Use o formato nativo da API Gemini por meio da CometAPI para geração de texto, entrada multimodal, thinking/reasoning, function calling, grounding com Google Search, modo JSON e streaming.
from google import genai
client = genai.Client(
api_key="<COMETAPI_KEY>",
http_options={"api_version": "v1beta", "base_url": "https://api.cometapi.com"},
)
response = client.models.generate_content(
model="gemini-2.5-flash",
contents="Explain how AI works in a few words",
)
print(response.text){
"candidates": [
{
"content": {
"role": "<string>",
"parts": [
{
"text": "<string>",
"functionCall": {
"name": "<string>",
"args": {}
},
"inlineData": {
"mimeType": "<string>",
"data": "<string>"
},
"thought": true
}
]
},
"finishReason": "STOP",
"safetyRatings": [
{
"category": "<string>",
"probability": "<string>",
"blocked": true
}
],
"citationMetadata": {
"citationSources": [
{
"startIndex": 123,
"endIndex": 123,
"uri": "<string>",
"license": "<string>"
}
]
},
"tokenCount": 123,
"avgLogprobs": 123,
"groundingMetadata": {
"groundingChunks": [
{
"web": {
"uri": "<string>",
"title": "<string>"
}
}
],
"groundingSupports": [
{
"groundingChunkIndices": [
123
],
"confidenceScores": [
123
],
"segment": {
"startIndex": 123,
"endIndex": 123,
"text": "<string>"
}
}
],
"webSearchQueries": [
"<string>"
]
},
"index": 123
}
],
"promptFeedback": {
"blockReason": "SAFETY",
"safetyRatings": [
{
"category": "<string>",
"probability": "<string>",
"blocked": true
}
]
},
"usageMetadata": {
"promptTokenCount": 123,
"candidatesTokenCount": 123,
"totalTokenCount": 123,
"trafficType": "<string>",
"thoughtsTokenCount": 123,
"promptTokensDetails": [
{
"modality": "<string>",
"tokenCount": 123
}
],
"candidatesTokensDetails": [
{
"modality": "<string>",
"tokenCount": 123
}
]
},
"modelVersion": "<string>",
"createTime": "<string>",
"responseId": "<string>"
}| Configuração | Padrão do Google | CometAPI |
|---|---|---|
| URL base | generativelanguage.googleapis.com | api.cometapi.com |
| Chave de API | $GEMINI_API_KEY | $COMETAPI_KEY |
x-goog-api-key quanto Authorization: Bearer são compatíveis para autenticação.thinkingLevel para controlar a profundidade do reasoning. Níveis disponíveis: MINIMAL, LOW, MEDIUM, HIGH.curl "https://api.cometapi.com/v1beta/models/gemini-3.1-pro-preview:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $COMETAPI_KEY" \
-d '{
"contents": [{"parts": [{"text": "Explain quantum physics simply."}]}],
"generationConfig": {
"thinkingConfig": {"thinkingLevel": "LOW"}
}
}'
thinkingBudget para controle refinado em nível de token:0 — desabilita o thinking-1 — dinâmico (o modelo decide, padrão)> 0 — orçamento específico de token (por exemplo, 1024, 2048)curl "https://api.cometapi.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $COMETAPI_KEY" \
-d '{
"contents": [{"parts": [{"text": "Solve this logic puzzle step by step."}]}],
"generationConfig": {
"thinkingConfig": {"thinkingBudget": 2048}
}
}'
thinkingLevel com modelos Gemini 2.5 (ou thinkingBudget com modelos Gemini 3) pode causar erros. Use o parâmetro correto para a versão do seu modelo.streamGenerateContent?alt=sse como operador para receber Server-Sent Events à medida que o modelo gera conteúdo. Cada evento SSE contém uma linha data: com um objeto JSON GenerateContentResponse.
curl "https://api.cometapi.com/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $COMETAPI_KEY" \
--no-buffer \
-d '{
"contents": [{"parts": [{"text": "Write a short poem about the stars"}]}]
}'
systemInstruction:
curl "https://api.cometapi.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $COMETAPI_KEY" \
-d '{
"contents": [{"parts": [{"text": "What is 2+2?"}]}],
"systemInstruction": {
"parts": [{"text": "You are a math tutor. Always show your work."}]
}
}'
responseMimeType. Opcionalmente, forneça um responseSchema para validação estrita de schema:
curl "https://api.cometapi.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $COMETAPI_KEY" \
-d '{
"contents": [{"parts": [{"text": "List 3 planets with their distances from the sun"}]}],
"generationConfig": {
"responseMimeType": "application/json"
}
}'
googleSearch:
curl "https://api.cometapi.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $COMETAPI_KEY" \
-d '{
"contents": [{"parts": [{"text": "Who won the euro 2024?"}]}],
"tools": [{"google_search": {}}]
}'
groundingMetadata com URLs de origem e pontuações de confiança.
