February 27, 2025

The model is currently only available as a research preview through an OpenAI API account (Chat Completions API, Assistants API, and Batch API) or as a chatbot with a ChatGPT Pro Users account on web, mobile and desktop.

Only text+image inputs and text output is supported at this time.

The model accepts text and image inputs and outputs text responses, but doesn't specify technical details like maximum context window length, token limits, or specific formatting requirements.

Proprietary

The system card lacks comprehensive usage instructions. It doesn't provide specific examples, recommendations, hardware/software dependencies, or detailed interaction guidelines. It mentions that GPT-4.5 follows an "Instruction Hierarchy" that prioritizes system messages over user messages, but doesn't offer practical guidance for effective model usage.

The OpenAI website has easily-accessible help articles that guide hands-on use. The system card covers performance and safety evaluations in detail.

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

The OpenAI Usage policies can be found [here](https://openai.com/policies/usage-policies/).

User Data is used to train GPT models, unless users explicitly opt-out. (https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance)

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. Copyright-related disputes can be submitted here: https://openai.com/form/copyright-disputes/.

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

OpenAI 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).

GPT-4.5 is a general-purpose text generation model that can be used for creative and nuanced tasks like writing and solving practical problems. It has an extensive knowledge base, improved ability to follow user intent, and greater “EQ” that previous generations. However, the model lacks chain-of-thought reasoning abilities and may be slower due to its size. It, also, performs significantly worse than reasoning models (i.e. o1) on complex coding tasks or tasks require detailed logic or multi-step reasoning.

The exact architecture details like layer counts or attention mechanisms are not available. This model prioritizes scaling over chain of thought reasoning. This results in a pre-trained model that claims "broader knowledge and deeper a understanding of the world, leading to reduced hallucinations and more reliability". Almost no details are available at this time.

This model uses the typical approach for state of the art large foundation models: pre-training from large-scale unsupervised learning on a large corpora and post-training for fine tuning with supervision (SFT), and reinforcement learning (RLHF). Benchmark performance metrics are shared, but no detailed information about the training methodology is provided.

They claim that this model improves on GPT-4's computational efficiency by "more than 10x" without any further details.

No specific information about carbon emissions or total energy consumption associated with training the model

No information is provided about the total size of the training dataset.

The system card says the model was trained using publicly available data, proprietary data from data partnerships and custom datasets developed at OpenAI. No details about dataset composition or origin are provided.

The system card says the model was trained using publicly available data, proprietary data from data partnerships and custom datasets developed at OpenAI. No details about dataset composition or origin are provided.

No information is provided.

The data processing pipeline claims rigorous filtering to maintain data quality and mitigate potential risks as well as advanced data filtering processes to reduce processing of personal information. Beyond these considerations, no details are available.

A moderation API and safety classifiers are used to prevent the use of harmful or sensitive content, including explicit content involving minors, but doesn't detail specific handling of toxic or hateful language in the training data. No details are provided beyond mention that this was addressed.

GPT-4.5 was evaluated on knowledge-related and coding benchmarks. It showed improved performance over GPT-4o and o1 on several benchmarks, like MMMLU (multilingual) and Simple-QA; however, it is significantly worse than o1 at coding (SWE-Bench Verified) and math (AIME ‘24 (math)) assessments.
On human preference testing, GPT-4.5 outperformed GPT-4o by 7 to 13% depending on task type. Additional qualitative examples suggest that GPT-4.5 may have a better understanding of subtle cues or implicit expectations in prompts, potentially, making it a more useful tool for human collaboration.

GPT-4.5 was evaluated for hallucinations, biases and safety.

Bias: GPT-4.5 showed low bias on BBQ, a benchmark that assess whether known social biases override the ability for the model to produce the correct answer.

Hallucinations: Metrics for the SimpleQA, PersonQA, and some domain-specific benchmarks which indicate this model is a substantial improvement over previous OpenAI models, though the system card does acknowledge that hallucinations persist and remain a significant limitation. The remainder of the discussion of limitations is primarily reporting benchmark performance, noting there are some improvements, but adding no further discussion about model limitations.

Safety was evaluated by measuring correct refusals to unsafe requests (using existing benchmarks like WIldChat and custom multi-modal benchmarks) and robustness against jailbreaks (using a custom dataset of human-sourced jailbreak prompts).

Multiple adversarial tests were performed, including jailbreak evaluations against human-sourced jailbreaks, red teaming, third party assessment from Apollo Research, and generic stress-testing according to OpenAI's Preparedness Framework. The Safety Advisory Group classified GPT-4.5 as overall medium risk, including medium risk for CBRN and persuasion and low for cybersecurity and model autonomy.

Some safety mitigations are discussed in general terms. The mitigations include SFT and RHLF during post-training to align with human preferences, safety training for political persuasion tasks, monitoring and detection systems, and enhanced content moderation. In addition, the model was taught to adhere to an [Instruction Hierarchy](https://arxiv.org/abs/2404.13208) that explicitly defined how the model should behave given conflicting instructions. These are claims made in the system card with no further information for public assessment.