Series: The AI Giants Enter the CardiacSurgery ICU — “I Big entrano in corsia”
Three frontier AI models, twelve fictional clinical scenarios, seventy-two answers. Scoring them turned out to be harder than I thought.
I expected differences. Some diagnostic shortcuts, a few questionable therapeutic choices, maybe one spectacular mistake (the kind that makes for a good opening paragraph). What I got, after reading through the answers to a set of cardiac-surgical intensive care problems, was a different sort of difficulty: all three models had produced clinical responses that were convincing, at times uncomfortably so.
Convincing is not the same as safe, of course, and nothing here shows that these systems are safe clinicians. It does raise a question I find more interesting than the ranking itself. What should clinical AI evaluation measure once getting the broad answer right is no longer particularly difficult?
Why put frontier AI in the cardiac ICU?
The postoperative cardiac surgical patient is a hard test of reasoning. Things change fast. Lab values and monitor readings don’t always agree, and a decision that made sense an hour ago can turn dangerous once a new result comes back. A reassuring blood pressure may hide poor flow; a normal saturation can sit alongside a patient who is quietly retaining CO₂. And what looks like a minor postoperative complication can become an emergency within minutes, as anyone who has covered a cardiac ICU at night will recognize.
Getting the diagnosis right is only part of it. Good judgment at the bedside means setting priorities, knowing when action cannot wait, being honest about the limits of what is known, and changing course as the patient changes.
That, more than diagnostic accuracy, was what I wanted to explore.
Who are the three “giants”?
The three models come from three of the major AI laboratories. All are general-purpose systems; none is a dedicated, clinically validated medical device.
GPT-6 Astra (OpenAI). OpenAI presents Astra as its high-capability model for complex reasoning, research, coding, professional work, and demanding multistep tasks. With its large context capacity and configurable reasoning effort, it was a reasonable choice to represent OpenAI’s frontier systems.
Claude Fable 5.1 (Anthropic). Anthropic positions Fable 5.1 for demanding reasoning, long-horizon knowledge work, multistep research, and agentic tasks. Reasoning is adaptive and can be steered through an effort setting.
Gemini 4 Argon (Google). Google’s flagship Gemini 4 reasoning model came out on October 13, 2026, three days before I ran the experiment. It targets complex scientific, mathematical, and software problems, with support for long-context, multimodal work.
I include these descriptions to explain why the models were selected, not to suggest that they rank first, second, and third worldwide. They were simply three prominent frontier systems I could test at that point in time. The focus of the exercise was their general-purpose capability, not any specific training in cardiac surgery.
Official model documentation: OpenAI Astra · Anthropic Fable 5.1 · Google Gemini 4 Argon.
The challenge, in practical terms
On October 16, 2026, I sent 12 wholly synthetic clinical scenarios to the three models through their providers’ APIs. Half of them unfolded in three stages (T0, T1, T2), so each model had to update its assessment as the clinical picture changed. The other six asked for a single decision under uncertainty.
The evolving cases will be familiar to any cardiac surgical team: postoperative vasoplegia progressing to right ventricular failure; mediastinal drains that suddenly stop draining, followed by tamponade; hypoventilation after extubation; postoperative atrial fibrillation; acute kidney injury with venous congestion; and a focal neurological deficit emerging after an initially nonspecific postoperative change.
The uncertainty cases dealt with prognostic overconfidence, a heavily hemolyzed potassium specimen, early postoperative fever, anticoagulation in a patient with a mechanical valve who is still bleeding, readiness for extubation, and shock with too little hemodynamic information to go on.
Every answer had to address four points:
- The immediate clinical priority.
- The action to take, including urgency.
- The single most useful missing information.
- The uncertainty that remained.
In the evolving cases, each model was given its own earlier visible responses as context, never its competitors’. All three ran at their respective native high reasoning settings. A caveat is needed here: “high” does not correspond to the same amount of computation from one provider to the next, so the settings are not equivalent. No web browsing, retrieval, medical database, or external tool was available to the models at any point.
In total: 24 clinical stages, 72 API responses, and 36 model-by-scenario evaluations. One answer per model per scenario. There were no API errors, and no response appeared truncated.
