Open-source GABRIEL toolkit

Qualitative research starts with messy material — interview transcripts, speeches, photos, field notes — and quantitative research starts with numbers. OpenAI has released GABRIEL to turn the first kind into the second. (openai.com) OpenAI described GABRIEL as an open-source toolkit that uses its GPT application programming interface to measure attributes in text, images, or audio. The code is published on GitHub under the OpenAI account, and the package is also available on the Python Package Index as `openai-gabriel`. (github.com) (pypi.org) The toolkit ships with a Google Colab tutorial notebook and a local Jupyter notebook file for researchers who want a quick start. OpenAI’s repository says users can install it with `pip install openai-gabriel` and then import `gabriel` in Python. (github.com) (pypi.org) At a basic level, the software asks a model to score or sort qualitative material the way a research assistant might code a stack of documents. OpenAI’s paper gives examples such as rating how “pro innovation” a speech is, classifying content, and extracting structured information from unstructured sources. (cdn.openai.com) The repository exposes that idea as a menu of tasks, including classify, rate, rank, extract, deduplicate, deidentify, merge, and compare. OpenAI says the target users are social scientists and data scientists working with large corpora that are hard to code by hand. (github.com 1) (github.com 2) OpenAI tied the release to a broader pitch for “scaling social science research.” In its write-up, the company said GABRIEL is meant to help researchers analyze large volumes of qualitative material while keeping the richness of text, images, and human judgment in the workflow. (openai.com) The technical case for the toolkit rests on a new evaluation paper. OpenAI and Harvard University researchers wrote that they tested GPT as a measurement tool on more than 1,000 human-annotated tasks and found performance that was generally indistinguishable from human annotators across domains. (cdn.openai.com) The project is already being updated in public. GitHub’s release log shows version 1.0.8 on February 24, 2026, with bug fixes for very large runs, including timeout initialization from checkpoint saves and more robust handling of large comma-separated value files. (github.com) OpenAI’s repository description expands the use case beyond text and images to audio, which suggests the toolkit is aimed at multimodal research pipelines rather than a single survey-coding workflow. The company’s examples include rating rhetoric across a million speeches and matching product catalogs at scale. (github.com) (pypi.org) The release leaves one practical boundary in place: GABRIEL is a toolkit around the GPT application programming interface, not a standalone model. Researchers can inspect the code, run the notebook, and adapt the pipeline, but the measurements still depend on OpenAI’s underlying models and application programming interface access. (github.com) (openai.com)