Up to five in a single ordered list per bot. Most production deployments use three: a premium primary like GPT-4o, a strong fallback like Claude 3.5 Sonnet, and a budget tier like Gemini 1.5 Flash. Adding OpenRouter as a fourth gives you access to dozens of underlying providers as a final safety net.

If all configured models fail to respond within their respective timeouts, SleekAI returns a friendly error message to the user explaining there's a temporary issue. This is rare with three or more tiers but possible in genuine widespread outages. The user sees a graceful message instead of a broken widget.

Only during actual failures. In normal operation the first tier responds and the others are never called. When the first fails, the retry against the second adds typically 1-2 seconds of latency. A full cascade through three tiers in a worst case adds 3-4 seconds total, still faster than most timeout windows.

Only the successful model is billed. Failed attempts typically don't consume tokens because providers don't charge for errored requests. The admin log shows per-turn which model handled it, so cost breakdowns stay accurate. Switching tiers mid-conversation does not double-bill.

Yes. Each tier has its own timeout and retry conditions. You might give the premium tier 5 seconds to respond before falling through, and the budget tier 8 seconds. This tunes latency budgets to match the relative speed of different models without holding up the user when a tier is clearly stalled.

Yes. The full conversation history, resolved variables, and system prompt carry over to every retry. The fallback model sees exactly what the primary saw, so the reply remains coherent. Even mid-conversation tier changes do not reset state. The next user turn picks up where the last bot turn left off.

Quality varies by model. The cascade order should reflect both reliability and quality preference. Most users put their preferred quality model first and accept that fallback tiers may produce slightly different phrasing or depth. For mission-critical accuracy, keep tier quality close (Sonnet then Opus, not Sonnet then a small open model).

No. The cascade works with any combination of direct OpenAI, Anthropic, and Google keys. OpenRouter is one option for the final tier because it itself spans many providers, giving extra depth. But a three-tier setup with three direct keys is fully supported and commonly used.