Monitor semantic governance policies with built-in metrics (Preview)

Built-in Cloud Monitoring metrics for the semantic governance policy engine are available in Preview. You can now observe request throughput, evaluation counts, latencies, verdict distribution (ALLOW versus DENY), and LLM token consumption for the policy engine directly in Metrics Explorer, query them through the Cloud Monitoring v3 API and PromQL, and use them in alerting policies.

For more information, see Monitor semantic governance policies.

Gemini 3.7 Flash is generally available

Gemini 3.7 Flash is now generally available (GA) and available for production use. This model is our first model to enable agentic video processing enabled by default.

For more information on 3.7 Flash, see the model page.

CodeMender CLI: Sandbox enabled by default

This release updates the CodeMender CLI default behavior:

• Sandbox enabled by default: The CLI now runs commands inside the process-level sandbox by default to protect your workstation. You can disable the sandbox in your config.yaml, by passing --sandbox=false to CLI commands, or bypass it using the --unrestricted flag.

For more information, see Install the CLI and configure.

Process-level sandboxing and auto-updates (Preview)

This release introduces process-level sandboxing, automatic update checks, and other improvements to the CodeMender CLI (Preview):

• Process-level sandboxing: To safeguard your workstation against unintended file modifications or unexpected tool side effects, you can now execute all agent-proposed tools (such as compiling code, running tests, or executing shell scripts) inside an OS-level sandbox. Sandboxing is disabled by default to allow seamless dependency resolution, but can be enabled persistently in config.yaml or per-command using the new --sandbox flag.
• Automatic updates: The CLI now automatically checks for updates in the background (at most once every 24 hours) when running in an interactive terminal. You can also run cm update to forcefully check for and apply updates immediately.
• Token usage statistics: You can now use the cm stats subcommand to view a summary of token usage (input, output, cached, thought, tool-use) for your local sessions.
• Improved Windows support: The standalone CLI now has improved support for running on Windows.

For more information, see Install the CLI and configure.

Anthropic's Claude Opus 5

Claude Opus 5 is available in Model Garden.

Gemini 3.6 Flash and 3.5 Flash-Lite are generally available (GA)

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are now generally available (GA) and available for production use. These models are designed to improve upon their predecessors' capabilities, including improved token usage and improved document understanding. See the linked model information pages for more information.

This release includes some potentially breaking changes from previous Flash and Flash-Lite models:

• Sampling parameters: Custom values for temperature, top-K, and top-P are not supported and will be ignored if set.
• Penalty parameters: Custom values for frequency and presence penalty parameters are not supported. Setting these will result in an API error.
• API turn structure: API requests where the last input turn has a role of Model are no longer supported and will return an error:
  • Interactions API: Requests where the last object in the input array has "type": "model_output" will fail.
  • GenerateContent API: Requests where the last object in the contents array has "role": "model" will fail.

Feature

CodeMender available in Public Preview

CodeMender is now available in Preview. CodeMender is an AI code security agent that can find, verify, and fix deep vulnerabilities in your codebase. CodeMender wraps a fine-tuned harness around a large language model (LLM), using Google DeepMind-engineered prompts, skills, and orchestration logic to turn the model into an agentic system specialized in code security.

Key capabilities include:

• Find vulnerabilities: Scan your codebase using an agent-guided LLM with specialized security tools, or import findings from external static analysis tools.
• Verify vulnerabilities: Build code and execute proof-of-concept (PoC) exploits in your local sandbox to confirm exploitability and reduce false positives.
• Fix vulnerabilities: Automatically generate, validate, and apply source code patches tested against local builds and unit tests to prevent regressions.
• Manage sessions and diffs: Track scan and remediation attempts as stateful sessions, inspect unified diffs, and export detailed findings reports in HTML, Markdown, JSON, or SARIF formats.

For more information, see CodeMender overview.

Deprecated

Open model endpoint deprecations

The following open model endpoints are deprecated and will be retired on October 21, 2026. For more information, see Open model deprecations.

• deepseek-ocr-maas
• deepseek-r1-0528-maas
• deepseek-v3.2-maas
• deepseek-v3.1-maas
• glm-5-maas
• glm-4.7-maas
• gpt-oss-20b-maas
• kimi-k2-thinking-maas
• llama-3.3-70b-instruct-maas
• minimax-m2-maas
• multilingual-e5-large-instruct-maas
• multilingual-e5-small-maas
• qwen3-235b-a22b-instruct-2507-maas
• qwen3-coder-480b-a35b-instruct-maas
• qwen3-next-80b-a3b-instruct-maas
• qwen3-next-80b-a3b-thinking-maas

Announcement

CodeMender is available in Public Preview

CodeMender is now available in Public Preview for select customers. CodeMender is an AI-powered code security agent that runs on your local workstation to find, verify, and fix security vulnerabilities in your codebase.

