---
schema_version: "newruntime-reader-dossiers-v1"
type: "reader_dossier"
slug: "gptbot"
kind: "ai_training"
operator: "OpenAI"
researched_at: "2026-10-23"
confidence: "high"
json_url: "https://newruntime.com/readers/gptbot.json"
canonical_url: "https://newruntime.com/readers/gptbot/"
---

# GPTBot

> OpenAI's crawler for content that may be used to train its generative AI foundation models.

## Dossier

- Kind: AI model crawlers
- Operator: OpenAI (https://openai.com/)
- Documentation: https://developers.openai.com/api/docs/bots
- robots.txt tokens: GPTBot
- Feeds: model training
- Respects robots.txt: Yes, per its documentation. Disallowing GPTBot signals content should not be used for training; controlled separately from OAI-SearchBot, and one crawl may serve both if both are allowed.
- Verification: Published IP ranges (https://openai.com/gptbot.json)
- Labels matched in our logs: gptbot

## Why it matters to New Runtime (our interpretation)

Training reads mean New Runtime's framing can end up baked into future OpenAI model weights — long-horizon influence, no attribution.

## Observed behaviour

- Baseline 2026-10-23 (93-day window): 2,236 requests on 43 days; first seen 2026-07-23, last seen 2026-10-22; Returning.
- Reading mix (crawler journeys): 0 content reads, 2 feed or discovery reads, 0 machine-consumption reads; 0 unique materials.
- Identity check: claimed unverified 8, verified 2 requests.
- Live numbers: https://newruntime.com/bot-traffic-history.json and https://newruntime.com/crawler-journeys.json

## Sources

- https://developers.openai.com/api/docs/bots

Identity: A family name is what a client claims in its user agent. It is verified only when its request came from the operator's published IP ranges or passed forward-confirmed reverse DNS, as reported per family in /crawler-journeys.json. Anyone can claim any name.

All dossiers: https://newruntime.com/readers.md