Skip to main content

Anyscale CLI and SDK release notes

View Markdown

Anyscale CLI and SDK release notes

This page documents changes to the Anyscale CLI and SDK, including new features, behavior changes, and bug fixes.

For installation and configuration, see Get started with the Anyscale CLI and Get started with the Anyscale SDK.

Version 0.26.102 release notes​

Version 0.26.102 is available as of May 19, 2026.

Behavior changes​

CLI​

  • anyscale logs job now defaults to a per-job log download endpoint. Downloads complete faster on large clusters because the new endpoint avoids scanning all cluster logs. Cluster-level filters such as --head-only, --worker-only, --node-ip, --instance-id, and glob patterns fall back to the previous cluster-log path. --ttl is ignored on the job-scoped path and surfaces a warning.

CLI and SDK​

  • Submitting a job with both timeout_s and job_queue_config no longer returns an HTTP 400 error. Per-job timeouts apply to jobs running in a job queue, and the job queue cluster remains alive after a single timed-out job.

Bug fixes​

SDK​

  • anyscale.schedule.apply() now forwards job_config.tags to the created schedule so spawned jobs are tagged, and anyscale.schedule.status() returns the tag dict instead of None.
  • ComputeConfig.to_dict() and from_dict() now preserve explicit CPU=0 and GPU=0 resource pinning on head and worker nodes. Previously, explicit 0 resources were stripped during round-trip, which converted an explicitly unschedulable head node into a schedulable one. The load path also collapses the backend-injected unschedulable default back to resources=None so user-authored YAML stays clean.

Version 0.26.101 release notes​

Version 0.26.101 is available as of May 11, 2026.

New features​

CLI​

  • Added optional --cloud-resource-name to anyscale cloud verify to verify a single named cloud resource. When unset the function verifies every cloud resource sequentially.

Bug fixes​

CLI​

  • anyscale cloud verify now generates declarative CPU and memory compute configs for verification on Kubernetes-backed cloud resources, fixing instance type not found errors. Operator health checks also retry transient connection failures after port-forwarding.
  • anyscale cloud verify now runs static verification across all resources before functional verification to surface configuration errors earlier. Service verification is skipped on non-primary cloud resources with a warning.
  • anyscale job-queue list --cluster-status now applies the filter instead of silently ignoring the filter.

SDK​

  • anyscale.job.status(...).config.timeout_s now returns the timeout value when a job is submitted with --timeout-s. The field previously always returned None.

Version 0.26.100 release notes​

Version 0.26.100 is available as of May 4, 2026.

Behavior changes​

CLI and SDK​

  • anyscale service deploy --versions now requires Anyscale to enable the enable-multi-version-services feature flag for your organization. Anyscale enforces this flag server-side, so the requirement applies to all CLI and SDK versions, not just 0.26.100. Version 0.26.100 adds the translated CLI error message Multi-version services (--versions) are not enabled for your organization. Please contact Anyscale support to request access.. Operations on existing multi-version services, including weight updates, restarts, in-place upgrades, and rollbacks, continue to work without the feature flag. See Multi-version services move to private beta for the platform announcement.

Version 0.26.99 release notes​

Version 0.26.99 is available as of April 29, 2026.

New features​

Cloud setup​

  • You can now set redis_endpoint under kubernetes_config in the cloud resources YAML to configure a Kubernetes-managed Redis endpoint to enable head node fault tolerance for Anyscale services on cloud resources deployed with Kubernetes.

Behavior changes​

SDK​

  • list_clouds now returns DecoratedcloudListResponse instead of CloudListResponse. The decorated response includes an optional creator field identifying who deployed the cloud. CloudsQuery also accepts new creator_id and state_filter parameters to filter clouds by deployer or state.

Bug fixes​

CLI and SDK​

  • Corrected help text and docstrings across the auth, cloud, compute-config, job, project, schedule, service, service-account, and workspace command groups.

Version 0.26.98 release notes​

Version 0.26.98 is available as of April 28, 2026.

Version 0.26.97 release notes​

Version 0.26.97 is available as of April 27, 2026.

Bug fixes​

  • Fixed anyscale cloud verify reporting NGINX Ingress: FAILED on Kubernetes clouds that use gateway-based networking. The output now includes a single Networking check that reads the operator config and verifies the configured gateway when gateway support is enabled, or the NGINX ingress controller otherwise. Invalid or unreadable operator config now fails the check instead of skipping it.

Version 0.26.96 release notes​

Version 0.26.96 is available as of April 22, 2026.

New features​

  • Added anyscale skills CLI commands and anyscale.skills SDK module to download versioned skill bundles and install them into Claude Code and Cursor. See Anyscale agent skills.

