Gemini to OpenAI API Converter

The converter is a deterministic transformation of a published schema — there is no model call. It reads your Gemini generateContent payload, applies a fixed set of rules transcribed from the official Google and OpenAI API references, and emits the equivalent OpenAI Chat Completions JSON. Every rule is a static mapping, so the same input always produces the same output, and two worked examples are reconciled field-by-field on every load.

Requests are rebuilt turn by turn:

1. System prompt. The top-level systemInstruction.parts[].text is hoisted into a leading {role:"system"} message — OpenAI has no separate instruction field.
2. contents → messages. Order is preserved. role:"model" becomes assistant; user stays user. Text parts become a content string, or a {type:"text"}/{type:"image_url"} content-part array when an inlineData part is present (encoded as a data: URL).
3. functionCall → tool_calls. A model functionCall{name,args} becomes an assistant tool_calls entry. The args object is JSON-stringified into arguments. Gemini calls carry no id, so a stable call_0, call_1… is synthesized in order.
4. functionResponse → tool message. A user-turn functionResponse{name,response} becomes a separate {role:"tool",tool_call_id,content} message, linked back to the matching synthesized call id by name and order.
5. tools & toolConfig. tools[].functionDeclarations[] is unwrapped into tools[].function (upper-case Type enums like STRING are lower-cased to JSON Schema). functionCallingConfig.mode: AUTO→"auto", ANY→"required" (or a named function), NONE→"none".
6. generationConfig. maxOutputTokens → max_tokens (or max_completion_tokens per the toggle), topP → top_p, stopSequences → stop, candidateCount → n (flagged), and responseMimeType/responseSchema → response_format.

Responses wrap candidates[] into a chat.completion object: each candidate becomes a choice{index,message,finish_reason}, part text becomes message.content, and finishReason maps to finish_reason (STOP→stop, MAX_TOKENS→length, SAFETY/RECITATION→content_filter; forced tool_calls when calls are present). usageMetadata maps to usage — promptTokenCount→prompt_tokens, candidatesTokenCount→completion_tokens, totalTokenCount→total_tokens. Fields with no clean target — topK, safetySettings, thinking parts, extra candidates — are listed in the notes panel, never silently dropped.

• Google Gemini API — generateContent reference
• Google Gemini API — function calling (functionCall, functionResponse, functionCallingConfig)
• OpenAI API Reference — Chat Completions

The mapping rules were last cross-checked against the official Gemini and OpenAI references on 2026-07-12. This page is reviewed whenever either provider ships a breaking request- or response-schema change.