Skip to content
aicypress.xyz

Bring your own API keys

Live benchmarks ask the real AI assistants your buyers use — ChatGPT, Claude, Gemini, and Perplexity — and that happens on your own API keys. You pay each provider directly for exactly what your runs use; Cypress never resells AI usage or marks it up.

Simulated runs are free and need no keys — they’re for exploring the product. Live runs are where the real answers come from.

Each model in a run needs its provider’s key. Add them in Settings → ai keys — keys are validated the moment you save them.

ProviderModels benchmarkedCreate your key atKey looks like
OpenAI (ChatGPT)GPT-5.5, GPT-5.4 miniplatform.openai.com/api-keyssk-...
Anthropic (Claude)Sonnet 4.6, Haiku 4.5console.anthropic.com/settings/keyssk-ant-...
Google (Gemini)Gemini 3.5 Flash, 3.1 Flash-Liteaistudio.google.com/apikeyAIza...
PerplexitySonar Pro, Sonarperplexity.ai → Settings → APIpplx-...

Every provider requires a funded account (or active free tier) for API usage. You only need keys for the providers you want benchmarked — start with one and add more anytime.

Live models answer with web search enabled — that’s the point: you see which answers are grounded in live web results (and which sites they cite) versus answered from the model’s training memory.

Rough per-answer cost: $0.01–$0.03 including the provider’s web-search fee (about a cent per searched answer) plus model tokens. A typical benchmark — 16 buyer prompts across all four providers’ flagship models — lands around $1–2 total, spread across your four provider accounts. Cheaper models (Haiku, GPT-5.4 mini, Flash-Lite, Sonar) cost a fraction of that.

  • Encrypted at rest with AES-256-GCM; never returned to your browser after saving (you’ll only ever see a masked prefix).
  • Used exclusively for the live benchmark runs you launch — nothing runs in the background without you pressing the button.
  • Removable anytime in Settings; runs in flight will error their remaining cells rather than use a removed key.
  • Cypress’s own API key is used only for onboarding (drafting your profile and prompts) and for the per-answer judge that scores recommendation strength — never for your benchmark answers.