{
"candidates": [
{
"content": {
"role": "model",
"parts": [{"text": "Hello"}]
},
"finishReason": "STOP",
"avgLogprobs": -0.0023
}
],
"usageMetadata": {
"promptTokenCount": 5,
"candidatesTokenCount": 1,
"totalTokenCount": 30,
"trafficType": "ON_DEMAND",
"thoughtsTokenCount": 24,
"promptTokensDetails": [{"modality": "TEXT", "tokenCount": 5}],
"candidatesTokensDetails": [{"modality": "TEXT", "tokenCount": 1}]
},
"modelVersion": "gemini-2.5-flash",
"createTime": "2026-03-25T04:21:43.756483Z",
"responseId": "CeynaY3LDtvG4_UP0qaCuQY"
}
thoughtsTokenCount em usageMetadata mostra quantos tokens o modelo gastou em raciocínio interno, mesmo quando a saída de thinking não está incluída na resposta.| Recurso | Gemini nativo (/v1beta/models/...) | Compatível com OpenAI (/v1/chat/completions) |
|---|---|---|
| Controle de thinking | thinkingConfig com thinkingLevel / thinkingBudget | Não disponível |
| Grounding com Google Search | tools: [\{"google_search": \{\}\}] | Não disponível |
| Grounding com Google Maps | tools: [\{"googleMaps": \{\}\}] | Não disponível |
| Modalidade de geração de imagem | responseModalities: ["IMAGE"] | Não disponível |
| Cabeçalho de autenticação | x-goog-api-key ou Bearer | Apenas Bearer |
| Formato da resposta | Gemini nativo (candidates, parts) | Formato OpenAI (choices, message) |
Your CometAPI key passed via the x-goog-api-key header. Bearer token authentication (Authorization: Bearer <key>) is also supported.
The Gemini model ID to use. See the Models page for current Gemini model IDs.
"gemini-2.5-flash"
The operation to perform. Use generateContent for synchronous responses, or streamGenerateContent?alt=sse for Server-Sent Events streaming.
generateContent, streamGenerateContent?alt=sse "generateContent"
The conversation history and current input. For single-turn queries, provide a single item. For multi-turn conversations, include all previous turns.
Show child attributes
System instructions that guide the model's behavior across the entire conversation. Text only.
Show child attributes
Tools the model may use to generate responses. Supports function declarations, Google Search, Google Maps, and code execution.
Show child attributes
Configuration for tool usage, such as function calling mode.
Show child attributes
Safety filter settings. Override default thresholds for specific harm categories.
Show child attributes
Configuration for model generation behavior including temperature, output length, and response format.
Show child attributes
The name of cached content to use as context. Format: cachedContents/{id}. See the Gemini context caching documentation for details.
Successful response. For streaming requests, the response is a stream of SSE events, each containing a GenerateContentResponse JSON object prefixed with data:.
The generated response candidates.
Show child attributes
Feedback on the prompt, including safety blocking information.
Show child attributes
Token usage statistics for the request.
Show child attributes
The model version that generated this response.
The timestamp when this response was created (ISO 8601 format).
Unique identifier for this response.
from google import genai
client = genai.Client(
api_key="<COMETAPI_KEY>",
http_options={"api_version": "v1beta", "base_url": "https://api.cometapi.com"},
)
response = client.models.generate_content(
model="gemini-2.5-flash",
contents="Explain how AI works in a few words",
)
print(response.text){
"candidates": [
{
"content": {
"role": "<string>",
"parts": [
{
"text": "<string>",
"functionCall": {
"name": "<string>",
"args": {}
},
"inlineData": {
"mimeType": "<string>",
"data": "<string>"
},
"thought": true
}
]
},
"finishReason": "STOP",
"safetyRatings": [
{
"category": "<string>",
"probability": "<string>",
"blocked": true
}
],
"citationMetadata": {
"citationSources": [
{
"startIndex": 123,
"endIndex": 123,
"uri": "<string>",
"license": "<string>"
}
]
},
"tokenCount": 123,
"avgLogprobs": 123,
"groundingMetadata": {
"groundingChunks": [
{
"web": {
"uri": "<string>",
"title": "<string>"
}
}
],
"groundingSupports": [
{
"groundingChunkIndices": [
123
],
"confidenceScores": [
123
],
"segment": {
"startIndex": 123,
"endIndex": 123,
"text": "<string>"
}
}
],
"webSearchQueries": [
"<string>"
]
},
"index": 123
}
],
"promptFeedback": {
"blockReason": "SAFETY",
"safetyRatings": [
{
"category": "<string>",
"probability": "<string>",
"blocked": true
}
]
},
"usageMetadata": {
"promptTokenCount": 123,
"candidatesTokenCount": 123,
"totalTokenCount": 123,
"trafficType": "<string>",
"thoughtsTokenCount": 123,
"promptTokensDetails": [
{
"modality": "<string>",
"tokenCount": 123
}
],
"candidatesTokensDetails": [
{
"modality": "<string>",
"tokenCount": 123
}
]
},
"modelVersion": "<string>",
"createTime": "<string>",
"responseId": "<string>"
}