Labs Redesign

- Access Controls: Delete access for Monitors can now be granted separately from write access, so custom roles can edit monitors without being able to delete monitors or incident comments.

• AI Data Plane: Operators can now set a sampling rate for session classification, which reduces classification volume while keeping each sampled session complete. Usage data is still forwarded in full.
• AI Data Plane: The session cache and telemetry buffer now require password authentication by default, so conversation text in your cluster is protected from unauthenticated access. Before upgrading, create a Kubernetes Secret holding the password and reference it in your Helm values; the chart will not install without one.
• AI Signals: AI dimensions now show a short description on hover, so you can see what each dimension represents before grouping or filtering by it.
• AWS Connection: CUR 2.0 connections can now read data files from beside their manifests, so exports you copy to another bucket or prefix ingest successfully. The setting is off by default and available in the connection detail panel and the API.
• Cloud Connections: You can now update an existing connection through the API, including its configuration and credentials, without deleting and recreating it. AWS, Google Cloud, and Azure connections are not yet supported.
• Cloud Connections: AnyCost uploads with cost values of one trillion or more are now rejected with an error that names the field and the limit, so implausible values are caught at upload instead of distorting reports later.
• CloudZero MCP: When a telemetry stream cannot be queried because its keys match no dimension, the error now names the stream and the unmatched keys, so you can fix the stream setup directly.
• Collector: In macOS version 1.1.2, the menu bar status card shows ChatGPT alongside Claude and CloudZero, so you can tell whether an OpenAI outage explains unusual or missing usage. You can also minimize the setup wizard with Cmd-M and return to the same step from the Dock.
• GitHub Connection: GitHub connections, including GitHub Enterprise, can now authenticate with a GitHub App registered under your own account instead of a personal access token, so organizations that prohibit classic tokens can still ingest GitHub costs. Existing token-based connections keep working.
• Monitors: You can now change the cost type on the built-in Account, Service, and AI User monitors from the Edit Monitor drawer, so each one tracks the cost view your team reviews.
• Monitors: Incident APIs now identify whether an incident came from billing or telemetry data and return an hourly cost series for telemetry incidents, and roles with View Incidents access can see those hourly charts, so AI spend incidents can be investigated without broader permissions.
• Single Sign-On: Administrators can now enable Demonstrating Proof of Possession (DPoP) on OpenID Connect integrations that use back-channel authentication, so stolen access tokens cannot be replayed by another client.
• Telemetry Streams: Uploading a CSV to an existing stream now lets you replace whole dates, update matching rows only, or add to existing values, so you can correct data without leaving stale or doubled rows.
• User Settings: The Number format preference now applies to cost values, so currency amounts use the separators and grouping you selected instead of always using US style.
• Views: Anomaly detection, with its threshold and minimum cost impact, can now be turned on or off when editing a View.

- AI Signals: The Anomalies card on Overview now respects every page filter, such as Department, so it lists only anomalies for the people your filters select.

• Analytics: Dashboard links that include a plain numeric filter value, such as a code or ID, now apply that filter when opened, and dimension names ending in a backslash no longer stop your Looker project from loading.
• Anomalies: Large batches of reviewed anomalies now process completely, so notifications for organizations with many reviewed anomalies are no longer interrupted.
• AWS Connection: Payer accounts linked through automatic setup now appear in Settings > Connections as soon as the CloudFormation stack completes, and CUR 2.0 exports delivered as CSV/GZIP now show a clear message that only Parquet is supported instead of a raw error.
• AWS Connection: Trusted Advisor data now loads completely for checks that flag very large numbers of resources, and ingestion continues when one account in a large organization is slow or fails, so recommendations stay current.
• Azure Connection: Connecting an Azure account now retries automatically when Azure briefly throttles requests, so setup no longer fails with a server error.
• Cloud Connections: Rejected credentials and other fixable setup problems on Databricks, Snowflake, Oracle Cloud, and Anthropic connections now show specific remediation steps instead of a generic error, so you can correct credentials or permissions without contacting support.
• Cloud Connections: The Test button now honors "Use fixed IP egress", so credentials restricted to an IP allowlist, such as OpenAI keys, validate successfully. OpenAI connections with very high daily request volumes now ingest reliably instead of timing out, and older AnyCost Bucket connections that previously failed to delete can now be removed.
• CloudZero MCP: Cost, dimension value, and telemetry queries that run longer than 30 seconds now return a task ID instead of a service error, so you can retrieve the result when the query finishes. Connector sessions now last the full duration your administrator configured.
• Collector: The automatic Claude.ai session start option has been removed. The Collector now only reads Claude.ai usage and no longer opens sessions on your behalf. Your other Collector preferences are kept.
• Dimension Studio: Rules that match an empty tag value with an Equals condition now pass validation and publish, and validation errors shown at Publish now point to the correct line in the editor.
• Explorer: Choosing Global View or clearing the View chip now keeps the saved View deselected, so it no longer reapplies itself.
• Labs: After an organization-wide disable, an experiment no longer stays marked as locked by an admin.
• Monitors: Monitors created through the API with custom dimension IDs now evaluate correctly instead of being skipped. Monitors using standard dimensions were not affected.
• Optimize: Kubernetes over-provisioned workload, CPU, and memory recommendations now measure utilization against requests in consistent units, so CPU utilization is no longer understated and memory is no longer read as near zero. Savings estimates on affected recommendations may change as they refresh, and fewer inaccurate recommendations appear.
• Optimize: Recommendations now refresh and the recommendations list now loads for organizations with very large custom dimensions. Each multiple-value cell shows up to 20 values.
• Organization Settings: New organizations created from Microsoft 365 sign-ups are now named after the company tenant instead of "onmicrosoft".
• Resource Details: Resource lookups filtered on metadata keys containing hyphens, such as instance-id, now return results instead of failing.
• Sign-In: Your session now renews while you are actively working, and an expired tab renews itself instead of signing you out of every open tab.
• Snowflake Data Sharing: Adding the Billing Connection dimension to a data share view now saves instead of failing with a "dimensions not found" error.
• Telemetry Streams: Large telemetry payloads that compress well are now accepted during ingestion instead of being rejected.
• User Interface: Warning text and icons are now easier to read in light mode, using a darker amber in place of pale yellow.