Cloud setup​

  • You can now set cluster_instance_profile_id under aws_config in the cloud resources YAML to specify an AWS instance profile ARN when the profile name differs from the cluster IAM role name. Omitting the field preserves existing behavior.

Behavior changes​

CLI​

  • anyscale cloud resource delete now tears down Anyscale-managed cloud resources, including AWS CloudFormation stacks and Google Cloud infrastructure. Previously, the command returned an error for managed resources. The CLI warns that teardown is irreversible and rejects primary or ambiguous resource targets.

CLI and SDK​

  • TERMINATED is now a distinct job state. anyscale job list --v2 accepts --state TERMINATED as a filter, and the SDK's JobState enum exposes TERMINATED separately from FAILED. Previously, the CLI and SDK reported terminated jobs as FAILED.

Version 0.26.95 release notes​

Version 0.26.95 is available as of April 15, 2026.

Bug fixes​

  • Fixed anyscale job list and job.list() ignoring the --cloud filter. The call now correctly returns only jobs from the specified cloud.
  • Log output downloaded from S3 now decodes as UTF-8. Unicode characters display correctly and non-UTF-8 bytes are replaced instead of raising an error.

Version 0.26.94 release notes​

Version 0.26.94 is available as of April 14, 2026.

Version 0.26.93 release notes​

Version 0.26.93 is available as of April 7, 2026.

Bug fixes​

CLI​

  • anyscale cloud create-cloud-resource now automatically skips verification when adding a Kubernetes compute stack and logs a message to run verification separately after installing the Anyscale Operator.
  • Fixed the KubernetesConfig field name from anyscale_operator_iam_id to anyscale_operator_iam_identity, matching the underlying API.
  • The CLI now strips leading and trailing whitespace from image URIs before validation, preventing errors when a URI is pasted with leading or trailing spaces.
  • Fixed anyscale job submit --help to consistently show the -f / --config-file flag in all usage examples.

Version 0.26.92 release notes​

Version 0.26.92 is available as of March 25, 2026.

Bug fixes​

CLI​

  • anyscale cloud get --name now returns a clean "Cloud not found." error instead of a raw traceback when the specified cloud doesn't exist.

SDK​

  • JobSubmissionClient now correctly scopes project resolution to a specific cloud using the cloud kwarg or ANYSCALE_CLOUD environment variable. This fixes a silent error resulting in jobs targeting the wrong project when the same project name exists in multiple clouds.

Version 0.26.91 release notes​

Version 0.26.91 is available as of March 17, 2026.

Bug fixes​

SDK​

  • Fixed asyncio.run() failures when using the SDK inside an already-running event loop, such as in Jupyter notebooks.

Version 0.26.90 release notes​

Version 0.26.90 is available as of March 9, 2026.

New features​

CLI​

  • Added --migrate-dm-to-im to anyscale cloud update to migrate eligible cloud resources deployed on Google Cloud from Deployment Manager to Infrastructure Manager.

Version 0.26.89 release notes​

Version 0.26.89 is available as of March 4, 2026.

Bug fixes​

  • When the image builder is hibernated, the CLI now surfaces a clear user-facing error message.

Version 0.26.88 release notes​

Version 0.26.88 is available as of March 2, 2026.

Behavior changes​

CLI​

  • Workspace create, start, and update now surface a single clear error message for compute config validation failures instead of a stack trace.
  • EFS mount target verification now shows a clearer warning when a mount target exists in the expected subnet but uses a different security group, with guidance to allow inbound NFS (port 2049) from worker nodes.
  • Performance improvements for anyscale schedule list and anyscale schedule status.
  • Compute config validation is stricter: memory must include a unit suffix (for example, 8Gi), empty required_resources: {} is rejected, and you can't combine instance_type with required_resources. Logical resources cannot exceed physical required_resources. Out-of-range values (for example, CPU > 512, memory > 24 TiB) are rejected with explicit errors.
  • When you add labels in required_labels that don't affect VM instance selection, the CLI emits a warning.

SDK​

Version 0.26.87 release notes​

Version 0.26.87 is available as of February 19, 2026.

Behavior changes​

CLI​

  • You can now delete clouds backed by Google Cloud VMs provisioned with Infrastructure Manager as well as Deployment Manager. anyscale cloud delete tears down resources correctly for both provisioning paths.
  • anyscale cloud update no longer updates the IAM external ID policy when the resulting policy matches the existing one. This avoids failures when you don't have permission to modify IAM policies.

Bug fixes​

  • Fixed cloud deletion for AWS and Google Cloud VM stacks so that the CLI now cleans up all managed cloud resources attached to a cloud when you run anyscale cloud delete. Previously, some managed resources in multi-resource or multi-region setups sometimes remained.

