What are the limitations of GPT-5 for real-world multimodal medical diagnosis?

A 161-case test built for open-ended diagnosis

The study asks whether adding reasoning depth, web retrieval, and multi-step agentic control makes a generalist model a more reliable diagnostic aid. To test that, the authors built a benchmark from MedXpertQA, keeping only reasoning-style questions about diagnosis, examination, or treatment, then having medical experts retain 395 cases with concrete, actionable clinical decisions. They further restricted the set to diagnosis cases and required each body system to make up at least 10% of the final benchmark, leaving 161 cases: cardiovascular (40), digestive (34), respiratory (34), musculoskeletal (31), and nervous (22).

Each case pairs an open-ended diagnostic question with a patient narrative, at least one medical image (CT, MRI, X-ray, PET, pathology, EEG/ECG, or clinical charts), and a clinically validated ground truth. Ten GPT-5 configurations each answered all 161 cases independently, giving 1610 responses, all collected through the ChatGPT web interface between August 9 and 20, 2025 with identical zero-shot prompts.

Accuracy here is not string matching against the original label. Three physicians independently judged each response on semantic equivalence, consistency with clinical evidence, diagnostic specificity, anatomical correctness, and relevance, and could credit a model answer that fit the evidence better than the benchmark label. Inter-rater reliability was moderate (Gwet's AC1 0.56–0.64, mean 0.60) with high observed agreement (0.76–0.80, mean 0.79), and a majority vote formed the consensus label. So the percentages below are physician-adjudicated accuracy, not exact-match scores.

Reasoning depth raised accuracy more than tools did

The paper separates three kinds of added complexity: model complexity (reasoning strength), tool complexity (adding web search), and structural complexity (moving from standard chat to agentic or deep-research workflows). Only the first showed a clear, monotonic gain. Accuracy rose from Thinking Mini at 32.9% to Auto at 38.5%, Thinking at 41.6%, and Pro at 46.0%.

Tool complexity was mixed. Web search sharply reduced accuracy in Auto mode, from 38.5% to 26.7%, while producing only small gains for Thinking Mini, Thinking, and Pro. Structural complexity gave no clear advantage either: ChatGPT Agent reached 36.6% and Deep Research 29.8%, both below Auto without web search.

Performance also varied by body system. Neurologic cases consistently achieved the highest accuracy across configurations, while digestive, respiratory, and cardiovascular cases were generally harder. In most body systems accuracy improved as reasoning capability increased, and repeated-run robustness analyses on a stratified 20% sample showed broadly consistent patterns across models and body systems.

Web search mostly reinforced the model's first guess

Retrieval did not fail because the models found bad sources. Most citations came from professional medical and educational sites, and the stronger models were especially likely to pull from specialized medical domains or academic resources. Auto was the exception: over half of its cited sources were non-specialist sites such as Wikipedia.

The problem was how retrieved material was used. Across configurations, most cases either stayed incorrect or stayed correct after search. The share rescued by search was small, from 6.2% in Auto to 11.2% in Thinking, while some initially correct cases got worse, especially in Auto at 18.0%. The authors describe this as confirmation bias: retrieval typically reinforced the model's initial guess rather than prompting reflection or correction. For the Thinking model, retrieved content generally followed the base trajectory, strengthening both correct and incorrect predictions, while Auto often retrieved low-relevance or distracting material.

Higher-capacity models also tended to prioritize visually similar images over lesion-specific cues, so even domain-relevant retrieval was poorly converted into a corrected diagnosis. This is why the paper treats retrieval as a mechanism that needs redesign for reflection and error correction, not as a lever that reliably adds accuracy.

Agentic modes underperformed and hid their reasoning

Agent and Deep Research are designed to interleave reasoning with action planning, tool invocation, and iterative refinement. Despite that design, both underperformed the standard models on this benchmark. The transparency cost was as notable as the accuracy cost: Agent mode returned only a final diagnosis without visible intermediate reasoning, making the output hard for a clinician to audit.

Deep Research produced more output detail but still showed limited traceability. In 29 of 161 cases it provided no citations at all, and in the remaining cases it leaned less on professional sources and more on community or case-sharing sites, a pattern the authors connect to its lower accuracy.

