16 April 2025

The models can be accessed through the ChatGPT API

The models can take text, files, code, and images as input. They are able to access tools within the ChatGPT catalogue (web browsing, Python, image analysis and generation etc) and produce text.

The context window for this model is 200,000 tokens and it can output a maximum of 100,000 tokens.

Proprietary

Instructions for using this model can be found both on the OpenAI website and on the ChatGPT page.

The documentation is clear and easy to read. Most key information is available but too general. Some information is missing entirely, such as details on the system architecture or training processes.

You can find the changelog [here](https://platform.openai.com/docs/changelog); however, it may not contain all the details related to minor changes.

Usage Policies are available on Open AI's website.

User data is used to train ChatGPT models. This includes log data (e.g. your IP address), usage data, device information, location information, and cookies.

You can find how to opt out of model training and remove your data [here](https://help.openai.com/en/articles/7730893-data-controls-faq)

OpenAI describe their principles in their [OpenAI Charter](https://openai.com/charter/).

ChatGPT has a reporting feature that you can use to give feedback and report incidents. You can find more information [here](https://chatgpt.com/g/g-Jjm1uZYHz-incident-reporting).

The models are capable of responding to text, image, code, and file input. They have access to the full range of ChatGPT tools, including searching the internet, image analysis and generation, and Python.

Not explicitly stated in the provided documents.

OpenAI reasoning models are trained using reinforcement learning on chains of thought to encourage reasoning.

The two models were trained on diverse datasets. This included information that is available publicly online, information from third parties, and information from users, human trainers and researchers. Data is pre-processed to maintain quality and mitigate potential risks.

The data used to train these models was sourced from third parties, publicly available information on the internet, and from ChatGPT users, researchers, and human trainers.

Human labour is used in the production of this data. This includes data produced and generated by researchers and human trainers.

Data was filtered to maintain quality and mitigate a series of identified risks. Personal information was reduced from the training data. Moderation API and safety classifiers were also used to help prevent the use of harmful or sensitive content, including explicit material.

Personal information is removed from the dataset through pre-training filtering.

The models are tested against a variety of safety and performance evaluations. These include in-house evaluations, third-party evaluations by groups including Apollo Research and METR, PersonQA, PaperBench, SWE-Lancer, and MMLU. The results of each evaluation are listed in the system card and compared to other OpenAI models. Many of these benchmarks are reported with clear explanations for how and why the evaluation was conducted, but this is not the case for all of the evaluations.

The models can hallucinate, be jailbroken (i.e. prompted to produce inappropriate content) and produce incorrect refusals. Hallucinations are a significant concern with the models hallucinating 33% and 48% percent of the time on PersonQA (a benchmark that asks questions about public figures). Detailed results related to these limitations are reported in the System Card.

The model is tested extensively for safety risks. This includes through jailbreak testing to evaluate the robustness of the model using adversarial prompts. These jailbreaks are either human sourced or they are from the StrongReject database. Other safety risks are evaluated, including harmful image generation, production of disallowed content, hallucinations, and bias, among others. The measures taken to prevent each risk, the evaluations used to test them, and the results of each evaluation are included in the system card.

The model mitigations included post-training to teach the model about refusal behavior for harmful requests and using moderation models for the most egregious content. The final models are tested for a variety of safety risks including fairness and bias, personal identification, and deception by both OpenAI and third parties.