This promptfoo example compares OpenAI, Anthropic, and Groq models routed through Cloudflare AI Gateway, demonstrating gateway-based caching, rate limiting, and analytics on top of multiple providers. The same "explain a topic simply" prompt runs against gpt-4o-mini, claude-haiku-4-5-20251001, and llama-3.1-8b-instant, each addressed via a cloudflare-gateway:<provider>:<model> provider ID.

Use this example when you already route (or plan to route) LLM traffic through Cloudflare AI Gateway and want a promptfoo eval that exercises multiple providers through the same gateway, or when you want to compare gateway-routed provider outputs on a shared prompt. It is not useful if you call providers directly without Cloudflare AI Gateway, since it depends entirely on gateway configuration (CLOUDFLARE_ACCOUNT_ID, CLOUDFLARE_GATEWAY_ID, and optionally CF_AIG_TOKEN for authenticated gateways).

Each provider config supplies accountId and gatewayId from environment variables, plus optional cfAigToken for authenticated gateways and per-model generation settings (temperature, max_tokens). Provider API keys (OPENAI_API_KEY, ANTHROPIC_API_KEY, GROQ_API_KEY) can be set directly, or omitted entirely if Bring Your Own Keys (BYOK) is configured in the Cloudflare dashboard. Tests supply a topic variable and assert the response contains at least one expected keyword:

providers:
  # Route OpenAI requests through Cloudflare AI Gateway
  - id: cloudflare-gateway:openai:gpt-4o-mini
    config:
      accountId: '{{env.CLOUDFLARE_ACCOUNT_ID}}'
      gatewayId: '{{env.CLOUDFLARE_GATEWAY_ID}}'
      temperature: 0.7
      max_tokens: 200

Run it with npx promptfoo eval -c examples/cloudflare-gateway/promptfooconfig.yaml.

Routes through Cloudflare AI Gateway to OpenAI (gpt-4o-mini), Anthropic (claude-haiku-4-5-20251001), and Groq (llama-3.1-8b-instant), and supports Cloudflare's BYOK feature so provider keys can live in the Cloudflare dashboard instead of local environment variables.

Teams already using Cloudflare AI Gateway for caching, rate limiting, or analytics across multiple LLM providers who want a promptfoo eval template that exercises that gateway setup directly.