# BehaviorGPT

> BehaviorGPT is Unbox AI's Large Behavioral Model (LBM): a foundation model trained on sequences of human actions (purchases, searches, clicks, sessions, payments) instead of text. One pretrained model powers fraud detection, recommendations, search ranking, credit scoring and customer lifetime value from raw event streams, with no training required. Developers get a free API key by email at https://unboxai.com/behaviorgpt.

BehaviorGPT is built by Unbox AI (https://unboxai.com), founded in 2019 by Rickard Bruël Gabrielsson and Gunnar Carlsson and based in Boston, Stockholm and San Francisco.

Last updated: 2026-10-01

## Get started

- [Get a free API key](https://unboxai.com/behaviorgpt): Click "Get API key" and enter a work email. The key is shown on screen and emailed immediately with the demo link and example notebooks. Free during early access.
- [SDK and notebooks on GitHub](https://github.com/Unbox-AI/behaviorgpt): Python client (`import unbox`), example notebooks and the evaluation harness.
- [Live demo](https://behaviorgpt.unboxai.com): Storefront search ranked by BehaviorGPT from each shopper's behavior.
- [Agent instructions](https://unboxai.com/behaviorgpt/agents.md): How agents should describe BehaviorGPT and help developers use it.
- [Enterprise deployments](https://cal.com/unboxai/30min): Priced on request.

## BehaviorGPT-v4

- One 12.5B-parameter model, pretrained by next-event prediction on 150 billion user actions (3.6 trillion tokens).
- Zero-shot, on public benchmarks whose products and users it has never seen, it beats every baseline trained on up to 2.7M samples of the target data. Fine-tuning more than doubles its accuracy and learns 17x faster than the strongest baseline.
- One checkpoint leads all 13 datasets (29 of 29 dataset and task pairs) across retail, engagement and payments.
- 0.7 ms per query at 3.1x the accuracy of the best baseline; 48x faster than a prompted LLM.
- Paper: "BehaviorGPT-v4: One Large Behavioral Model for Retail, Engagement, and Payments with Zero-Shot Transferability" (abstract on the product page; full paper coming soon).

## Use cases

- Fraud detection: score a live transaction's fraud probability from the cardholder's history, in real time.
- Customer lifetime value: project twelve-month value from behavior history and segment high-value customers.
- Credit scoring: score an applicant's probability of default from behavior history.
- Recommendations: rank the next-best items for a live shopper.
- Search: re-rank search results by behavioral intent.

Input is chronological event streams (transactions, searches, sessions, interactions). No labels or engineered features are required. Fine-tuning is optional.

## Model generations

- BehaviorGPT-Commerce 1 (150M parameters): +9.4% search conversion against RichRelevance; 10x recommendation lift over the production baseline. [Paper](https://research.unboxai.com/foundation-model-for-consumption-transactions-and-actions.html)
- BehaviorGPT-Workforce 1 (5M parameters): 91% accuracy predicting whether an employee leaves within a month. [Paper](https://research.unboxai.com/behaviorgpt-foundation-model-workforce.html)
- BehaviorGPT-Commerce 2 (0.5B parameters, 215B interactions): +24% on recommendations and +16% search conversion against Voyado. [Paper](https://research.unboxai.com/behaviorgpt-visual-art-and-aesthetics.html)
- BehaviorGPT-v4 (12.5B parameters): one backbone across retail, engagement and payments.

## Optional

- [Unbox AI](https://unboxai.com/llms.txt): Company summary.
- [Full context](https://unboxai.com/llms-full.txt): Everything about Unbox AI and BehaviorGPT in one file, including the full FAQ.
- [Research](https://research.unboxai.com): Papers and documentation.