GPTZero Release Notes: AI Detection Model & API Updates

Alex Adam

10 min read

This page covers the history of all GPTZero AI detection model releases, describing the architecture and data changes, as well as the resulting user benefits.

Added New features
Improved Enhancements
Fixed Bug fixes
Removed Deprecated

Model 4.11b (aka 4.1o)

· modelVersion2026-10-07-base
Improved
  • Improved recall on sub-classes AI paraphrased, mixed and polished.
  • Improved performance on third party benchmarks.
  • Improved document chunking and aggregation.

Model 4.10b (aka 4o)

· modelVersion2026-09-13-base
Improved
  • Decreased false positive rate on a variety of domains
  • Improved performance on third party benchmarks

Model 4.9b

· modelVersion2026-08-09-base
Improved
  • Improved recall on frontier LLMs including Claude 5, GPT 5.6, Gemini 3.6, and Grok 4.5 series models

Model 4.8b

· modelVersion2026-08-01-base
Added
  • Robustness to headers; headers are now excluded from our model’s input

Model 4.7b

· modelVersion2026-07-04-base
Improved
  • Improved performance on academic benchmarks
  • Reduced FPR on customer-aligned data
  • Improved advanced scan consistency

Model 4.1m

· modelVersion2026-06-10-multilingual
Improved
  • Data balance adjustment for news domain

Model 4.6b

· modelVersion2026-05-11-base
Improved
  • Data balance made to be more relevant for common writing domains

Model 4.5b

· modelVersion2026-05-06-base
Improved
  • More granular mixed text predictions and identification of interleaving mixed texts

Model 4.4b

· modelVersion2026-03-30-base
Improved
  • Improved recall on frontier LLMs and open source LLMs
  • Reduced false positive rate

Model 4.3b

· modelVersion2026-03-11-base
Improved
  • Improved performance on the AI polished class

Model 4.2b

· modelVersion2026-02-14-base
Improved
  • Decreased reliance on special characters

Model 4.1b

· modelVersion2026-01-20-base
Improved
  • Improved robustness to named entities
  • Improved performance on classic texts, Q&A datasets, reviews, legal, and government documents.

Model 3.15b

· modelVersion2025-12-18-base
Improved
  • Improved robustness to AI paraphrasers

Model 3.7m

· modelVersion2025-12-04-multilingual
Improved
  • Improved performance on latest language model providers for French, Spanish, German, Portuguese.
  • Reduced FPR on formally written multilingual documents

Model 3.14b

· modelVersion2025-11-28-base
Improved
  • Increased robustness against word swapping

Model 3.6m

· modelVersion2025-11-20-multilingual
Improved
  • Reduced FPR on formally written multilingual documents

Model 3.13b

· modelVersion2025-11-13-base
Improved
  • Reduced FPR on documents with dates

Model 3.5m

· modelVersion2025-11-09-multilingual
Added
  • Added support for Turkish, Hindi, Dutch, Vietnamese and Indonesian languages.

Model 3.12b

· modelVersion2025-10-30-base
Improved
  • Higher AI recall and reduced FPR on creative writing and scientific articles.

Model 3.4m

· modelVersion2025-10-24-multilingual
Improved
  • Improved AI detection on French, Spanish, Portuguese, German.
Added
  • Added support for Arabic, Italian, Korean, Chinese and Japanese languages.

Model 3.11b

· modelVersion2025-10-17-base
Improved
  • Higher accuracy on scientific text and text from web pages
  • Improved robustness to bypassers

Model 3.3m

· modelVersion2025-09-29-multilingual
Improved
  • Upgraded to larger multilingual model with increased cross-language understanding.
  • Higher AI recall and reduced FPR on French, Spanish, Portuguese, German.

Model 3.10b

· modelVersion2025-09-20-base
Improved
  • False positives have been decreased further across a variety of domains

Model 3.9b

· modelVersion2025-09-04-base
Improved
  • Improved accuracy across the board, reducing false positives on human documents, and increasing recall on AI documents

Model 3.8b

· modelVersion2025-08-22-base
Improved
  • Improved performance on GPT5 models

Model 3.7b

· modelVersion2025-08-11-base
Improved
  • Improved performance on the latest language models from top providers
  • Improved robustness to openai deep research data
  • Improved robustness to adversarial prompts

Model 3.6b

· modelVersion2025-07-23-base
Improved
  • Decreased false positive rate
  • More consistent results on documents with special characters
  • Higher robustness to AI bypassers

Model 3.5b

· modelVersion2025-06-02-base
Added
  • Ability to detect “polished” texts that are lightly edited by AI
Improved
  • Robustness to AI bypassers
  • Robustness to text formatting changes

Model 3.2m

· modelVersion2025-05-13-multilingual
Improved
  • Added training data for the supported languages

Model 3.4b

· modelVersion2025-05-02-base
Improved
  • Increased output stability when rescanning multiple times

Model 3.1m

· modelVersion2025-04-01-multilingual
Improved
  • Improved AI detection on French, Spanish, German, and Portuguese text

Model 3.3b

· modelVersion2025-03-13-base
Improved
  • Improved recall on texts that are paraphrased

Model 3.2b

· modelVersion2025-03-04-base
Improved
  • Improved AI detection robustness to text formatting changes

Model 3.1b

· modelVersion2025-01-09-base
Improved
  • Improved AI detection on o1 and paraphrased texts

Model Update

· modelVersion2024-12-17-base
Improved
  • Improved AI detection on Claude and Gemini models

