SCIM 2.0 provisioning (Users and Groups → roles, per-tenant, escalation-guarded so an IdP can never grant super-admin), IdP group → role mapping, and SAML Single Logout — automated identity lifecycle from your IdP
Self-service MFA enrollment (TOTP with one-time recovery codes), active-session management, and refresh-token rotation with reuse detection
Seven more providers — Novita, SiliconFlow, Nebius, Baseten, Hyperbolic, NVIDIA NIM, and Jina (embeddings + rerank) — bringing the catalog to 49 adapters
Full OpenTelemetry pipeline — traces, metrics, and logs — across the API and worker, plus new Prometheus metrics (cache hit rate, auth failures, rate-limit rejections, streaming time-to-first-token, queue depth) and a bundled Grafana dashboard and alert rules
Developer experience: a typed TypeScript SDK, a gatewyse CLI, and a generated OpenAPI 3 spec at /openapi.json
GDPR data-subject export and erasure, a pluggable KMS/BYOK key provider, budget-alert delivery over email / webhook / Slack, backup restore with sha256 integrity verification, and a schema-migration framework
Cross-replica Redis circuit breaker (no more per-pod tripping), discounted prompt-cache-read pricing, and internal showback CSV export
v2.2
Frontier catalog refresh + routing correctness
OpenRouter meta-provider and MiniMax added — 40 providers total, with OpenRouter importing live per-model pricing for least-cost routing
Catalog reconciled against every vendor’s own docs: current OpenAI GPT-5.x, Anthropic Claude Opus 4.8 / Sonnet 5 / Fable 5, Google Gemini 3.x, xAI Grok 4.x, DeepSeek V4, Moonshot Kimi K3, GLM, Qwen and more — with corrected prices, context windows, and decommissioned models pruned
Reasoning-model request correctness across providers: Claude adaptive thinking, OpenAI / GPT-5 max_completion_tokens, Kimi and Grok sampling-parameter handling — eliminating a class of provider-side 400 errors
Distributed tracing (OpenTelemetry) plus per-model token and cost Prometheus counters for Grafana spend dashboards
Unified guard hierarchy with in-process ML toxicity / PII plus new jailbreak and secrets-detection guards
v2.1
Assistants surface + provider sweep
Files, Batches, Responses, Vector Stores, and Realtime APIs land — full OpenAI-SDK parity for assistants and agent workloads
11 new providers: Moonshot, Zhipu, DashScope (Qwen), DeepInfra, Cloudflare Workers AI, Lambda Labs, Voyage, Deepgram, Cartesia, AWS Bedrock, Vertex AI — total 38
Async batch poller with HMAC-signed completion webhooks; 50% provider discount on supported endpoints
Gateway-owned vector stores with cosine-similarity search and chat-side file_search built-in tool
Realtime WebSocket multiplexer with single-use ticket exchange, sliding idle timeout, per-tenant concurrency cap
v2.0
The Passage
New marketing site with a photoreal Blue Ridge parallax hero
Interactive routing simulator on the home page
Commercial model: lifetime license + managed tiers with 36-month amortization