roh26it · Oct 15, 2026 · Oct 14, 2026 · Oct 14, 2026
diff --git a/changelog/2026/august.mdx b/changelog/2026/august.mdx
@@ -0,0 +1,85 @@
+---
+title: "August"
+---
+
+August brings a native OCR endpoint, a Headroom plugin for reducing input token costs, and OAuth Client Credentials for keyless upstream authentication.
+
+We've also added Deepgram for speech workloads, new guardrail partners, and shipped improvements across Anthropic, Bedrock, and Vertex AI.
+
+See what's new:
+
+## Summary
+
+| Area                     | Key highlights                                                                                              |
+| :----------------------- | :---------------------------------------------------------------------------------------------------------- |
+| **Gateway**              | • Native OCR endpoint <br /> • OAuth Client Credentials for upstream auth <br /> • MCP tool call rate limits |
+| **Guardrails**           | • Headroom plugin <br /> • Singulr <br /> • Header forwarding                                               |
+| **Models and providers** | • Deepgram <br /> • Qwen <br /> • Databricks <br /> • Anthropic                                             |
+
+## Gateway
+
+### Native OCR Endpoint
+
+OCR requests now get the full gateway treatment — retries, fallbacks, load balancing, caching, logging, and pricing — through a unified `POST /v1/ocr` endpoint. Supports Mistral AI, Azure AI Foundry, and Vertex AI, with cost tracked per page processed.
+
+Learn more: [OCR API](https://portkey.ai/docs/api-reference/inference-api/ocr)
+
+### OAuth Client Credentials for Upstream Auth
+
+Upstream providers now authenticate without a static API key. The gateway acquires and caches a bearer token from your configured token endpoint using the **OAuth Client Credentials** grant and forwards it upstream automatically.
+
+Learn more: [OpenAI integration](https://portkey.ai/docs/integrations/llms/openai)
+
+### MCP Tool Call Rate Limits
+
+Tool invocation volume is now governable—rate limits apply to MCP tool calls just like they do to LLM requests. Set limits per config to prevent runaway agent loops or unexpected tool usage spikes.
+
+Learn more: [MCP rate limits](https://portkey.ai/docs/product/mcp-gateway/rate-limits)
+
+## Guardrails
+
+### Headroom Plugin
+
+Input token costs go down without changing your prompts. The **Headroom plugin** compresses request context before it reaches the model, configured through the standard guardrails workflow.
+
+Learn more: [Headroom plugin](https://portkey.ai/docs/integrations/guardrails/headroom)
+
+** Other updates:**
+
+- **Singulr** — now available as a guardrail partner for request and response scanning
+- **Header forwarding** — specific client headers (e.g. `x-session-id`, `traceparent`) now flow through to guardrail upstream providers. Credential headers are never forwarded.
+
+## Models and providers
+
+### New providers
+
+- **Deepgram** — speech-to-text and text-to-speech, with full observability and reliability features
+
+### Qwen
+
+- Responses API and Messages API now supported, along with additional non-OpenAI parameters
+
+### Databricks
+
+- Responses API now supported for Databricks-served models
+
+### Anthropic
+
+- New beta parameters now passed through: `inference_geo`, `diagnostics`, `fallbacks`, and `context_management`
+- `thinking_tokens` now correctly mapped to OpenAI's `reasoning_tokens` for unified token tracking
+
+## Bug fixes and improvements
+
+- MCP Registry requests now proxied through the gateway
+- Meshy and Tripo3D: credit-based cost tracking added
+
+## Support
+
+<CardGroup cols={2}>
+<Card title="Need Help?" icon="bug" href="https://github.com/Portkey-AI/gateway/issues">
+Open an issue on GitHub
+</Card>
+<Card title="Join Us" icon="discord" href="https://portkey.wiki/community">
+Get support in our Discord
+</Card>
+</CardGroup>
diff --git a/changelog/2026/july.mdx b/changelog/2026/july.mdx
@@ -0,0 +1,69 @@
+---
+title: "July"
+---
+
+July brings cross-API routing between Messages and Responses API, expanded agentic client support in Prisma AIRS, and a soft deny mode for guardrails that keeps sessions alive for agentic clients.
