searchquery
purchaseparty decor
tripvacation
streamlullabies
returnswaddles
card paymentpharmacy
BehaviorGPT-Commerce 1
Purchase and session sequences modeled as the language of consumption.
- Parameters
- 150M
The foundation model for behavior
BehaviorGPT turns actions into real-time predictions.
Introducing BehaviorGPT-v4
It reads behavior the way an LLM reads language
An LLM predicts the next token from a fixed vocabulary. BehaviorGPT predicts the next action, where each event can carry text, imagery, a time, a place or a price, and the candidates form an open catalog.
Languagetokens from text
Behavioractions from the real world
Zero-shot, it beats models trained on your data
One 12.5B-parameter model, pretrained on 150 billion user actions, evaluated on public benchmarks whose products and users it has never seen.
Built on the SDK, running on BehaviorGPT-v4
Pre-trained foundation model for making predictions on behavioral data
No training required. Fine-tuning is optional.
import unbox
client = unbox.UnboxAI(api_key="your_prod_key")
# the customer's live transaction from the auth stream
def authorize(event):
# 1. Load the cardholder's transaction history
user = client.history.load(user_id=event.card_hash)
# 2. Score how likely this transaction is fraud
result = client.score(history=user, action=event)
# 3. Act on the Unbox intelligence score
if result.probability > 0.95:
return "DECLINED: BEHAVIORAL_ANOMALY"
return "APPROVED"import unbox
client = unbox.UnboxAI(api_key="your_prod_key")
# project a customer's value from their behavior
def lifetime_value(user_id):
# 1. Load the customer's full behavior history
user = client.history.load(user_id=user_id)
# 2. Predict expected value over the next 12 months
result = client.predict(history=user, target="clv_12m")
# 3. Act on the projected lifetime value
if result.value > 5000:
return "SEGMENT: HIGH_VALUE"
return "SEGMENT: STANDARD"
import unbox
client = unbox.UnboxAI(api_key="your_prod_key")
# assess default risk for a credit application
def underwrite(application):
# 1. Load the applicant's behavior history
user = client.history.load(user_id=application.user_id)
# 2. Score the probability of default
result = client.score(history=user, action=application)
# 3. Act on the Unbox risk score
if result.probability > 0.30:
return "DECISION: DECLINE"
return "DECISION: APPROVE"
import unbox
client = unbox.UnboxAI(api_key="your_prod_key")
# rank the next-best items for a live shopper
def recommend(session):
# 1. Load the shopper's behavior history
user = client.history.load(user_id=session.user_id)
# 2. Rank the catalog by predicted intent
result = client.rank(history=user, candidates=session.catalog)
# 3. Return the top items to surface
return result.top(k=10)
import unbox
client = unbox.UnboxAI(api_key="your_prod_key")
# re-rank search results by behavioral intent
def search(query, session):
# 1. Load the user's behavior history
user = client.history.load(user_id=session.user_id)
# 2. Retrieve and rank results for the query
result = client.search(history=user, query=query)
# 3. Return the ranked results
return result.top(k=20)
One recipe, each generation trained on a new domain
BehaviorGPT-Workforce 1
Employee action sequences, predicting workforce dynamics from behavior rather than surveys.
- Parameters
- 5M
BehaviorGPT-Commerce 2
The second commerce generation, learning taste from what people do rather than from pixels.
- Parameters
- 0.5B
BehaviorGPT-V4
One backbone unifying behavior across retail and payments, transferring zero-shot across domains.
- Parameters
- 12.5B
FAQs
What is BehaviorGPT?
BehaviorGPT is a foundation model for human behavior. It learns directly from sequences of actions such as purchases, searches, and clicks, with no hand-engineered features. One pretrained model transfers across catalogs, companies, and tasks. The current version, BehaviorGPT-v4, has 12.5B parameters and was pretrained on 150 billion user actions.
What is a Large Behavioral Model?
A Large Behavioral Model (LBM) is a foundation model trained on chronological sequences of human actions instead of text. It predicts the next action from the actions before it, and what it learns transfers to tasks like fraud detection and recommendations. Unbox AI introduced the term, and BehaviorGPT is its LBM.
How is this different from a recommendation engine?
A recommender scores items against a user profile or matches similar users. BehaviorGPT models the sequence itself, so earlier actions change how the next one is read. That context lets it identify intent instead of scoring each action in isolation.
How does BehaviorGPT compare to using an LLM?
An LLM reads text and generates its answer token by token. BehaviorGPT reads the event sequence directly and ranks a whole catalogue in one forward pass: 0.7 ms per query, 48 times faster than a prompted LLM, and ahead of LLM baselines on accuracy in the BehaviorGPT-v4 benchmarks. Use an LLM for language, and BehaviorGPT for predicting what people do next.
Can BehaviorGPT be used for fraud detection?
Yes. The same action can be routine or suspicious depending on what preceded it, and BehaviorGPT models that sequence directly. It scores risk in real time as a session unfolds, and the model that powers recommendations adapts to fraud and risk without retraining from scratch.
What results has BehaviorGPT shown?
Zero-shot, BehaviorGPT-v4 beats every baseline trained on up to 2.7 million samples of the target data, and one checkpoint leads all 13 public datasets it was tested on across retail, engagement, and payments. In production A/B tests, BehaviorGPT models have increased sales by double-digit percentages against incumbent systems, including +24% on recommendations and +16% search conversion.
What data does it need?
Chronological event streams: transactions, searches, sessions, and interactions. No labels or engineered features are required.
Is it available as an API or product yet?
Yes. Click Get API key, enter your work email, and your API key arrives by email together with a link to the demo and the example notebooks. Enterprise deployments are available separately.
How much does BehaviorGPT cost?
The developer API key is free during early access. Enterprise deployments are priced on request: book a call.
How fast do I get an API key?
Immediately. The key is shown on screen as soon as you sign up and emailed to you at the same time, so keep that email. If you sign up again with the same address, we point you back to the original email.
Is there a technical paper or documentation?
Yes. Our papers and documentation are on the research page. The BehaviorGPT-v4 abstract is on this page, and the full paper is coming soon.
Curate my wall