Reliability report — how Nox handles failure

How Nox is engineered to stay useful when things go wrong — the fallbacks, refunds, and degradation paths built into the product — and, just as honestly, the reliability work that is not done yet.

Automatic model fallback

Nox talks to OpenAI through its newest Responses API. If that API fails mid-request, Nox automatically retries the same message through the older Chat Completions API instead of surfacing an error — so a provider-side hiccup usually costs you a moment, not your answer.

Usage is refunded when generation fails

Nox meters AI usage behind the scenes, and usage is reserved before an answer is generated. If the generation fails after the reservation — a network error, a provider outage — the reserved usage is automatically returned to your allowance. You are not charged for answers you never received.

Graceful degradation, not hard failure

When a premium capability hits its cap, Nox steps down instead of stopping: the Nox Agent's live web search degrades to its Deep Research mode when the monthly live-search allowance runs out, so research requests keep working at reduced freshness rather than failing outright.

Leo does not depend on the AI

Leo's red-flag safety screen is deterministic code that runs before any model is called. It has no AI dependency, so even if every AI provider is degraded, emergency guidance for recognized red-flag messages still fires exactly the same way.

Streaming that survives interruptions

Answers stream token-by-token over server-sent events with careful buffer handling, and the app reconciles the final message with the server afterwards — so a dropped connection mid-answer doesn't corrupt your conversation history.

Your data lives in a managed database

Conversations, memories, and account data are stored in a managed PostgreSQL database with the platform's standard backup and recovery mechanisms. Incognito conversations are the deliberate exception: they are never written to the database at all.

What we don't have yet (honestly)

Nox does not yet publish a public uptime dashboard, a formal SLA, or third-party reliability audits. Incidents are handled by the team directly. As the product matures we intend to publish uptime history the same way we already publish safety-detector metrics — measured, not promised.