# New ChatGPT API, prompt-to-reply as structured JSON

> Source: https://scrape.do/documentation/changelog/2026-05-12-chatgpt-api/

2026-05-12 · feature

A new dedicated endpoint sends a prompt to `chatgpt.com` and returns the assembled assistant message as structured JSON. One GET call. No account. No cookies. No streaming protocol to handle on your side.

## What you get back

The response is the full message envelope, ready to walk:

| Path | What's there |
|---|---|
| `data.message.content.parts[0]` | The assistant's reply text, already cleaned of inline citation delimiters |
| `data.message.metadata.model_slug` | Which model produced the reply (e.g., `gpt-5`) |
| `data.message.metadata.finish_details.type` | Stop reason (`"stop"` for a normal completion) |
| `data.message.metadata.content_references[]` | Inline citations with resolved `safe_urls[]` |
| `data.message.id` | Unique message id |
| `data.conversation_id` | Conversation id for logging or audit |

## Why it matters

- **Stateless from your side.** Each request is independent. Build prompt batches, evaluation harnesses, or research pipelines without juggling conversation IDs or session tokens.
- **Citations already parsed.** When the model cites a source, you get the resolved URL in `content_references[].safe_urls[]`. Inline citation markers in the raw stream are stripped, so `parts[0]` reads cleanly.
- **Flat per-call price.** Every successful request is 25 credits, regardless of prompt length or response size. Failed calls don't consume credits.

Each request costs 25 credits. See the [ChatGPT API documentation](/documentation/chatgpt-api) for the full request shape, response envelope, and citation reference fields.