Changelog

Latency routing mode now uses live P95 telemetry per region rather than static averages

Added `x-budget-usd` header for per-request cost guardrails with automatic model fallback

Semantic cache layer: identical and near-duplicate prompts reuse cached completions (opt-in)

Proxy overhead reduced from 18ms to 6ms average through WASM routing core

Failover chain now executes in parallel after 2s primary timeout instead of sequential

Streaming responses no longer drop final `[DONE]` token on long outputs

DPA auto-attachment now correctly includes APAC SCCs for Singapore-homed accounts

GPT-4o and GPT-4o Mini added to the official model registry with real-time pricing data

EU-West (Frankfurt) PoP launched, reducing median latency for European customers by 38%

Webhook alerts for budget threshold breaches (configurable at 50%, 80%, 100%)

routing_metadata object now includes `failover_chain` array showing evaluated candidates

API key rotation generates replacement before invalidating old key (zero-downtime rotation)

`top_p` and `frequency_penalty` parameters were being dropped on cost-routed requests

SOC2 Type II audit completed. Certificate available in Trust Center.

Python SDK (openai-compatible) now documented with ModelSwitch header examples

Cost routing mode: token budget projection now estimates output tokens using input length heuristic

Request logs are now queryable via the dashboard with up to 30-day retention on Scale+ plans

Correct HTTP 429 propagation when upstream provider rate limits trigger on failover chain tail

Multi-region request routing: requests are now load-balanced across the nearest available PoP

`x-failover-chain` header allows per-request custom model priority ordering

Dashboard redesigned with bento-style layout and live cost attribution by project

Model registry expanded with context window and per-capability metadata

Complete rewrite of the routing engine in Rust/WASM for sub-10ms overhead

Three routing modes introduced: cost, performance, latency

`routing_metadata` field added to all chat completion responses

Custom response headers: `x-modelswitch-latency`, `x-modelswitch-model`, and others

API key format changed from `sk-ms-` prefix to `ms_live_` / `ms_test_` prefix. Old keys deprecated March 2025.

Streaming endpoint path changed from `/v1/stream` to unified `/v1/chat/completions` with SSE