Key capabilities in this release include:

• Scan and find: Scan your codebase for application logic, memory management, and authentication vulnerabilities, or import findings from external scanners.
• Verify: Build your code and run tests locally to confirm whether findings are exploitable, minimizing false positives.
• Fix: Auto-generate security patches and verify them locally before applying them to your workspace.
• Safety controls: Enforces a Human-in-the-Loop workflow with mandatory manual confirmations for all disk writes and command executions by default.

For more information, see the CodeMender overview.

Security update for Server-Side Request Forgery (SSRF) in Agent Studio

This release fixes a Server-Side Request Forgery (SSRF) vulnerability in the auto-generated /api-proxy backend endpoint for web applications created before July 1, 2026, using Agent Studio.

If you downloaded, generated, or deployed web application code from Agent Studio before July 1, 2026, regenerate the app from Agent Studio and deploy the new version. For more information, see Quickstart: Deploy your Agent Studio prompt as a web application

The updated backend code includes strict domain allowlist validation, ensuring that destination hostnames for the /api-proxy endpoint end with allowed Google Cloud domains, such as *-aiplatform.clients6.google.com

Gemini 3.1 Flash Image Preview and 3 Pro Image Preview are retired

The Nano Banana preview models gemini-3.1-flash-image-preview and gemini-3-pro-image-preview have been retired and are no longer accessible. Update your code to use either gemini-3.1-flash-image or gemini-3-flash-image instead.

Memory Bank memory profiles are generally available (GA)

Memory profiles in Memory Bank are generally available (GA). Memory profiles allow you to generate structured profiles, which are data structures with static schemas populated and updated using LLMs. By defining a fixed schema, you ensure your agents have immediate, low-latency access to evolving information without the need for expensive search operations during a session.

For details, see Memory profiles.

Retirement for preview models for 2.5 Flash, 2.5 Flash-Lite, and 3.1 Flash-Lite

The following preview model endpoints have been retired and are no longer accessible:

• gemini-2.5-flash-lite-preview-09-2025
• gemini-2.5-flash-preview-05-2025
• gemini-3.1-flash-lite-preview

See Migrate to the latest Google models for information on how to migrate your project.

Grok 4.1 models deprecation

The Grok 4.1 model family (including xai/grok-4.1-fast-reasoning and xai/grok-4.1-fast-non-reasoning) on the Gemini Enterprise Agent Platform is deprecated and will be shut down on August 20, 2026. After this date, Google Agent Platform Model as a Service (MaaS) will no longer serve these models. Specifically, API requests directed to the https://aiplatform.googleapis.com/v1/projects/<your-project>/locations/global/endpoints/openapi/chat/completions endpoint using the Grok 4.1 model IDs will fail and return a 400 error. To maintain service, migrate your applications to newer xAI models (such as Grok 4.2 or Grok 4.3) or choose an alternative model from the Google Cloud Model Garden.

Feature

Memory Bank IngestEvents is generally available (GA)

The Memory Bank IngestEvents API is generally available. The IngestEvents API decouples event ingestion from memory generation, letting you continuously stream content to Memory Bank and configure when memory generation is triggered.

This GA release includes the following features:

• Carry over context between generation windows: Use the overlap_event_count parameter to re-include already-processed events at the start of the next window to keep memories coherent.
• Configure memory revisions for ingested events: Use revision_labels, revision_ttl, or disable_memory_revisions to customize how generated revisions are managed.
• Attach metadata to memories: Use the metadata and metadata_merge_strategy configuration parameters to store structured information alongside generated memories.

For details, see Ingest events.

Feature

Memory Bank support for Gemini Embedding 2

Memory Bank supports the gemini-embedding-2 model for similarity search configurations.

When you configure gemini-embedding-2 as the embedding model, you must use one of the global, us, or eu endpoints in the model's resource name (for example, projects/{project}/locations/us/publishers/google/models/gemini-embedding-2). Memory Bank does not support using regional locations (for example, us-central1) for gemini-embedding-2.

For details, see Similarity search configuration.