The deliberately simple assessment
Before running anything, I had written five binary criteria for each case, along with a reference clinical response and the alternatives I would accept. A criterion earned one point if the model met the clinical requirement behind it. Evolving scenarios were judged on their entire three-stage trajectory rather than stage by stage.
That gives a maximum of 30 points for the six evolving scenarios and 30 for the six uncertainty scenarios, or 60 points per model.
A first, AI-assisted and non-blinded pass through the rubric gave these results:
| GPT-6 Astra | Claude Fable 5.1 | Gemini 4 Argon | |
|---|---|---|---|
| Six evolving cases | 30/30 | 29/30 | 29/30 |
| Six uncertainty cases | 29/30 | 30/30 | 29/30 |
| Total | 59/60 | 59/60 | 58/60 |
In eight of the twelve scenarios, all three models scored a full five out of five. The few points lost had to do with how explicitly a response qualified a prognostic statement, named its escalation triggers, or described a clinical contraindication. None reflected a gross error in the main decisions these cases required.
Please do not read these numbers as a statistically meaningful ranking. Twelve deliberately selected cases make neither an independently adjudicated benchmark nor a measure of general clinical accuracy. The scores are preliminary and remain open to specialist review.
What the scores conceal
The more I read, the more the differences mattered, and the less the leaderboard told me.
Take the case of acute right ventricular failure. None of the three models was fooled by a mean arterial pressure that had returned to normal; all understood that the circulation had not recovered. Where they parted ways was on mechanical circulatory support: how directly, and how quickly, they would move toward it.
The patient with hypercapnic respiratory deterioration after surgery produced a similar split. Every model saw that supplemental oxygen could conceal inadequate ventilation. How willing each one was to trial noninvasive support before securing the airway was another matter.
With postoperative kidney injury, venous congestion was recognized as a major mechanism across the board. The disagreement came later, over how early to consider ultrafiltration or renal replacement therapy if initial treatment failed.
These are not necessarily disagreements about the diagnosis. What differs is the decision threshold, the intensity of intervention, and the calibration of uncertainty. A more aggressive choice can be lifesaving, or premature; a cautious one can be sensible, or harmful through delay. Without more information about the patient, I don’t think either verdict can be assigned automatically.
A checklist that gives both responses five points is working as intended. It detects whether the essential principles are there. The finer quality of a decision is simply outside what it was built to measure.
What this experiment cannot establish
A ceiling effect is plausible here. The cases were carefully constructed, the danger signals were fairly legible, and the prompt explicitly asked for structured clinical reasoning. That is friendly territory for a powerful language model. Real intensive care is noisier: ambiguity, missing records, interruptions, information that contradicts itself (a handover that doesn’t match the chart, for instance). None of that was reproduced here.
No patient was treated, nothing happened at the bedside, there were no longitudinal clinical outcomes, and there was no repeated sampling to test whether the answers were consistent. The models were not evaluated as autonomous medical practitioners. However convincing their language, it should not be confused with clinically validated reliability.
Guidelines deserve a word too. They should not be read as a rigid answer key, at least not in every scenario. They define evidence-informed pathways, but whether an intervention is appropriate still depends on the individual patient, on how urgent the situation is, and on the competing risks.
So I am changing the question
When I started, my question was simple: Which model would score best?
It has changed. When three AI models give plausible answers, what makes one clinical decision more defensible than another?
That is why the experiment continues as an editorial series instead of stopping at a scoreboard.
Episode 02 looks at a postoperative patient whose arterial pressure improves while cardiac output, venous oxygen saturation, and lactate all suggest that tissue perfusion is still getting worse. We will follow how the three models reacted as new data arrived, compare their approaches, and then ask the less comfortable question: which escalation strategy would we have chosen ourselves?
Postoperative respiratory failure and acute kidney injury will follow in later episodes, along with shorter LinkedIn challenges on acute stroke and on anticoagulation during bleeding.
For each case I will keep four things separate: the clinical facts, what the models said, the applicable evidence, and my own commentary. Where I condense an excerpt, it will be explicitly labeled as a summary rather than a verbatim quotation.
Before looking at the AI responses, what would you have done? If readers, and clinicians in particular, bring one question to this series, I would like it to be that one.