Version 0.26.86 release notes​

Version 0.26.86 is available as of February 17, 2026.

Behavior changes​

CLI​

  • anyscale cloud register --skip-verifications now works without local cloud provider credentials for VM clouds (AWS and Google Cloud). The backend looks up derived values (for example, availability zones and EFS mount targets) instead of requiring local creds.
  • Google Cloud VM cloud setup and cloud resource setup now provision resources through Google Cloud Infrastructure Manager instead of Deployment Manager. You can track progress in the Google Cloud console.
  • When adding VM resources to a cloud in the same region on Google Cloud, firewall policy names now use a unique deployment name to avoid conflicts. Workload Identity Federation setup no longer creates duplicate pools and returns the correct pool path.

Bug fixes​

  • Fixed anyscale cloud setup for Azure after breaking changes in the Azure management package. Install the CLI with the Azure extra (pip install anyscale[azure]) for Azure cloud setup. Azure cloud setup no longer incorrectly requires availability zones.
  • Fixed cloud resource setup bugs that could cause 409 conflicts and incorrect IAM bindings when creating multiple Google Cloud resources in the same region.

Version 0.26.85 release notes​

Version 0.26.85 is available as of February 11, 2026.

New features​

CLI​

  • Added --sort to anyscale schedule list for v2 schedules. Supported fields: id, name, created_at, next_trigger_time.

Behavior changes​

CLI​

SDK​

  • Removed legacy Service SDK methods: get_service, list_services, rollback_service, rollout_service, and terminate_service. Use anyscale.service.status(), anyscale.service.list(), anyscale.service.rollback(), anyscale.service.deploy(), and anyscale.service.terminate() instead. See Methods removed in version 0.26.85 for migration guidance.
  • Removed legacy cloud SDK methods: get_cloud, get_default_cloud, and search_clouds. Use anyscale.cloud.get, anyscale.cloud.get_default, and anyscale.cloud.list instead. See Methods removed in version 0.26.85 for migration guidance.
  • Removed legacy SSO SDK methods: partial_update_organization, upsert_sso_config, and upsert_test_sso_config. There is no replacement in the modern SDK.
  • When you submit a job without setting max_retries, the CLI and SDK emit a warning. The default is 1 in this release, but Anyscale plans to change the default to 0 in a later release. Set max_retries explicitly for consistent behavior.

Compute​

  • VM-backed clouds now reject TPU shapes with a clear error. Use a Kubernetes (GKE) cloud for TPU workloads, or use GPU resources instead.

Bug fixes​

  • Fixed resource fetch when uploading files on Anyscale-hosted clouds where the CLI can't resolve resource ID from name; the upload flow now proceeds without requiring the resource name.

Version 0.26.84 release notes​

Version 0.26.84 is available as of February 6, 2026.

Bug fixes​

  • Fixes an error for cloud-isolated projects in legacy Anyscale Connect.

Version 0.26.83 release notes​

Version 0.26.83 is available as of February 4, 2026.

New features​

SDK​

  • Added anyscale.job.delete() to permanently delete a job and all associated job runs. The job must be in a terminal state (SUCCEEDED or FAILED). Supports name, id, cloud, project, and include_archived parameters.
  • Added anyscale.job_queue.delete() to delete a job queue. All jobs must be in terminal state with no running clusters. Supports job_queue_id, name, project, cloud, and include_archived parameters.
  • Added anyscale.schedule.delete() to delete a schedule. Active schedules are automatically paused before deletion. Supports id, name, cloud, and project parameters.

CLI​

  • Added anyscale job delete command to permanently delete a job. Supports --name, --id, --cloud, --project, and --include-archived options.
  • Added anyscale job-queue delete command to delete a job queue. Supports --name or --id options.
  • Added anyscale schedule delete command to delete a schedule. Supports --name or --id options.
  • Added --include-archived flag to anyscale job commands (status, terminate, archive, delete, wait, logs, tags add, tags remove, tags list) to support finding archived jobs.
  • Added --include-archived flag to anyscale job-queue list and anyscale job-queue status.
  • Added anyscale cloud update-storage-cors command to update CORS configuration on cloud storage buckets for Anyscale UI features. Works with AWS, Google Cloud, and Azure clouds.

Behavior changes​

CLI​

  • Removed legacy anyscale cluster-env commands. Use anyscale image commands instead.
  • Removed legacy anyscale cluster start and anyscale cluster terminate commands. Use anyscale workspace_v2 commands instead.

SDK​

Version 0.26.82 release notes​

Version 0.26.82 is available as of January 27, 2026.

