When Quizlet killed its AI tutor after relying on a single LLM provider, we saw the lesson: never bet your product on one API. Here's how we built Voklex's 5-model cascade — Groq, Gemini, Claude, Llama, GPT — with automatic failover, and what it costs to run.
The problem with one AI provider
In June 2025, Quizlet shut down Q-Chat, its AI tutor. The root cause was architectural: the feature depended on a single LLM provider, and when that relationship changed, the feature died. Millions of learners lost a tool they relied on overnight.
When we started Voklex — our AI vocabulary platform with a 64,000-word dictionary — resilience was requirement number one. If a student is revising for their 11 Plus the night before the exam, "our AI provider is down" is not an acceptable answer.
The cascade design
The core idea is simple: five providers, tried in order, with automatic failover.
- Groq — first choice. Fastest inference (sub-second for most generations) and the cheapest per token. Handles ~80% of requests.
- Gemini — first fallback. Strong at structured JSON output, generous free tier for burst traffic.
- Claude — used for nuanced content: etymology explanations and example sentences where quality matters most.
- Llama 3.3 70B — open-weight workhorse for bulk generation jobs.
- GPT — final fallback. Rarely hit, but guarantees the chain never ends in an error page.
Each request carries a timeout and a schema validator. If a provider times out, rate-limits, or returns malformed JSON, the request moves down the chain automatically. The user sees a slightly slower response — never a failure.
What it costs
The counterintuitive result: the cascade is cheaper than single-provider premium pricing. Because ~80% of traffic lands on Groq's inexpensive tier, our blended cost per generated word profile is a fraction of a penny. That's what makes the £1.99/month price point — 68% cheaper than the incumbent — sustainable.
What we'd tell other builders
- Validate every LLM response against a schema. Models hallucinate structure, not just facts.
- Cache aggressively. 785 UK curriculum word sets are pre-generated; the LLM only runs for genuinely new content.
- Design for provider death, not provider failure. Any link in the chain can be removed without a deploy.
The result: zero AI downtime since launch. If you're building an AI feature for your business and want it to survive provider outages, talk to us — this is exactly the kind of engineering we do for clients.