AlphaGenome released for Gemini Enterprise Agent Platform

AlphaGenome, Google DeepMind's state-of-the-art genomics foundation model, is now available for deployment and use with Gemini Enterprise Agent Platform.

Designed to decipher the functional regulatory code of the human genome, AlphaGenome analyzes large-scale DNA sequences at single-base resolution to predict how genetic variations affect molecular and biological mechanisms like gene expression, chromatin accessibility, and RNA splicing. For information on how to use AlphaGenome in Agent Platform, see the documentation.

Provisioned Throughput: Multiple pending new orders GA

The ability to submit multiple pending orders is generally available. you can submit up to seven Google model orders for the same model and region. See Multiple pending orders.

Anthropic's Claude Sonnet 5

Claude Sonnet 5 is available in Model Garden.

Gemini 3.5 Flash default model for Memory Bank

The default model used for Memory Bank generation is now Gemini 3.5 Flash instead of Gemini 2.5 Flash. For information, see Set up Memory Bank.

Feature

Semantic Governance Policies available in Public Preview

Semantic governance policies and the policy engine are now available in Preview. Semantic governance policy provides an intelligent security and compliance layer that evaluates an AI agent's proposed tool calls against user intent and organizational business rules at runtime.

Key capabilities include:

• Natural Language Constraints (NLC): Author declarative business rules and security guardrails in plain English without needing to write code or redeploy agent applications.
• Layered Intent Gating: Intercepts agent tool calls at runtime to verify alignment with trusted user intent and prevent unauthorized actions, rogue tool use, and data exfiltration.
• Granular Scoping: Apply constraints globally across all tools for an agent or target specific tools and parameters (e.g., enforcing strict financial limits or geographic restrictions).
• Agent Skills Lifecycle Governance: Protects agents from context poisoning and supply-chain exploits by governing the dynamic loading of Agent Skills (tool packages) during sessions.
• Dry Run Mode: Test and observe policy verdicts in Log Explorer before enforcing them on active traffic.

For more information, see Semantic governance policies overview.

Feature

Provisioned Throughput: Email notifications GA

The ability to get email notifications for Provisioned Throughput events is generally available. See Get email notifications.

New Provisioned Throughput features

The following Provisioned Throughput features are generally available:

• Change an order: See Change a standard Provisioned Throughput order.
• Schedule a new order: Set a start date and time when you create an order. See Place a standard Provisioned Throughput order.
• Schedule a change to an order: Set a start date and time when you change an order. See Change a standard Provisioned Throughput order.
• Split an order: Divide an active order into two orders. See Split an order.

Administrative correction to Gemini Online Inference API on Gemini Enterprise Agent Platform SLA

We made an administrative correction to the Gemini Online Inference API on Gemini Enterprise Agent Platform Service Level Agreement (SLA), addressing a clerical issue.

AI security findings in Agent Platform

Viewing AI security findings and posture management summaries in Gemini Enterprise Agent Platform is generally available.

With this release, the Security dashboard introduces the Top security findings widget.

Also, specific features within the AI security widgets are available in Preview, including the following:

• Vulnerability findings and threat monitoring for agent runtimes (such as Cloud Run)
• Historical content violation trends (Violations over time)

For more information, see View security findings.

Feature

Model Armor for Agent Gateway in General Availability

You can enable Model Armor on an Agent Gateway resource to apply your organization's content security guardrails to prompts and responses that pass through the gateway. This feature is generally available (GA).

For more information, see Configure Model Armor on a gateway.

Supervised fine-tuning available for Gemini 3.1 Flash Lite and Gemini 3.5 Flash in Public Preview

Supervised fine-tuning is now available for the gemini-3.1-flash-lite and gemini-3.5-flash models in Preview. Model tuning for Gemini 3.1 Flash Lite and Gemini 3.5 Flash is restricted to us-central1 and europe-west4 and tuned model serving is restricted to the us and eu multi-region endpoints.

See About supervised fine-tuning for more information.

Feature

Provisioned Throughput support for supervised fine-tuned Gemini 3 model inference.

Provisioned Throughput can be used to assure supervised fine-tuned inference using the same quota. Supervised fine-tuned inference for Gemini 3 models incurs a higher burndown rate compared to base model inference. Learn more here.

Agent Gateway in General Availability

Agent Gateway is the networking component of the Gemini Enterprise Agent Platform ecosystem. It secures and governs connectivity for all agentic interactions, whether they occur between users and agents, agents and tools, or among agents themselves.

For details, see Agent Gateway overview.

