canonical: https://jentic.com/apis/openai.com/openai

# OpenAI API

Generate text, images, embeddings, and audio with 126 endpoints spanning chat completions, assistants, fine-tuning, batch processing, and moderation. Supports GPT-4o, GPT-4, DALL-E 3, Whisper, and TTS models through a single unified REST interface with bearer-token authentication. Handles multi-turn conversations, function calling, vision inputs, and JSON mode for structured outputs.

## For AI agents

Generate text completions, create embeddings, transcribe audio, and produce images using OpenAI models. Supports function calling, multi-modal inputs, and structured JSON output for agent workflows.

## Scope

Does not handle vector storage, fine-tuned model hosting infrastructure, or real-time streaming WebSocket connections - use for AI model inference and training only.

## Capabilities

- Generate multi-turn chat completions with function calling and tool use
- Create vector embeddings for semantic search and retrieval-augmented generation
- Transcribe and translate audio files using Whisper models
- Synthesize natural-sounding speech from text with multiple voice options
- Generate and edit images using DALL-E 3
- Fine-tune base models on custom training datasets with checkpointing
- Run batch inference jobs for high-volume asynchronous processing

## Use cases

### AI Agent Tool Calling via Jentic

AI agents discover and invoke OpenAI chat completions through Jentic's intent-based search, enabling multi-step reasoning with function calling. Agents search for the operation by intent, receive the input schema, and execute completions without manual SDK setup. Supports tool_choice parameters for deterministic function invocation and parallel tool calls for complex agent workflows.

Example prompt: Search Jentic for 'generate a chat completion with function calling', load the /chat/completions schema, and execute a request with model gpt-4o and a weather tool definition

### Retrieval-Augmented Generation Pipeline

Build RAG pipelines by generating vector embeddings with the /embeddings endpoint and combining them with chat completions for grounded answers. The text-embedding-3-small model produces 1536-dimensional vectors at low cost, while text-embedding-3-large offers higher accuracy for production retrieval systems. Agents can index thousands of documents and query relevant chunks in a single orchestration flow.

Example prompt: Generate embeddings for 10 document chunks using text-embedding-3-small via POST /embeddings and store the resulting vectors

### Audio Transcription and Translation

Transcribe audio files in 57 languages using the Whisper model through the /audio/transcriptions endpoint, or translate non-English audio directly to English text via /audio/translations. Accepts mp3, mp4, mpeg, mpga, m4a, wav, and webm formats up to 25 MB. Returns timestamped segments for subtitle generation or plain text for downstream processing.

Example prompt: Transcribe a 5-minute WAV audio file using POST /audio/transcriptions with model whisper-1 and return timestamped JSON segments

### Batch Inference for High-Volume Processing

Process thousands of requests asynchronously at 50% reduced cost using the Batch API. Submit JSONL files containing chat completion, embedding, or moderation requests via POST /batches, then poll for completion. Results are returned as downloadable output files. Ideal for nightly report generation, bulk content moderation, or large-scale embedding jobs that do not require real-time responses.

Example prompt: Upload a JSONL file with 500 chat completion requests via POST /files, create a batch job via POST /batches, and poll GET /batches/{batch_id} until completion

### Custom Model Fine-Tuning

Fine-tune GPT-4o-mini or GPT-3.5-turbo on domain-specific training data to improve accuracy for specialized tasks like classification, extraction, or tone matching. Upload training JSONL via /files, launch a fine-tuning job via POST /fine_tuning/jobs, monitor with checkpoints, and deploy the resulting model for inference. Supports validation files and hyperparameter configuration.

Example prompt: Upload a training JSONL file via POST /files with purpose 'fine-tune', then create a fine-tuning job via POST /fine_tuning/jobs with model gpt-4o-mini-2024-07-18

## Key endpoints

| Method | Path | Description |
| --- | --- | --- |
| POST | /chat/completions | Generate a chat completion with optional function calling |
| POST | /embeddings | Create vector embeddings from input text |
| POST | /audio/transcriptions | Transcribe audio to text using Whisper |
| POST | /audio/speech | Generate speech audio from text input |
| POST | /images/generations | Generate images from text prompts |
| POST | /fine_tuning/jobs | Create a model fine-tuning job |
| POST | /batches | Create an asynchronous batch processing job |
| GET | /models | List all available models |

