May

May brings more control for platform teams: flexible config parameters, new guardrail partners, and expanded provider coverage including Claude Platform on AWS. See what’s new:

​

Summary

​

Gateway

​

Default, override, and drop parameters for config targets

Platform teams get three levers to control exactly what reaches the model—all in one place, without touching client code:

• default_params — fills in a value only when the client hasn’t already set it.
• drop_params — strips fields before they reach the model, by dot-notation path.

These inherit through nested targets, so org-wide defaults and workspace overrides layer cleanly on top of each other. Learn more: Configs

​

Usage limit policies

Teams that need more precision than a budget policy offers now get a dedicated usage limit policy type—cap consumption exactly how your org plans it, across teams or workloads.

​

Endpoint guardrail for proxy requests

Endpoint allow/block policies now apply consistently across both the standard and proxy path—no more gaps when traffic routes differently. Learn more: Guardrails

​

Circuit breaker: provider slug references

Breaker rules stay stable even as providers are renamed or shared across catalogs—no more reconfiguring when a provider slug changes.

​

Guardrails

​

Cato Networks

Cato Networks is now available as a guardrail to inspect LLM prompts and responses for security risks—with monitor, anonymize, or block actions depending on the severity of what’s detected. Learn more: Cato Networks

​

Lakera Guard

The new Lakera Guard integration detects prompt injection and redacts PII. When only PII is detected, Lakera redacts it and lets the request through; any other policy hit blocks the request. Use project_id to select a specific Lakera policy and apiBase to target regional endpoints. Learn more: Lakera Guard

​

Tavily Online

The new Tavily guardrail runs web-grounded checks—verifying claims against live web results before returning them to users, so your guardrail pipeline catches outdated or fabricated information. Learn more: Tavily

​

CrowdStrike AIDR: richer context

Requests to CrowdStrike AIDR now include user ID, user name, LLM provider, and model from your request metadata—giving CrowdStrike more context for policy decisions and audit trails. Learn more: CrowdStrike AIDR

​

Prompt Security

Prompt security checks are extended with additional coverage and configuration options, including provider- and deployment-specific settings.

​

Models and providers

​

Claude Platform on AWS

Anthropic’s full-platform capabilities are now accessible through AWS via the new claude-platform-aws provider, with accessKey, assumedRole, and serviceRole authentication modes. Learn more: Claude Platform (AWS)

​

Vertex AI

• Multi-region endpoints for batches, embeddings, Anthropic Claude on Vertex, and Gemini models
• Anthropic Claude batch workloads through the unified /v1/batches API
• Gemini text-to-speech through the unified speech APIs
• OpenAI-style service_tier forwarded for tiered inference and usage attribution

Learn more: Vertex AI

​

Amazon Bedrock

• service_tier support for correct mapping and usage attribution on tiered inference
• Prompt cache pricing now accounts for 1-hour cached context on Bedrock Mantle

Learn more: Amazon Bedrock

​

Azure AI Foundry

• Native /v1/responses endpoint with agent references

Learn more: Azure AI Foundry

​

Azure

• Workload identity for Azure OpenAI, Azure AI Foundry, and Azure Cache for Redis

​

New providers

• Fal AI — image, video, and audio generation models
• BytePlus — chat completions
• Scenario AI — proxy passthrough with Basic auth

​

Provider updates

• Cloudflare Workers AI — handles OpenAI-compatible response format for newer models (Kimi K2.5, K2.6)
• Together AI — cost now reported correctly on streamed responses
• Anthropic Messages — tool-call argument deltas after the finish chunk are no longer dropped
• Perplexity — long Sonar streams no longer end prematurely

​

Support