The sources reviewed here do not establish an officially confirmed parameter count for GPT-4o. A January 2025 secondary article reports an estimate of about 200 billion parameters, attributing it to Figure 2 of a Microsoft paper. Treat that as an attributed, unverified estimate—not an OpenAI specification.
The evidence retrieved for this article does not confirm that Figure 2 is from the MEDEC paper, or that an inspectable MEDEC table explicitly lists GPT-4o at 200 billion parameters. The MEDEC paper’s version 1 identifies its Microsoft and University of Washington authors, but the retrieved text does not show the alleged GPT-4o entry. (arxiv.org)
Where does the 200-billion estimate come from?
The January 2025 article says Figure 2 of a Microsoft paper gives GPT-4o at approximately 200 billion parameters, and its author expresses skepticism about the figures. However, the captured article text does not expose the figure or provide a usable destination for its “[Source]” reference. The connection between that figure and MEDEC therefore remains unverified in the material reviewed.
The MEDEC paper’s indexed version 1 text says that most model parameter counts were estimates included for context, rather than publicly disclosed specifications. A later clarification reported in the supplied research says those estimates were mined from public articles and the authors could not vouch for their accuracy. Neither passage, as retrieved here, establishes a GPT-4o-specific table entry or confirms the approximately 200-billion figure. (arxiv.org)
The careful takeaway is: a secondary article reports GPT-4o at around 200 billion parameters, but the primary-source table and the link between that table and the secondary article’s figure have not been verified here. Do not describe the number as officially confirmed or as a measurement by Microsoft researchers.
Don’t confuse GPT-4 with GPT-4o
Parameter estimates for GPT-4 concern a different model; they do not establish GPT-4o’s size. Keep each estimate attached to the specific model it names.
Should the estimate guide production API selection?
No. An unverified parameter estimate is not enough to determine whether a model fits your application. Check the official API documentation for the exact model and the capabilities your workload requires, then assess those against your own task.
OpenAI’s GPT-4o API page currently labels GPT-4o as deprecated and lists text and image inputs, text output, and a 128,000-token context window. For a new production integration, check the current model status and documentation before committing to it. (developers.openai.com)
Bottom line: about 200 billion is an estimate reported by a secondary source, not a confirmed GPT-4o count in the evidence reviewed here. Choose an API model based on verified, task-relevant product information—not an inferred parameter total.
Sources
The Number of Parameters of GPT-4o and Claude 3.5 Sonnet
MEDEC: A Benchmark for Medical Error Detection and Correction in Clinical Notes