GPT-4 by OpenAI

GPT-4¹ is the most factual and steerable Multimodal Model available via Poe, Bing and ChatGPT+ and OpenAI API.

Introduction

OpenAI released today GPT-4, their next generation AI system accepting both image & text inputs. It was trained on Microsoft Azure² as a large scale multimodal model.

OpenAI offers no details related to model architecture, size, hardware or datasets — justifying it with both with competitive market & safety reasons².

The main differences from the GPT-3.5 are:

• GPT-4 is multimodal with both text and images.
• GPT-4 is more reliable, creative and handles better nuances.
• GPT-4 context length is 25 000 words.

It offers State-of-the-Art (SOTA) performance on Natural Language Understanding (NLU) tasks, including MMLU with 86.4%. This result is 15.7% better than PaLM model (5-shot).

GPT-4 Pricing

Is GPT-4 available to users?

OpenAI offers access to the text version through paid ChatGPT+. Microsoft confirmed Bing uses GPT-4 and Poe-app already includes GPT-4.

The API will be available starting today through waitlist. It may be quicker to pass this waitlist by providing support to OpenAI Evals⁴ — used to share and crowdsource benchmarks.

GPT-4 be available in two context lengths: 8,192 tokens and 32,768 tokens:

GPT-4 price is for the 8k-version (“gpt-4”):

• 0.03$ / 1k tokens for input texts.
• 0.06$ / 1k tokens for output text.

GPT-4 price is for the 32k-version (“gpt-4–32k” or “gpt-4–32k-0314” supported until June 2023):

• 0.06$ / 1k tokens for input text.
• 0.12$ / 1k tokens for output text.