The OCR-based text input uses far fewer tokens than image-based multimodal input for the same slide, making it more cost-effective¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: tool and model accuracy changes quickly

Subclaims¶

q2 i? For the example slide in Figure 3(a), OCR-based text input consumed 392 tokens for 1124 characters, whereas image-based input of a 1500× 844 px image consumed 1105 tokens. → Patel 2025

Evidence¶

Patel 2025¶

Patel, K., Lin, Y.-Z., Raul, G., Shih, B. P.-J., Redondo, M. W., Latibari, B. S., Pacheco, J., Salehi, S., & Satam, P. (2025). Integrating Generative AI into Cybersecurity Education: A Study of OCR and Multimodal LLM-assisted Instruction. arXiv. https://arxiv.org/abs/2509.02998 doi:10.48550/arxiv.2509.02998

q2 · i? · design · r2

Token-cost analysis of one example slide (Figure 3(a)) comparing OCR text input (392 tokens for 1124 characters) against image input (1105 tokens for a 1500× 844 px image). The article concludes the OCR method "is more cost-effective"; no aggregate cost statistics are printed.

"In terms of cost, the OCR-based text input method in this example uses only 392 tokens for 1124 characters, but if image-based input is used to multimodal LLM, an image file is 1500× 844 px, and the number of tokens used is 1105 tokens"

Discussion¶

Related Claims¶

• The text-based LLM compensates for noisy OCR input, generating contextually accurate explanations even when OCR produces jumbled text from UI screenshots — related
• Multimodal LLMs outperform the OCR-based pipeline on visually dense slides, while the OCR pipeline produces comparable instruction on text-centric slides — related
• Students rate LLM-generated simplified instructions positively, averaging 7.83 on a 1-10 usefulness scale across 42 ratings in a live cybersecurity course — related
• Multimodal input consistently improves accuracy but not F1, model size correlates moderately with performance, and reasoning models show mixed results — reports the opposite
• Purpose-trained single-pass generation completes a slide in a median of 17 seconds and an interactive page in 59 seconds over 220k production requests, and delivers artifacts at 15–22× lower per-artifact API cost than flagship models — related