New features​

SDK​

  • Added anyscale.job_queue.archive() to archive a job queue and prevent new job submissions. Existing pending and running jobs continue to completion. Supports job_queue_id or name parameters.
  • Added anyscale.job_queue.terminate() to terminate a job queue and all its pending and running jobs. Supports job_queue_id or name parameters.
  • Added anyscale.schedule.url() to get the web UI URL for a schedule. Supports id or name with optional cloud and project parameters.

CLI​

  • Added anyscale job-queue archive command to archive a job queue. Supports --id or --name options.
  • Added anyscale job-queue terminate command to terminate a job queue. Supports --id or --name options.
  • Enhanced anyscale job list --v2 with pagination (--page-size, --interactive/--no-interactive), sorting (--sort), JSON output (--json), and --max-items limit.
  • Enhanced anyscale schedule list --v2 with pagination (--page-size, --interactive/--no-interactive), JSON output (--json), and user filtering (--include-all-users/--only-mine).
  • Added --name option to anyscale job-queue status and anyscale job-queue update as an alternative to --id.
  • Added --v2 flag to anyscale schedule url with support for name resolution using --name, --cloud, and --project.

Compute config​

  • Added validation for accelerator type consistency between ray.io/accelerator-type in required_labels and cloud.google.com/gke-accelerator in nodeSelector. Mismatched GPU types now produce a helpful error message.
  • Added validation requiring GPU count in required_resources when specifying a non-TPU accelerator type in required_labels.
  • Added validation requiring tpu_hosts in required_resources and ray.io/tpu-topology in required_labels when specifying a TPU accelerator type.
  • CLI output for compute configs now includes required_resources and required_labels fields.

Behavior changes​

SDK​

  • Deprecated the AnyscaleSDK class with a warning. Migrate to import anyscale and use anyscale.<module>.<function>(). Anyscale will remove the legacy class on February 28, 2026. See Get started with the Anyscale SDK.

CLI​

  • anyscale workspace_v2 push and anyscale workspace_v2 pull now use HTTPS WebSocket tunnel by default for file transfers. Use --direct-ssh to force legacy SSH (port 22) connection.
  • CLI upgrade warnings now only appear when your version is significantly outdated: major or minor version behind, or more than 5 patch versions behind.
  • anyscale compute-config list command description improved and no longer marked as limited support.

Cloud setup​

  • anyscale cloud setup for AWS (VM and Kubernetes) now configures additional CORS headers (Accept-Ranges, Content-Range, Content-Length) on storage buckets to support the optimized File Viewer feature.

Version 0.26.81 release notes​

Version 0.26.81 is available as of January 14, 2026.

New features​

SDK​

  • Added anyscale.compute_config.get_default() to retrieve the default compute config. Supports optional cloud and cloud_resource parameters for filtering. See Compute config SDK.
  • Added anyscale.compute_config.list() with filtering, sorting, and pagination. Supports cloud_id, cloud_name, sort_by, sort_order, max_items, and next_token parameters. Returns a ComputeConfigListResult object. See Compute config SDK.
  • Added anyscale.job.list() with filtering and pagination. Supports filtering by name, project, cloud, state_filter, tags_filter, and more. See Job SDK.
  • Added anyscale.schedule.list() with filtering and pagination. Supports filtering by name, project, cloud, and creator_id. See Schedule SDK.
  • Added anyscale.image.archive() to archive container images and all versions. See Image SDK.
  • Added cloud parameter to anyscale.compute_config.get() to filter by cloud name when resolving compute configs.

CLI​

  • Added anyscale image archive --name <image> command to archive container images. See Image CLI.
  • Enhanced anyscale compute-config list with pagination (--next-token), filtering (--cloud-id, --cloud-name), sorting (--sort-by, --sort-order), and JSON output (--json).
  • Enhanced anyscale compute-config get with --cloud-id and --cloud-name options for filtering.

Compute config​

  • Added cpu_architecture field to PhysicalResources. Valid values: x86_64 (default) and arm64. See PhysicalResources.

  • GKE now automatically derives node selectors from Ray labels for TPU workloads. Specify ray.io/accelerator-type and ray.io/tpu-topology in labels instead of explicit node selectors:

    labels:
      ray.io/accelerator-type: TPU-V6E
      ray.io/tpu-topology: 2x2

Behavior changes​

CLI​

  • Deprecated anyscale cloud edit. Use anyscale cloud update instead.
  • Refactored cloud setup, cloud update, and cloud delete to use new cloud resource APIs.
  • Log downloads now create the temp directory inside the target directory to avoid cross-partition issues. Error messages for disk space limits now include the target directory, available space, and required space.

Bug fixes​

  • Fixed bug where JobStatus returned by anyscale.job.list() had created_at=None despite the API returning the correct timestamp.
  • Fixed anyscale user list-permissions command.