## Key resources

- **Chat Completions** — Generate text responses with multi-turn context, function calling, vision, and JSON mode
- **Embeddings** — Create vector representations of text for search and clustering
- **Audio** — Transcribe, translate, and synthesize speech
- **Images** — Generate and edit images from text descriptions
- **Fine-tuning** — Train custom models on domain-specific datasets
- **Assistants** — Build stateful AI assistants with tool use, file retrieval, and code execution
- **Batch** — Process bulk requests asynchronously at reduced cost

## Why Jentic

- **Setup:** Wiring the OpenAI API by hand means setting the Bearer token on every request and shaping the chat, embeddings, audio, and image payloads yourself. Through Jentic you install once, import the OpenAI API from the API Directory, store the token once, and your agent calls it.
- **Permission scoping:** OpenAI takes its inputs in the request body rather than a resource id in the URL path, so limit the agent to the operations it needs, such as creating a chat completion or generating embeddings. You credit the agent only with the operations you allow, so fine-tuning jobs stay out unless you add them.
- **Credential handling:** Your OpenAI token is stored once, encrypted, by your own Jentic One instance and injected at execution time. It never enters the agent's prompt, logs, or context.
- **Discovery method:** Agents search Jentic by intent such as 'generate a chat completion' or 'create text embeddings', and Jentic returns the matching OpenAI operation with its input schema so the agent calls the right endpoint without browsing the reference docs.

## Related APIs

- **Anthropic API** — Claude models with 200K context window, focused on safety and instruction-following
- **Cohere API** — Enterprise-focused LLM API with native RAG support and reranking
- **Mistral AI API** — Open-weight models with competitive performance at lower cost
- **Pinecone API** — Vector database for storing and querying OpenAI embeddings
- **ElevenLabs API** — Advanced voice synthesis and cloning beyond OpenAI's TTS capabilities

## FAQ

### What authentication does the OpenAI API use?

The OpenAI API uses Bearer token authentication. You pass your API key in the Authorization header as 'Bearer sk-...'. Through Jentic, your OpenAI API key is stored encrypted in your Jentic One instance and agents receive scoped access tokens, so the raw secret key never enters the agent context.

### Can I generate structured JSON output with the OpenAI API?

Yes. The POST /chat/completions endpoint supports a response_format parameter set to 'json_object' which constrains the model to output valid JSON. For stricter schemas, use function calling with a JSON Schema definition to guarantee the output matches your expected structure.

### What are the rate limits for the OpenAI API?

Rate limits vary by model tier and organization. GPT-4o allows up to 10,000 requests per minute on Tier 5 accounts, while lower tiers start at 500 RPM. Embedding models support higher throughput. The API returns x-ratelimit-remaining and x-ratelimit-reset headers with each response for programmatic tracking.

### How do I generate embeddings for a RAG pipeline through Jentic?

Search Jentic for 'create text embeddings' to discover the POST /embeddings operation. Load the schema, then execute with model 'text-embedding-3-small' and your input texts. Jentic returns the operation schema with parameters so your agent can call it directly without browsing OpenAI documentation. Install the SDK with pip install jentic to get started.

### What models are available through the OpenAI API?

The GET /models endpoint lists all available models. Current options include GPT-4o and GPT-4o-mini for chat, text-embedding-3-small and text-embedding-3-large for embeddings, whisper-1 for transcription, tts-1 and tts-1-hd for speech synthesis, and dall-e-3 for image generation. Fine-tuned models also appear in this list after training completes.

### Can I use function calling to connect GPT to external tools?

Yes. Pass a tools array to POST /chat/completions with function definitions including name, description, and a JSON Schema for parameters. The model returns a tool_calls array when it decides to invoke a function. You execute the function and pass results back in a follow-up message with role 'tool'. Supports parallel tool calls and forced tool_choice.

### Is the OpenAI API free to use?

OpenAI offers a free tier with limited credits for new accounts. After that, pricing is pay-per-token: GPT-4o costs $2.50 per million input tokens and $10 per million output tokens. Embedding models are significantly cheaper at $0.02 per million tokens for text-embedding-3-small. The Batch API provides a 50% discount for non-real-time workloads.

### Can I limit what my agent is allowed to do with the OpenAI API?

Yes. Because you run Jentic One yourself, your own rules decide which OpenAI operations and credentials the agent may use. Since OpenAI takes its inputs in the request body rather than a resource id in the URL path, you can credit the agent with only the operations it needs, such as creating a chat completion or generating embeddings. Operations you do not grant, like creating fine-tuning jobs, stay off limits unless you add them.