ChatGPT Versus DeepSeek: Assessing Artificial Intelligence Performance on Radiation Oncology Examination Questions
Bibliographic record
Abstract
Purpose: Large language models have been assessed for their ability to receive and answer medical questions. Recently, there has been a new large language model named DeepSeek released, which has not been assessed for medical accuracy. This is the first study to assess DeepSeek for accuracy in responding to medical questions. Methods and Materials: We prompted DeepSeek-R1 and several models of ChatGPT with 600 radiation oncology examination questions from national radiation oncology in-service multiple-choice examinations. These questions are used by medical residents in preparation for their certifying board examination and assess knowledge on anatomy, treatment planning, cancer epidemiology, and landmark trials. We recorded each model's accuracy, total prompt and completion tokens used, and total run time. Accuracy was compared across question categories and between models. Type I error was set at 0.05. Results: = .012) and was least accurate for questions about landmark studies (74.2% accuracy). ChatGPT o1 answered 89.0% of questions correctly, requiring 10 seconds per question. ChatGPT o1's accuracy did not significantly differ across question categories (93.5% accurate on questions about landmark studies). DeepSeek-R1 used 7.2% more tokens than ChatGPT o1. At February 2025 prices, DeepSeek-R1 costs up to $1.56, compared with ChatGPT's $37.96. Conclusion: DeepSeek-R1 is less accurate and answers more slowly compared with ChatGPT o1, but is less costly at the time of this manuscript preparation. Careful analysis and consideration of the current landscape and performance of each model is needed before implementation of DeepSeek-R1 or ChatGPT o1 to determine if the added financial costs of ChatGPT o1 are within the intended goals of improved accuracy and efficiency.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".