GPT-4 adds multimodal, human-level exam performance
OpenAI's model reads images as well as text and scores in the top 10 percent on a bar exam
Quick facts
- Institution
- OpenAI
- Capability
- Multimodal (text and image input)
- Benchmark
- Top 10% on simulated bar exam
What happened
OpenAI released GPT-4, described in its technical report as a large-scale, multimodal Transformer that accepts both image and text inputs and produces text outputs, a departure from earlier GPT models that handled text only. OpenAI reported that GPT-4 achieved human-level performance on a range of professional and academic benchmarks originally designed for people, including a simulated bar exam on which it scored around the top 10 percent of test takers, compared with GPT-3.5's bottom-10-percent score on the same test. A significant part of the underlying engineering effort, according to the report, went into building training infrastructure that behaved predictably across scales, letting OpenAI forecast aspects of GPT-4's eventual performance using much smaller models trained with a tiny fraction of GPT-4's total compute. OpenAI did not disclose GPT-4's parameter count, architecture details, or training data in the technical report, a departure from the fuller disclosures of earlier GPT papers.
Why it matters
GPT-4's jump in exam performance and its addition of image understanding expanded what a single general-purpose model could plausibly be used for, from tutoring to document analysis, but OpenAI's decision to withhold architectural details also marked a shift toward less open publication norms in frontier AI research, a change some researchers criticized.
How we know
The GPT-4 technical report describes the benchmark scores and multimodal capability directly; OpenAI's own announcement page corroborates the same claims about bar-exam performance and multimodal input.
Sources
- OpenAI. GPT-4 Technical Report · Primary source (author-declared)arxiv.org · Cited as a "primary" source (no stronger domain match). · Link is live and its text matches the event's key terms (Jul 2026)
- OpenAI. GPT-4 · Primary source (author-declared)openai.com · Cited as a "primary" source (no stronger domain match). · Link is live and its text matches the event's key terms (Jul 2026)
See something wrong? . Corrections with a source get fixed fastest.
Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →