Currently, all OpenAI API paying users (with a successful payment history) can access GPT-4 with 8K context memory. And the company plans to open API access to new developers at the end of this month, and then adjust the available limits through computing power availability.
OpenAI announced on Thursday (July 6) that its latest large-scale language model, GPT-4, will be officially open for use through its API.
Currently, all OpenAI API paying users (with a successful payment history) can access GPT-4 with 8K context memory. And the company plans to open API access to new developers at the end of this month, and then adjust the available limits through computing power availability.
This means that developers around the world can use the GPT-4 large language model to enhance their applications or develop new generative AI applications.
"Since March, millions of developers have requested access to the GPT-4 API, and the range of innovative products leveraging GPT-4 grows every day," OpenAI wrote in a blog post Thursday.
In March of this year, OpenAI officially released GPT-4, a model that can accept image and text input to generate text (including code) and reach "human level" on various professional and academic benchmarks. This is an improvement on its predecessor GPT-3.5, which only accepts text input. Like previous GPT models, GPT-4 is trained using publicly available data, including data from public web pages and data licensed by OpenAI.
GPT-4 retains its previous crown in terms of context windows, with a default of 8k tokens (tokens are the basic unit of text processing for models) and an upper limit of 32k tokens. Generally speaking, models with larger context windows are more likely to remember the content of recent conversations, while models with smaller windows tend to forget the previous content during the conversation, causing them to stray from the topic.
01 Support any use case
The blog post also states, “We envision a future where chat-based models can support any use case.”
OpenAI also plans to release API interfaces for GPT-3.5 Turbo, DALL·E, and Whisper.
GPT-3.5 Turbo is another one of OpenAI’s text generation models, but it’s less powerful than GPT-4; DALL-E 2 is OpenAI’s image generation model; and Whisper is the company’s speech-to-text model.
The image understanding feature is not yet available to all OpenAI customers. OpenAI is reportedly testing it with a partner called "Be My Eyes," but it has not yet indicated when it will be available to a wider customer base.
OpenAI also said that in the future, it will allow developers to fine-tune GPT-4 and GPT-3.5 Turbo with their own data, a feature that may be available later this year.
02 Elimination and replacement
It’s worth noting that even one of the best generative AI models today, GPT-4, is not perfect. It misperceives facts, sometimes makes “confident” reasoning errors, and fails on some complex problems.
In response, OpenAI also announced that it will retire several older models in the Completions API, and future models will focus on the Chat Completions API as part of the company's efforts to optimize computing power.
The following figure shows the old models that OpenAI announced to be eliminated, and the new models that replaced them:
OpenAI will retire these old models of the Completions API within 6 months. While the API is still accessible, starting today, the company will mark it as "legacy" in the developer documentation. Starting January 4, 2024, the old Completions API will no longer be available.