Feature

Agent Observability is generally available (GA)

This release provides visibility into the performance, behavior, and health of deployed agents and Model Context Protocol (MCP) servers directly within the agent management workflow.

Key updates in this release include:

• Default-On Tracing: OpenTelemetry tracing is now enabled by default for newly deployed Agent Development Kit (ADK) agents on Agent Engine, simplifying the observability setup process without requiring manual configuration.
• Storage Prioritization: Google Cloud Storage (GCS) is the default storage choice in the Google Cloud Console, instead of Cloud Logging. We recommend that you store your multimodal prompt and response payloads in a Cloud Storage (GCS) bucket. This solution provides robust support for large payloads and it enables fine-grained lifecycle management.
• Enhanced Tracing: Inspect step-by-step session execution and view directed acyclic graphs (DAGs) of trace spans.

For more information, see the following:

• Observability overview
• View agent traces
• Set up tracing

Feature

Agent Registry is generally available (GA)

Agent Registry is generally available (GA). Agent Registry is a centralized catalog for discovering and registering agents and Model Context Protocol (MCP) servers.

The following features are available in Agent Registry for the GA launch stage:

• API v1 and client libraries: The v1 version of the Agent Registry API is available. Cloud client libraries are available in C#, Go, Java, Node.js, PHP, Python, and Ruby.
• A2A v1 support: Agent Registry supports Agent-to-Agent (A2A) protocol version 1.0, letting you explicitly declare transport endpoints and bindings inside the supportedInterfaces array, in addition to the existing 0.3 schema support.
• Terraform support: Terraform scripts for Application Default Credentials (ADC) have graduated to General Availability. You can use Terraform to configure and manage your agents, MCP servers, endpoints, and bindings.

Known limitations:

• Access Transparency and Access Approval: Access Transparency logs, which provide visibility into when Google personnel access your content, and Access Approval controls aren't available for Agent Registry configurations.
• Data Residency: If you configure the resource location constraint in your organization policy, Agent Registry enforces the constraint when you register a resource. However, detective controls for data residency compliance reporting are limited.

For more information, see the Agent Registry overview.

Feature

The Agent Identity API (agentidentity.googleapis.com) is available in Preview. This new API replaces the legacy IAM Connectors API (iamconnectors.googleapis.com) for managing auth providers and agent identities.

During the preview migration period, both APIs operate side-by-side. Existing auth providers are automatically mirrored to the new V2 resource hierarchy (authProviders/), allowing you to migrate your IAM policies, agent code, and client applications without downtime.

Memory Bank and Sessions global and multi-regional endpoints GA

Memory Bank and Sessions support for multi-regional and global endpoints is now in General Availability (GA). For more information, see Supported locations for agents. Note that Customer-Managed Encryption Keys (CMEK) cannot be used if your Memory Bank or Sessions instance is configured to use the global endpoint.

Reinforcement Learning fine-tuning is available for Gemini 3.5 Flash in Public Preview

Reinforcement Learning fine-tuning is now available for the gemini-3.5-flash models in Preview. Model tuning for Gemini 3.5 Flash is restricted to us-central1 and europe-west4, and tuned model serving is restricted to the us and eu multi-region endpoints.

See About reinforcement learning fine-tuning for more information.

Anthropic's Claude Fable 5

Claude Fable 5 is available in Model Garden.

Updates to abuse monitoring and zero data retention documentation

Documentation for abuse monitoring, zero data retention, and responsible AI has been updated to align with the Advanced AI Safety Addendum. These updates include new details regarding Advanced AI safety, partner-specific terms, and request-response logging for models like Claude Mythos and Opus.

For more information, see:

• Abuse monitoring
• Zero data retention
• Responsible AI

Gemini 2.0 Flash and Gemini 2.0 Flash-Lite are discontinued

Gemini 2.0 Flash and 2.0 Flash-Lite are discontinued and are no longer available. This includes both model serving and Provisioned Throughput. Use Gemini 3.1 Flash-Lite, Gemma 4, or more recent Gemini releases.

Agent Platform Gemini 3.1 Flash Image and Gemini 3 Pro Image

Gemini Enterprise Agent Platform Gemini 3.1 Flash Image and Gemini 3 Pro Image are Generally Available.

With this release, Gemini 3.1 Flash Image and Gemini 3 Pro Image support 4K image outputs in Preview.

Also supported in this release, Gemini 3.1 Flash Image supports video inputs in Preview. You can use video inputs to generate thumbnails or representative images of videos.

