Weekly Update — June 1 – 8, 2026

[missBerg]

v0.7.0 shipped this week — multi-tenant hostname routing, two new provider translation paths (Anthropic → Bedrock Converse and Azure OpenAI Responses), OpenAI audio transcription/translation endpoints, and the first runtime step toward quota-aware rate limiting. Thanks to everyone who landed code, reviewed, and weighed in on the road to GA. 🎉

🛣️ Road to v1.0 GA — v0.7.0 is out

v0.7.0 shipped on June 6, the last feature release before the mid-June GA. Full notes are on the release page and the docs site. The headline items:

• Multi-tenant hostname routing — AIGatewayRoute.spec.hostnames scopes a route (and its /v1/models listing) to specific hosts, so teamA.ai.example.com and teamB.ai.example.com each see only their own model catalog through one Gateway. Wildcards follow Gateway API matching rules.
• Anthropic /v1/messages → AWS Bedrock Converse — Anthropic-native clients can now reach any Bedrock model (text, images, tool use, thinking, streaming) without changing their integration. Azure OpenAI Responses API and OpenAI audio transcription/translation endpoints land in the same release.
• Quota-aware rate limiting, step one — attaching a QuotaPolicy to an AIServiceBackend now injects a backend rate limit filter for per-backend throttling. Full quota-aware routing across backends is the follow-on work.
• Claude Opus 4.7 reasoning — the display parameter and the new xhigh effort tier are supported for long-horizon agentic and coding tasks.

Two breaking changes to flag before you upgrade:

• AIGatewayRoute.spec.rules is now capped at 15 (down from 128) to match the Gateway API HTTPRoute limit. Split larger routes across multiple AIGatewayRoute resources on the same Gateway.
• Opus 4.7 display defaults to omitted (earlier Claude models defaulted to summarized). Set display: "summarized" explicitly if you want thinking content back.

Next on the GA track: the MCPBackend CRD and full quota-aware routing across multiple backends — see the proposals below.

✨ What's new

Everything that shipped this cycle is in v0.7.0 above — main has only just reopened post-release, so there's no separate post-release work to list yet. Expect the usual run of merges in next week's update.

💡 New design proposals & feature requests

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

• Provider-native API schemas as client input — #2204 by @PraveenKumarInjam. Accept Azure OpenAI, AWS Bedrock, and GCP Vertex AI request formats as client input — not just OpenAI/Anthropic/Cohere — so enterprises with hundreds of provider-native clients can adopt the gateway incrementally instead of forcing a big-bang migration. Picks up the open Gemini-native asks in #1960 and #2165.
• A stable error-handling contract — #2188 by @hustxiayang. Backends disagree on status codes for the same failure (Vertex AI recently flipped a 413 to a 400 and broke callers' error handling). Proposes the gateway present its own consistent error-processing promise so downstream retry logic stays stable across providers.
• Audio-duration cost signal for the new audio endpoints — #2186 by @PatilHrushikesh. Follow-up to the v0.7 audio endpoints (#2023): Whisper bills per minute, not per token, so transcription/translation currently bills at zero and rate-limit buckets never increment. Proposes input_audio_seconds/output_audio_seconds CEL variables so operators can price per minute.
• MCPBackend CRD — #2144 by @aishwaryaraimule21. Design doc (now merged) for a dedicated MCP backend resource that decouples MCP backend config from MCPRoute — flagged as next-up on the v0.7+ roadmap.
• OAuth 2.0 Token Exchange for MCP upstream auth — #2052 by @nacx. Design doc (now merged) for token-exchange-based upstream auth on MCP backends; reviewer eyes still welcome on the approach.

👀 PRs looking for review

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

• #2134 — Stop exposing plaintext API key secret refs in MCP backend HTTPRoutes by @aishwaryaraimule21 — the security fix called out in the maintainer sync as the leading patch candidate; reviewer assigned, close to landing.
• #1982 — OpenAI Files API implementation by @sivanantha321 — the file-upload work behind the sticky-routing discussion (uploads must pin subsequent requests to the same backend); needs Envoy-config review eyes.
• #1884 — Make ext_proc circuit breaker max_requests configurable via Helm by @PatilHrushikesh — lifts the hard-coded 1024 request cap so high-concurrency deployments can tune it.
• #2205 — Mirror feature for AIGatewayRoute by @mtparet — shadow live traffic to a second backend for safe testing and migration validation.
• #2169 — Time-to-first-token timeout with rollover to a new model by @albe2669 — failover trigger for slow-starting upstreams, not just hard errors.
• #2166 — Allow InferencePool as a backend reference on AIServiceBackend by @isztldav — routes inference-pool backends through the same AIServiceBackend surface as other providers.

🙏 Thanks to this week's contributors

@aabchoo, @hustxiayang, @yuzisun, @ajac-zero, @sivanantha321, @aishwaryaraimule21, @PatilHrushikesh, @arpitjain099, @anurags25, @nacx, @Killusions, @mtparet, @guohao, @sandangel, @kanurag94, @Hritik003, @albe2669, @isztldav, @ChrisJBurns, @immanuwell, @CodePrometheus, @xianml, @PraveenKumarInjam, @saixso, @missBerg — and everyone who showed up to triage, review, and discuss.

See you next week!