 500,000, targeting GA in November 2026.
Why it matters: For a 20,000-seat tenant with years of ungoverned SharePoint content, the previous 100K/day cap meant months of remediation backlog before Copilot grounding was safe. At 500K/day, full estate labeling becomes a weeks-long project rather than a quarterly one. This is directly upstream of Copilot governance posture — unlabeled files are unprotected files, and Copilot can surface them regardless of sensitivity.
What to do: If you've been deferring auto-labeling policy deployment because the throughput math didn't work at scale, revisit the project scope now. November GA is close enough to start scoping the simulation run today.
Next week, watch the Anthropic Frontier Red Team's binary exploitation benchmark — Claude Mythos Preview crossing the threshold on full control flow hijacks is a signal the security community will move on quickly, and your AI acceptable-use policy may need to catch up.