Model Update

· modelVersion2024-11-11-base
Improved
  • Model upgrade to decrease false positives on a variety of document types

Model Update

· modelVersion2024-11-04-base
Added
  • Increased robustness to paraphrasing

Model Update

· modelVersion2024-10-19-base
Improved
  • Minor tweak to decrease false positive rate across various document types

Model Update

· modelVersion2024-08-02-base
Improved
  • Improved detection on AI generated essays
  • Increased robustness to variations in document formatting

Model Update

· modelVersion2024-07-12-base
Improved
  • Model upgrades to have increased recall on open source LLMs

Model Update

· modelVersion2024-04-10-multilingual
Added
  • Dedicated multilingual model for serving AI detection requests on French and Spanish texts with higher accuracy

Model Update

· modelVersionbase-2024-04-04
Improved
  • Model upgrades to have increased recall on on AI-generated essays relative to March 28, 2024 model

Model Update

· modelVersion2024-03-28
Improved
  • Model upgrades to improve performance on academic writing and have better tuned confidence scores

Model Update

· modelVersion2024-02-06
Improved
  • Processing of long documents has been made more consistent. There should be less variation in predictions when cutting out a few sentences from long AI-generated documents

Model Update

· modelVersion2024-01-09
Improved
  • Predictions have been tuned to better represent the probability that the model is correct. That means that when we provide an AI score of 90%, there is a 90% chance that the text is entirely written by an AI, and 10% chance that it is either human or mixed.

Model Update

Added
  • Added a “mixed” class to our output. Now, we provide a probability for “human”, “mixed”, and “ai”, which add up to 1. We label a document as “mixed” if we think there is a combination of significant human passages and AI passages in the text. This enable users to differentiate between two cases where there is “50% AI”. It could be that:
    1. We are unsure if the entire document is AI or human, and it feels somewhere in-between
    2. We are confident 50% is AI-like, and 50% is human-like
  • Added a confidence score from 0 (least confident) to 1 (most confident). This enable users to see the model’s confidence in the prediction

Model Update

Improved
  • Improved splitting of sentences for numbered lists such that list numbers and corresponding content are treated as a contiguous span of text

Model Update

Improved
  • Consolidated ensemble predictions to improve prediction speed
  • AI probabilities are tuned to be more representative of prediction uncertainty

Model Update

Added
  • Improved sentence-level highlighting to make it more consistent with the document level prediction
  • Faster inference for longer documents that have many sentences

Model Update

Added
  • Deployed an updated deep learning model that performs better on several of our internal benchmarks including documents written by ESL students, resulting in fewer false positives
  • Optimized model for increased speed without sacrificing any accuracy
  • Made upgrades to our sentence-highlighting model to align more closely with our overall document probabilities

Model Update

Added
  • Deployed an updated deep learning model trained on much more data, including millions of rows of human data from community-generated articles and news, in order to lower false positives
  • Updated our sentence-labeling model to incorporate more powerful deep learning architecture, increasing its quality and alignment with the document-labeling model

Model Update

Added
  • Deployed a much more powerful deep learning model, trained on additional data, and able to distinguish much more subtle signs of AI vs. human generated text

Model Update

Added
  • Updated our set of documents to our internet archive search to verify that known human documents are labeled as human
Improved
  • Tuned our deep learning model to improve its performance on ESL writing, and generally
  • Updated our “Entirely AI” threshold from 0.85 to 0.88 to calibrate to the new model

Model Update

Added
  • Updated our API and dashboard to return a subtitle to the results message for times when more context is needed to describe the result
  • Added a result “Our ensemble of detectors have predicted different results for this text.” for when our statistical model and our deep learning model have different probabilities for the text
Improved
  • No longer highlighting any sentences if the document is deeply entirely AI by both the statistical and deep learning models

Model Update

Added
  • Added over 40,000 outputs from each of GPT-4, ChatGPT, and GPT3.5 (for a total of 120,000k additional rows), and over 400,000 outputs from human-written content to our training dataset. This helped us improve our performance on the challenging task of identifying GPT-4 text
  • Added a new deep learning model that does more complex pattern recognition on top of our existing detectors, increasing our accuracy

Model Update

Added
Fixed
  • Reduced the number of false negatives (AI detected as human) by improving the way we ensemble the results from our deep learning and statistical approaches

Model Update

Added
  • Added a new deep learning model that looks for more subtle signals of AI-generated writing (paragraph syntax, key phrases and word selection, formatting), increasing our overall accuracy. We combine the results from our statistical approach with this deep learning model
Fixed
  • Reduced the number of false positives (humans detected as AI) using our new model, in addition to beta testing our internet archive search to find documents that are known to be human-written, and classifying them as human

Keep reading

How to Fact-Check ChatGPT in 5 Simple Steps

How to Fact-Check ChatGPT in 5 Simple Steps

In five simple steps, we show you how to verify citations and check independent evidence when using AI.

Adele Barlow

· 8 min read
Watercolour-style illustration of a laptop computer split into blue and yellow halves, representing comparison of AI detection tools
Product Reviews

10 Best AI Detectors with Highest Accuracy (Updated October 2026)

Compare the best AI detectors of 2026 with these breakdowns of their accuracy rates, unique features, and pricing.

Adele Barlow

· 16 min read
How to Write a Cover Letter That Lands the Interview

How to Write a Cover Letter That Lands the Interview

A cover letter is more than a formality - it’s your chance to show employers why you’re right for the job. Learn to write a compelling, tailored letter with a simple structure, practical tips, templates, and samples to help you stand out.

Allie Watson

· 5 min read