+
+We've also added Lightning AI and Akto as new integrations, and shipped improvements across Anthropic, Vertex AI, and Bedrock.
+
+See what's new:
+
+## Summary
+
+| Area                     | Key highlights                                                                                                                                                              |
+| :----------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Gateway**              | • Messages API to Responses API routing                                                                                                                                     |
+| **Guardrails**           | • Prisma AIRS: agentic client support <br /> • Soft deny mode (HTTP 200) <br /> • Akto                                                                                      |
+| **Models and providers** | • Lightning AI <br /> • MCP user attribution in logs                                                                                                                        |
+
+## Gateway
+
+### Messages API to Responses API Routing
+
+Route Messages API requests to Responses API models—like OpenAI's Codex—without changing your client code. Requests and responses are transformed automatically in both directions.
+
+Learn more: [Responses API](https://portkey.ai/docs/api-reference/inference-api/responses/responses)
+
+## Guardrails
+
+### Prisma AIRS: Agentic Client Support
+
+False positives drop significantly for agentic clients like Claude Code, Cursor, and Cline. **`strip_scaffolding`** removes agent-harness content—system prompts, MCP tool instructions—before scanning, so Prisma AIRS inspects actual user content. **`scan_scope`** gives you control over which messages get scanned.
+
+Learn more: [Prisma AIRS](https://portkey.ai/docs/integrations/guardrails/palo-alto-panw-prisma)
+
+### Guardrails: Soft Deny
+
+Guardrail denials no longer crash agentic sessions. With **`soft_deny_200`**, a denial comes back as HTTP 200 formatted as a chat completion response—so clients like Claude Code that treat 4xx as fatal errors keep running and see the denial as a model response.
+
+Learn more: [Guardrails](https://portkey.ai/docs/product/guardrails)
+
+
+## Models and providers
+
+### New providers
+
+- **Lightning AI** — chat completions
+
+### MCP Gateway: user attribution in logs
+
+Every MCP request now traces back to a specific user—derived from OAuth claims or the API key name—for clearer per-user visibility in analytics.
+
+Learn more: [MCP observability]()
+
+## Bug fixes and improvements
+
+- Vertex AI: Qwen and xAI Model Garden routing added
+- Workers AI: reasoning support for K2.7 and GLM-5.2
+- Anthropic: cost attribution now differentiates batch, fast, and standard pricing tiers
+
+## Support
+
+<CardGroup cols={2}>
+<Card title="Need Help?" icon="bug" href="https://github.com/Portkey-AI/gateway/issues">
+Open an issue on GitHub
+</Card>
+<Card title="Join Us" icon="discord" href="https://portkey.wiki/community">
+Get support in our Discord
+</Card>
+</CardGroup>
diff --git a/changelog/2026/june.mdx b/changelog/2026/june.mdx
@@ -0,0 +1,81 @@
+---
+title: "June"
+---
+
+June brings deeper MCP capabilities and richer observability across the gateway.
+
+We've also shipped new guardrail partner, expanded provider support, and improvements across Anthropic, Bedrock, and Vertex AI.
+
+See what's new:
+
+## Summary
+
+| Area                     | Key highlights                                                                                                                                                                                |
+| :----------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Gateway**              | • Server-Side MCP Execution (Beta) <br /> • OTel guardrail spans <br /> • Token metering for MCP and agents <br /> • MCP OTel log export                                                    |
+| **Guardrails**           | • Alibaba Cloud AI                                                                                                                                                                            |
+| **Models and providers** | • Azure AI Inference: MAI image generation <br /> • MiniMax <br /> • Anthropic context management <br /> • Claude Code OAuth                                                                  |
+
+## Gateway
+
+### Server-Side MCP Execution (Beta)
+
+MCP tools now work through the AI Gateway on **any provider**—including AWS Bedrock and Vertex AI, which don't natively support remote MCP. The gateway handles fetching tool definitions and executing calls server-side, so you get MCP capabilities without any provider-level changes.
