GPT-4
GPT-4 accepts image and text input and passes a simulated bar exam around the top 10% of test takers, with architecture and data undisclosed.
Multimodal Transformer whose capabilities OpenAI forecast from runs using 1/1,000th of the compute. The technical report withheld parameter count, data and training details, setting the closed-report norm. Image input reached users later via GPT-4V.
- Date
- Tuesday, 14 March 2023
- Lab
- OpenAI
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Simulated bar exam | around top 10% of test takers simulated exam; per report abstract | company |
| Context window (default / extended) | 8K / 32K tokens API variants at launch | company |
arXiv v1 is dated 2023-03-15; the announcement was 2023-03-14. Altman said training cost over $100M and a 1T-parameter figure circulated via Semafor (unofficial, per Wikipedia).
Sources
- arxiv.org/abs/2303.08774
- en.wikipedia.org/wiki/GPT-4
- hidekazu-konishi.com/entry/openai_gpt_model_release_timeline.html
This record was checked against its sources on 6 October 2026. How we check