Grok API access through one OpenAI-compatible endpoint

Grok API is not Groq API

Grok refers to xAI model families, while Groq is a different inference platform. Search results and client configuration often mix the names, so verify the exact model identifier and provider before testing. CLODEX displays model slugs in the authenticated catalog and pricing interface.

Use https://clodex.xyz/v1 for OpenAI-compatible clients when the selected Grok channel supports that protocol. Keep model selection explicit and do not infer a provider from a shortened label.

Grok API integration workflow

Begin with model discovery, a short non-streaming request and an error case. Then test streaming, tools and long context if your application needs them. Capture the request ID so gateway and upstream failures can be correlated.

A 200 response that closes before the terminal event is not a completed stream. The client should surface partial output separately and avoid automatic side effects until completion is confirmed.

Grok API pricing and availability

Use the live CLODEX catalog for current pricing and availability. Model cost can include input, output, cache and fixed-usage components, and active channels can change. Avoid publishing hard-coded prices inside application UI unless they are refreshed automatically.

For high-volume workloads, measure success rate and total task cost as well as token price. A model that needs fewer retries or turns can be cheaper for the final outcome even when one unit price is higher.

Safe multi-model fallback

Fallback is an application decision, not only a gateway setting. Product owners should understand possible changes in answer style, safety behavior, latency and cost.

Benchmark Grok on the same dated prompt set used for other providers and record the model slug, sample size and scoring method. Re-run the comparison after material model or gateway changes instead of treating an old leaderboard result as permanent.

Include failure rate, first-token latency and completed-task cost so the comparison reflects production behavior rather than answer preference alone.

Repeat tests across short and long contexts.

• Document the approved fallback model and why it is acceptable.
• Do not switch models inside an already running stream.
• Preserve requested and actual model names in logs.
• Revalidate structured output after any fallback.
• Stop rather than substitute when policy requires a specific provider.