For more information, see the following:

• Gemini 3.1 Flash Image (Nano Banana 2)
• Gemini 3 Pro Image (Nano Banana Pro)

Deprecated

Agent Platform Gemini 3.1 Flash Image Preview and Gemini 3 Pro Image Preview deprecation

Gemini Enterprise Agent Platform Gemini 3.1 Flash Image Preview and Gemini 3 Pro Image Preview are deprecated. We recommend that you update your model endpoints before July 17, 2026, to avoid service disruption.

The following are the discontinued endpoints and recommended endpoint migration:

Feature

Anthropic's Claude Opus 4.8

Claude Opus 4.8 is available in Model Garden.

User ID logging now included with agent logs when you opt in to "Enable logging of prompt inputs and response outputs"

Prompt input and response output logging now includes the user.id field. This addition allows better tracking of anomalous tool interactions.

For details on configuration, see Write traces for an agent.

The Gemini Deep Research Agent released in Preview

The Gemini Deep Research Agent has been released in Preview. The Gemini Deep Research Agent is a managed AI agent that plans, executes, and synthesizes complex, multi-step research workflows across the public web and private enterprise data to generate comprehensive, cited reports.

For more information, see Use the Gemini Deep Research Agent.

Feature

Agent Platform Sandboxes

Additional Agent Platform sandbox features are now available:

• Computer use (Preview): Enables agents to automate browser-based tasks within an isolated web browser environment. You can control the browser using the API or connect directly using the Chrome DevTools Protocol (CDP).
• Custom container sandboxes (Preview): Bring your own container (BYOC) to run custom workloads with specialized dependencies hosted in Artifact Registry.
• Sandbox templates (Preview): Define sandbox specifications as reusable templates relying on pre-warmed pools to facilitate rapid, reliable startups.
• Sandbox snapshots (Preview): Save the exact state of your sandbox environment (including dependencies and file systems) and restore it to a new sandbox.

Feature

Identify the agents with the most content security violations

The Security dashboard displays the top 10 agents with the most content violations detected by Model Armor. The list shows the agent ID of each agent and the number of violations detected for that agent. For more information, see Monitor content security.

Supervised fine-tuning available for Gemini 3.1 Flash Lite (Preview)

Supervised fine-tuning is now available for limited support for the gemini-3.1-flash-lite model. During this period, model tuning for Gemini 3.1 Flash Lite is restricted to us-central1 and europe-west4 and tuned model serving is restricted to the us and eu multi-region endpoints.

See About supervised fine-tuning for more information.

Change

Set media resolution at a Part-level for data when using supervised fine-tuning

Supervised fine-tuning now supports Part-level mediaResolution declarations for images, videos, and PDFs. Part-level media resolution declarations also support the MEDIA_RESOLUTION_ULTRA_HIGH level.

See the following media type–specific pages for more information:

• Document tuning
• Image tuning
• Video tuning

Gemini 3.5 Flash is generally available (GA)

For details, see the model specifications page.

Feature

Manage agent revisions and traffic splitting

Agent revisions and traffic splitting are now available in public preview. You can create immutable revisions of deployed agents, and split traffic between the different active revisions. This enables canary deployments and safe testing of new agent versions. For more information, see Manage revisions and traffic.

Feature

Manage and discover agent skills with Skill Registry

Manage and discover agent skills with the Skill Registry, in public preview. This secure, private, and low-latency repository stores skills as self-contained packages, including instructions, code, and documentation, to enhance agent abilities.

For more information, see:

• Skill Registry overview
• Create and manage skills

Feature

Managed Agents API on Agent Platform released in Preview

The Managed Agents API on Agent Platform has been released in Preview.

This feature allows you to build and scale autonomous agents, including those built from configuration using the Antigravity harness. These agents run in a fully managed and isolated sandbox environment, equipped with tools and skills, and can be interacted with via a dedicated API.

For more information, see the following:

• Managed Agents API on Agent Platform overview
• Create and manage agents
• Interact with agents
• Managed Agents API on Agent Platform sandbox environment

Feature

AI Content Detection API available

AI Content Detection API is available in Preview. For details see AI Content Detection.

Feature

Provisioned Throughput for Gemini now supports latency SLA

Provisioned Throughput now provides a tokens per second latency SLA, covering generation speed from the first returned token to the last.

For more information, see the Gemini Enterprise Agent Platform Online Inference Service Level Agreement (SLA).

Change

Memory Bank and Sessions support for multi-regional and global endpoints is in Preview. For more, see Supported locations for agents in Agent Platform.