Product updates, new features, improvements, and fixes as we ship them. Subscribe via RSS to stay updated.
This changelog reflects the current mainline release. Availability in your environment depends on your organization's deployment cadence.
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.
You can now scroll back through previous months on the Usage page to review historical AI spend without leaving the Workbench. A CSV export brings that data into any spreadsheet for deeper analysis or finance reporting. The spend table also filters out system identities so the view reflects only real user activity.
The gateway now routes to Claude Opus 5 and Sonnet 5, Anthropic's latest and most capable models. They are configured as the Claude Code defaults, so developer tooling picks them up automatically without any reconfiguration. All traffic continues to run on private networking within your existing virtual key budgets.
A new Developer (Ultra) tier is now available with a $200 monthly AI spend budget, designed for teams running intensive or high-throughput workloads. It sits above the existing Developer tier and is provisioned through the same standard tier management tooling.
Company admins can now assign user-management bundles from the Settings page, giving specific team members the access they need to manage users without requiring a platform-level change. The assignment flow is built into the existing user settings screen so the rest of the workflow stays the same.
Governance date overrides and lifecycle events now log at the date they actually occurred rather than the date the entry was created. Authorized reviewers can also record a complete past committee review in a single step, capturing votes, decisions, and approval dates at their original times. Audit trails now accurately reflect what happened and when it happened.
The AIS Program tab now includes a built-in document library where you can upload and organize the policies, standards, and evidence your AI program depends on. Documents live alongside your program configuration rather than in a separate system, and each file shows its format in a dedicated column for quick scanning.
You can now mark an AI system as requiring training, and the compliance log will record those training entries automatically. Reviewers get a clear, timestamped history of which systems triggered a training requirement and when, with no manual bookkeeping needed. The flag fits into the existing review workflow and adds no extra steps.
The AI System Inventory now supports a user-assigned inventory number on each system record. Numbers are scoped to the user who sets them, imported values are gated per account, and the field appears in CSV exports alongside existing columns. Teams that maintain asset registries or map AI systems to internal control frameworks can now keep those identifiers in one place.
Two sets of high-severity vulnerabilities were patched this week. CVE-2026-66066 in Active Storage has been remediated, and high-severity advisories GHSA-r28c-9q8g-f849 and GHSA-mh99-v99m-4gvg affecting postcss and brace-expansion have been resolved. No action is required from users.