Changelogs: Week of June 8, 2026

Last week's product work focused on network routing intelligence, expanding transcription capabilities, and hardening platform trust across billing, security, and analytics.

Smarter routing: model-aware and load-aware

PolarGrid's autorouter now factors in model availability and real-time queue depth before routing a request. The router filters candidates by model availability first, then scores by a combination of distance, in-flight queue depth, and related factors. Same-data-center nodes load-balance freely. Spillover to a more distant data center only kicks in when a local node is genuinely at capacity. Model availability refreshes continuously from live health probes. Queue depth tracks the full lifetime of every streaming response and WebSocket session, not a snapshot.

The change is transparent to callers. Pass a model parameter, and the network handles the rest.

Live streaming speech-to-text

Partial transcripts now arrive while audio is still being captured. Connect to the transcription WebSocket endpoint, stream 16 kHz PCM audio frames as your microphone captures them, and receive rolling partials during the utterance. The authoritative final transcript arrives within 159ms of the speaker stopping, at p50.

Authentication is the same as the rest of the API. The session caps at 120 seconds; this surface is designed for conversational turns, not bulk audio processing.

STT inference is 2x faster

Server-side speech-to-text inference time dropped by roughly half this week, from 287ms to 145ms p50 on production hardware. The improvement applies automatically; no integration changes required.

The root cause was audio resampling. When audio arrives at a sample rate other than 16 kHz, the gateway calls out to ffmpeg via subprocess to convert it, a detour that added approximately 180ms per request. That conversion now runs in-process using scipy, taking about 1ms instead.

In practice, most production audio sources — WebRTC, telephony, and professional microphones — don't output native 16 kHz. If your pipeline was hitting this path, your STT latency just got cut in half.

The analytics dashboard is live

The analytics section of the management console now shows real usage data. Every metric — total requests, average latency, cost, request volume over time, latency by region, and top projects by activity — is sourced directly from your organization's usage events.

Time windows are configurable. Data reflects actual API calls as they complete.

A handful of metrics are marked coming soon: error rate, active users, and peak request volume. These require telemetry that isn't captured in the current usage events schema. We've left the placeholders visible rather than filling them with estimates.

API documentation is now off by default

Following a report from a beta tester, PolarGrid's edge gateway now ships with its interactive API documentation disabled by default. The Swagger UI, ReDoc interface, and raw OpenAPI schema were previously accessible unauthenticated on production nodes, exposing the full route catalog including admin endpoint schemas.

Every inference and admin endpoint independently enforces authentication and was unaffected — no credentials were at risk. The docs surface is now gated behind a server-side flag that defaults to off. All three endpoints return 404 on production nodes.

The published API reference at polargrid.mintlify.app remains the right place to explore the API.

Billing enforcement tightened

A gap that allowed organizations with zero credits and no payment method to make unlimited inference requests has been closed. The access check existed but wasn't being called from inference paths. It is now.

Payment failure handling is also more immediate. When Stripe reports a failed payment, inference access is blocked at the token layer before requests reach the edge — rather than only updating billing records and triggering a dunning email. Payment method verification now confirms an active card via the Stripe API directly, rather than treating the existence of a customer ID as sufficient proof.

Organizations with active credits are unaffected. Organizations with no credits and a card on file continue through — Stripe meters and charges at period close.