OpenAI has made GPT-4 generally available through its API, widening access to a model that developers had been requesting since March. The announcement also sets out plans for broader access, future fine-tuning, and a transition away from certain older models.
API access is expanding in stages
Existing OpenAI API developers with a history of successful payments can access GPT-4. OpenAI said it planned to open access to new developers by the end of the month, then increase availability limits depending on compute availability.
The staged rollout reflects the difference between making a model available and offering unlimited access. Developers may be able to build with GPT-4 now, while the amount of API capacity they can use may rise later as compute allows.
OpenAI said millions of developers had requested access since March and described a growing range of products using GPT-4. The company said it sees chat-based models supporting any use case, though the announcement did not specify a timetable for access beyond the initial rollout.
Text and image inputs broaden what the model can handle
GPT-4 can generate text, including code, and accept both text and image inputs. Its predecessor, GPT-3.5, accepted only text, according to the article. Image understanding, however, is not yet available to all OpenAI customers.
OpenAI is initially testing that capability with a single partner, Be My Eyes. The company has not indicated when image understanding will become available to a wider customer base, so developers should distinguish GPT-4's described capabilities from the features they can access today.
The model was trained on publicly available data, including public web pages, as well as data licensed by OpenAI. The article also reports that GPT-4 performs at “human level” on various professional and academic benchmarks, but that does not make its outputs reliable in every situation.
Model limits still matter in applications
GPT-4 can hallucinate facts and make reasoning errors, sometimes with confidence. It also does not learn from its experience, and the source notes that it can introduce security vulnerabilities into generated code when faced with hard problems.
For developers, those limits mean model output still needs review, especially in code and other contexts where mistakes have consequences. General availability creates an opportunity to integrate GPT-4, but it does not remove the need to check what the model produces.
OpenAI said it plans to let developers fine-tune GPT-4 and GPT-3.5 Turbo with their own data later this year. Fine-tuning can adapt a model using a developer's data; the article does not provide further details about how the planned capability will work.
Older API models have a transition deadline
In a related announcement, OpenAI made the DALL-E 2 image-generation API and Whisper speech-to-text API generally available. It also said it plans to retire older API models to optimize compute capacity.
Starting January 4, 2024, certain older models, specifically GPT-3 and its derivatives, will no longer be available. OpenAI said these will be replaced by new “base GPT-3” models, which the article says are presumed to be more compute efficient.
- Developers using the older models must manually upgrade their integrations by January 4.
- Those who want to keep using fine-tuned older models after that date will need to fine-tune replacements based on the new base GPT-3 models.
- OpenAI said it would support users through the transition and contact developers who had recently used the older models.
The change makes model migration part of the API announcement. Developers relying on GPT-3 derivatives need to plan for updates, while OpenAI says more information will follow when the new completion models are ready for early testing.
Competition is also changing the context window landscape
The article places GPT-4's availability against intensifying competition in generative AI. Anthropic had recently expanded the context window for Claude, its flagship text-generating model in preview, from 9,000 tokens to 100,000 tokens. GPT-4 had previously led on context window size, with up to 32,000 tokens.
A context window is the text a model considers before producing more text. Smaller windows can cause a model to forget even recent conversation content and drift off topic. The figures help explain one area of competition, but do not by themselves describe a model's overall usefulness.
For API developers, the announcement combines new capabilities with practical constraints: access is rolling out in stages, image understanding remains limited, model output can be wrong, and some older integrations have a fixed migration date. The ability to build with GPT-4 therefore comes with decisions about review, feature availability, and model updates.