As a reference point rather than a substitute for the within-ecosystem comparison, the authors also analyzed open medical agent frameworks with directly measurable workflow traces. Those frameworks showed substantially greater workflow complexity, including longer reasoning chains, more retrieval operations, and more agentic interactions, without consistent accuracy gains; in several settings more complexity came with reduced reasoning stability and more diffuse evidence integration.

Errors cluster in evidence use, not medical knowledge

The error analysis points to evidence integration as the main bottleneck. Physicians validated 928 error-type annotations drawn from 51 sampled cases and confirmed 827 (89.1%), supporting the precision of the model-generated error labels. That validation assessed precision only, not recall or completeness, so the reported patterns describe errors the framework identified rather than an exhaustive list of every reasoning error.

Across nearly all configurations, the largest error category was Evidence Errors, meaning failures to interpret or use case-specific evidence, accounting for roughly 27–48% of errors. Within that category, Misinterpretation was the most frequent subtype, appearing in roughly 45–65% of cases across most configurations, which the authors read as frequent misreading of clinical evidence, especially visual findings. Evidence-related errors were more strongly associated with body system and input evidence quality than knowledge-based errors were.

The models were not without strengths. They consistently extracted key patient history to guide coherent reasoning and produced structured differential diagnoses even when the final answer was wrong, and they increasingly linked history with imaging to give more specific labels and clearer explanations. The failure mode was a disconnect between what the model saw and what it concluded, not an absence of clinical facts.

Input quality set a hard ceiling on accuracy

Thirty-seven cases were classified as low-quality inputs because of insufficient or ambiguous questions, low-resolution images, patient motion, electrical noise, equipment artifacts, or other interference that obscured the relevant clinical signal; the other 125 were high-quality. Every configuration performed substantially worse on the low-quality group, with accuracy dropping by roughly 10–30 percentage points.

The gap was largest for the strongest configurations. Pro fell from 52.4% on high-quality cases to 24.3% on low-quality ones, Thinking from 47.6% to 21.6%, and Auto from 44.4% to 18.9%. ChatGPT Agent dropped from 43.5% to 13.5%. Thinking Mini was the exception in statistical terms, moving from 34.7% to 27.0% without a significant difference, and Thinking Mini with web search likewise showed no significant gap (37.9% to 24.3%).

This matters for interpreting the headline numbers: the 46.0% Pro result is an average over a case mix that includes degraded inputs, and the paper's own recommendation is to set practical standards for image resolution, noise, and question clarity rather than to assume the model will compensate.

Where these results do and do not transfer

Several boundaries qualify the comparison. The evaluated models were proprietary and lack architectural transparency, and all testing ran through a live interface, so background model updates may have introduced variability. The benchmark is 161 diagnosis-only cases drawn from MedXpertQA and filtered for reasoning-style questions and body-system balance, so it does not represent treatment planning, triage, or the full distribution of clinical work.

The authors are also explicit that the agentic finding is not an argument against agentic medical AI broadly. Their claim is narrower: within the evaluated generalist GPT-5 ecosystem, retrieval and agentic workflow complexity did not reliably improve open-ended multimodal diagnostic performance and often made reasoning less auditable. Future agentic systems, in their framing, would need evidence-grounded retrieval, transparent workflow control, and clinically auditable intermediate reasoning before added complexity could be expected to help.

Prioritize auditable reasoning over added tool layers

The paper's practical ordering follows directly from its comparisons. Reasoning strength was the one dimension that improved accuracy monotonically, from 32.9% at Thinking Mini to 46.0% at Pro, while web search and agentic structure failed to beat plain Auto. For a reader deciding what to look for in a medical AI system, that makes exposed, stepwise reasoning the first thing to check, ahead of retrieval breadth or the number of tools an agent can call.

The second priority is visual grounding. Misinterpretation appeared in roughly 45–65% of cases across most configurations, and the strongest models still sometimes favored visually similar images over lesion-specific cues. A system that cannot tie its textual conclusion back to the specific imaging finding is, on this evidence, the wrong place to add complexity.

The third is input discipline. Because accuracy fell by roughly 10–30 percentage points on the 37 low-quality cases, and because the best configuration still topped out at 46.0% overall, the realistic near-term role is decision support on well-captured cases with a clinician auditing the reasoning trace, not autonomous diagnosis on imperfect real-world data.

Sources used in this answer