Weekly Update — May 4 – 11, 2026

[missBerg]

v0.6.0 shipped on Tuesday — the first production-ready API surface, with the core CRDs (AIGatewayRoute, AIServiceBackend, BackendSecurityPolicy, GatewayConfig, MCPRoute) promoted to v1beta1. Main reopened after the cut, and a handful of provider-translation fixes plus the auth-aware MCP tools/list filter landed in the back half of the week. Thanks to everyone who shipped, reviewed, and tested the RC. 🎉

🛣️ Road to v1.0 GA

v0.6.0 is out: release notes · v0.6 docs. Headlines worth flagging if you skipped the RC:

• Core CRDs at v1beta1 — AIGatewayRoute, AIServiceBackend, BackendSecurityPolicy, GatewayConfig, and MCPRoute are now served at v1beta1. v1alpha1 stays registered with deprecation warnings during the upgrade window.
• Two breaking changes. AIGatewayRoute.spec.filterConfig is removed — move ExtProc config to a GatewayConfig resource referenced from the Gateway. The deprecated VersionedAPISchema.version-as-prefix behavior is also gone — use prefix (e.g. prefix: /v1beta/openai for Gemini's OpenAI-compatible API). Migration snippets are in the release notes.
• Cross-provider translation matured. Anthropic /v1/messages now works on any OpenAI-compatible backend, a single reasoning_effort knob (low/medium/high/xhigh) maps onto Anthropic, OpenAI, and Gemini thinking controls, and Gemini gets first-class embeddings plus Anthropic-style prefix context caching.
• MCP got per-backend forwardHeaders (with rename), JWT claim-to-header projection, and exclude/excludeRegex on tool selectors. Tool name now ships in dynamic metadata for access logs and dashboards.
• Operability + privacy — GKE Workload Identity via ADC, request/response body redaction, configurable webhook host networking, agent-session-id → session.id correlation by default, and a new ReasoningToken cost type for budgeting thinking tokens separately.

v1.0 tracking is still #2083 — end-of-June target unchanged.

✨ What's new

Post-release, main reopened and a small cluster of provider-translation polish and MCP work landed on top of v0.6.

MCP

• Auth-aware tools/list filtering — #2106 by @xadereq. MCPRoute now applies the same scope/CEL authorization rules to tools/list that already gate tools/call, so callers only see tools they can actually invoke. Closes #2105 — fixes both the wasted-turn UX and the information-disclosure footgun for consolidated /mcp endpoints.

Provider & translation coverage

• GCP finish-reason mapping cleanup — #2118 by @hustxiayang. Stops mapping every non-stop GCP outcome to content_filter and adds proper OpenAI-compatible mappings for MAX_TOKENS, RECITATION, MALFORMED_FUNCTION_CALL, and friends. Big quality-of-life win for anyone debugging Gemini responses from logs.
• Gemini streaming finish_reason chunk fix — #2107 by @hustxiayang. Gemini streaming responses now always emit a final finish_reason chunk with a delta set, matching OpenAI streaming semantics.

Developer experience

• Versioned docs for v0.6 — #2109 by @nacx, with install-doc corrections in #2117 and a multi-versioned API-reference link fix in #2096. The 0.6 docs are live alongside 0.5 at aigateway.envoyproxy.io/docs/0.6.

💡 New design proposals & feature requests

A few new threads worth a read if any of these are in your wheelhouse:

• Anthropic /messages → AWS Bedrock translator — #2102 by @Killusions, with a draft implementation in #2103. Lets you point an Anthropic-format client at any Bedrock model, not just Claude — the Anthropic-shaped counterpart to today's OpenAI → Bedrock path.
• Azure OpenAI Responses API support — #2110 by @azaretsk, with a draft in #2122 by @ajac-zero. The AzureOpenAI backend only routes /v1/chat/completions today; this adds /v1/responses translation against Azure's 2025-04-01-preview API. Unblocks web_search_preview and other Responses-only tools on Azure.
• AIGatewayRoute >15 rules silently bricks the downstream HTTPRoute — #2120 by @jenrik, with a proposed cap in #2123 by @siddharth1036. The Gateway API HTTPRoute.spec.rules maxItems: 16 limit combined with the controller's catch-all rule means 15-rule AIGatewayRoutes silently fail at admission with the status condition still reporting Accepted: True. Discoverability bug — design discussion is on the issue.
• Streaming gen_ai.client.token.usage histogram drops requests on context cancellation — #2115 by @sc7565. ~28% of long-running Sonnet streams on Bedrock never get recorded in the histogram even though the access log captures them all; root cause looks like the OTEL Histogram.Record call inheriting the cancelled request context. Fix is up in #2116.

👀 PRs looking for review

If you have time to help review, these are open and waiting:

• #2092 — OAuth 2.0 Token Exchange for MCP upstream auth by @nacx — flagged in the maintainer call as the next significant landing; uses the existing dynamic-module token-exchange path, ships with end-to-end coverage against Gemini.
• #2086 — PII redaction & rehydration processor by @tim-peyeeye
• #2099 — Reasoning and image support for Anthropic-to-OpenAI translation by @anurags25
• #2103 — Anthropic /messages → AWS Bedrock translator by @Killusions
• #2116 — Resilient streaming token-usage emission by @sc7565
• #2101 — validate command in AI Gateway CLI by @Sakalya

🙏 Thanks to this week's contributors

@missBerg, @nacx, @hustxiayang, @xadereq, @siddharth1036, @Killusions, @sc7565, @ajac-zero, @anurags25, @tim-peyeeye, @PatilHrushikesh, @Sakalya, @jenrik, @azaretsk, @xiaolin593, @gavrissh — and everyone who showed up to triage, review, and discuss.

See you next week!