Changelog

The Workbench usage report now captures activity from Claude Code working sessions, including parallel mode, and attributes it correctly alongside chat and API usage. Organizations get a more complete picture of AI consumption across developer tooling, not just the browser interface. The session sample in the report's About section also documents the counting method so the numbers are auditable.

When a user already holds a Swept grant through their identity provider group, company admins can now assign that same grant to them directly. This makes permission management more flexible for individual overrides without waiting for group membership changes to propagate. A clarifying banner explains the source of an existing grant so admins always know how access was granted.

Long numeric values in PDF usage reports now scale down to fit their cells rather than wrapping across lines. Hero bar charts draw bars before placing their labels, so labels on narrow bars no longer overlap neighboring bars. Both fixes improve readability when usage reports are shared with finance teams or exported for leadership review.

Spend data for requests routed through Google Vertex AI now uses the correct per-model pricing instead of falling back to a generic default. The spend log endpoint was also fixed to return complete data in production, so the Usage page and monthly reports reflect accurate totals across all models and users.

Workbench now produces a monthly AI usage report covering every model and every user the gateway served during the period. Reports include cited spend figures and recommendations, and arrive as a clean PDF ready to share with finance or leadership. Coverage spans all gateway users, not only those with a Workbench seat, so nothing is left out of the picture.

CVE-2026-9358 has been patched across the platform. Automated monthly external vulnerability scans are now part of the security workflow, producing a consistent signal that feeds into compliance reporting. Google Cloud Security Command Center findings are also now connected to the compliance dashboard, giving security teams a unified view of infrastructure findings alongside AI governance data.

All Gemini models now run through Google Vertex AI using Application Default Credentials, which means no API keys to rotate and full GCP IAM governance over every model call. As part of the same infrastructure update, Gemini 3.5 Transcribe is now available for audio note transcription through the AI gateway. Both changes apply automatically to existing deployments.

Cloud Lock now connects to Microsoft Entra ID over OIDC, so organizations already using Entra for identity management can extend their existing SSO policies to AI access without running a separate credential store. Users authenticate through Entra, and access to the AI environment follows the same groups and conditional-access rules already in place.

Admins can now assign an AI spend cap to individual users rather than relying on a single org-wide limit. Every time a cap is created or changed, the system requires a written reason, giving you a clear audit trail of who adjusted what and why. This makes it straightforward to enforce budget policies across teams without losing accountability.

Every prompt sent through the AI gateway and every model response is now logged to BigQuery, with a 180-day retention window for compliance and audit purposes. Organizations can trace exactly what was sent to any model, by whom, and when, without touching application code. Automated monitoring alerts fire if the log pipeline goes quiet, so gaps are caught before they affect audit coverage.