Short answer
There is no confirmed GPT-4 parameter count in OpenAI’s technical report. OpenAI says the report omits details about GPT-4’s architecture, including model size. So figures such as “1 trillion” or “1.8 trillion” should be treated as unverified estimates—not established specifications. (arxiv.org)
Why do different numbers appear online?
The estimates conflict, and the sources do not establish which figure is correct:
- Close to 1 trillion: In an article dated February 15, 2023, Cobus Greyling predicted GPT-4 would be near 1 trillion parameters. That was a forecast, not a confirmed count. Read the article.
- Around 1.8 trillion: A CometAPI article gives this estimate but explicitly labels it unconfirmed by OpenAI. The captured passage does not explain how it was calculated or whether it refers to total or active parameters. Read the article.
These are separate, unverified claims—not a reliable range. Don’t average them or repeat either as GPT-4’s known size. The primary-source answer is that OpenAI’s technical report does not provide the model-size details needed to confirm a count. (arxiv.org)
Some pages also give precise but unsupported figures. For example, a LinkedIn article claims GPT-4 has 175 billion parameters, while also assigning that same count to GPT-3 and describing GPT-4 as larger. That inconsistency makes it unsuitable evidence for GPT-4’s count. Read the article.
Parameters are not API token limits
Here, parameters means the model’s learned weights. That is different from API request settings—sometimes also called parameters—and from token limits such as context length or maximum output. OpenAI’s GPT-4 API documentation lists operational model details, but those are not a disclosure of the model’s learned-weight count. See the GPT-4 API documentation.
Does the count help you choose a production API model?
Not by itself. An undisclosed or speculative weight count does not tell you whether a model is currently available or fits your workload. Check current official model documentation for availability and the limits relevant to your application; evaluate candidate models on your own tasks rather than choosing by an unverified parameter estimate.
Sources
What Are Realistic GPT-4 Size Expectations?
How Many Parameters Does GPT-5 Have? Here's What We Actually Found