+
+
+### OTel: Guardrail execution spans
+
+Guardrail checks now emit their own OTel spans, so exactly what your guardrails did right alongside the model call in your traces.
+
+
+### Token metering for MCP and agent requests
+
+Every token from **MCP requests** and **agent endpoints** now shows up in your analytics and logs alongside LLM usage—one complete picture of consumption across all your AI workloads.
+
+
+### MCP gateway: OTel log export
+
+MCP traffic now appears in the same observability pipeline as your LLM workload exported to your OpenTelemetry collector alongside gateway logs.
+
+
+## Guardrails
+
+### Alibaba Cloud AI
+
+**Alibaba Cloud AI** is now available as a guardrail for content safety checks, configurable through the standard guardrails workflow.
+
+
+## Models and providers
+
+### Azure AI Inference
+
+- Image generation and editing now supported for MAI models (e.g., MAI-Image-2.5)
+
+### New providers
+
+- **MiniMax** — chat completions
+
+### Anthropic
+
+- **Context management:** `context_management` passed through to the upstream API, enabling memory tools and managed-context workflows on Claude models
+- **Claude Code OAuth:** Requests from Claude Code using an OAuth bearer token now forward `Authorization` to Anthropic correctly
+- **File downloads:** Binary responses (e.g., PPTX from Skills + Code Execution) now pass through untouched
+- **Thinking blocks in streaming:** Reasoning summary blocks are now included in streaming Responses API responses for Anthropic models
+
+## Bug fixes and improvements
+
+- Vertex AI: MoonshotAI model provider now supported
+- Modal: embeddings now supported alongside chat completions
+- Fireworks AI: `stream_options` support added
+- Amazon Bedrock Guardrails now only block when the action is blocked
+
+## Support
+
+<CardGroup cols={2}>
+<Card title="Need Help?" icon="bug" href="https://github.com/Portkey-AI/gateway/issues">
+Open an issue on GitHub
+</Card>
+<Card title="Join Us" icon="discord" href="https://portkey.wiki/community">
+Get support in our Discord
+</Card>
+</CardGroup>
diff --git a/changelog/2026/may.mdx b/changelog/2026/may.mdx
@@ -0,0 +1,140 @@
+---
+title: "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
+
+
+| Area                     | Key highlights                                                                                                                                                                                                                                                                          |
+| :----------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Gateway**              | • `default_params`, `override_params`, and `drop_params` for config targets <br /> • Usage limit policies <br /> • Prometheus config and API key labels <br /> • Proxy-path endpoint guardrail                                                                                          |
+| **Guardrails**           | • Cato Networks <br /> • Lakera Guard <br /> • Tavily online checks <br /> • CrowdStrike AIDR metadata forwarding <br /> • Prompt Security updates                                                                                                                                      |
+| **Models and providers** | • Claude Platform on AWS <br /> • Vertex AI: multi-region, Anthropic batches, Gemini TTS, `service_tier` <br /> • Amazon Bedrock: `service_tier`, prompt cache pricing <br /> • Azure AI Foundry: native Responses API <br /> • Multi-dimensional pricing <br /> • Fal AI, BytePlus, Scenario AI |
+
+
+
+
+## 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](/docs/product/ai-gateway/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](/docs/product/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](/docs/integrations/guardrails/cato)
+
+### 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](/docs/integrations/guardrails/lakera)
+
+### 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](/docs/integrations/guardrails/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](/docs/integrations/guardrails/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)](/docs/integrations/llms/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](/docs/integrations/llms/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](/docs/integrations/llms/bedrock)
+
+### Azure AI Foundry
+
+- Native `/v1/responses` endpoint with agent references
+
+Learn more: [Azure AI Foundry](/docs/integrations/llms/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
+
+<CardGroup cols={2}>
+<Card title="Need Help?" icon="bug" href="https://github.com/Portkey-AI/gateway/issues">
+Open an issue on GitHub
+</Card>
+<Card title="Join Us" icon="discord" href="https://portkey.wiki/community">
+Get support in our Discord
+</Card>
+</CardGroup>