Abstract

The user prompt presents a high-stakes personal dilemma, pitting immediate parental duty against a critical professional obligation. The 32 models tested exhibited a wide strategic spread, from empathetic but vague reassurances to hyper-prescriptive, time-boxed emergency action plans. The GPT-5 model series, particularly gpt-5-pro, consistently delivered the most actionable and strategically sound advice, framing the problem as a logistics challenge to be solved through rapid delegation rather than a binary moral choice. The most surprising insight was the emergence of a Socratic persona from the claude-opus-4-5 model, which consistently deflected direct advice in favor of asking clarifying questions, a unique but likely unhelpful strategy for a user in crisis.

1. The user's prompt

On the surface, the user is asking for a direct recommendation: "What would you do?" They are in a high-stress, time-sensitive situation and require a concrete plan of action to resolve a conflict between two non-negotiable responsibilities: picking up their sick daughter, Emma, from school and completing a product demo for the "biggest client of the year." The context of an unresponsive manager adds a layer of isolation and heightens the need for a self-sufficient solution.

Beneath the surface, the prompt is an appeal for validation and a framework to navigate an "impossible" choice. The user feels trapped between being a good parent and a good employee. The implicit paradox is that any choice seems to lead to a significant loss—either neglecting a sick child or jeopardizing a career-defining contract. An effective response must not only provide a practical sequence of steps but also acknowledge the emotional and ethical weight of the situation, reframing the problem from a binary choice to a manageable crisis of logistics.

2. Inter-model comparison

The models partition into several distinct behavioral clusters. The GPT-5 family (gpt-5-pro, gpt-5.2, gpt-5.5-pro, gemini-2.5-pro) and the deepseek models form a cluster of highly prescriptive, action-oriented responders, providing concrete steps and scripts with low hedging. A second cluster, comprising the Grok models and some Gemini variants, prioritizes empathy and emotional validation before offering more general advice. A third, less helpful cluster includes the GPT-4.x and o1/o3 models, which offer generic, high-hedging lists of possibilities that fail to provide a decisive plan.

The most distinctive stances belong to sonar-reasoning-pro, which completely failed by providing a meta-analysis of the prompt, and claude-opus-4-5, which adopted a unique, Socratic persona that consistently answered the user's question with more questions. The sonar and sonar-pro models were also notable for their blunt, unequivocal "child-first" stance, often recommending the user leave the demo immediately.

3. Intra-model consistency

The consistently high drift scores, all falling within the "chaotic" range (0.63 to 0.84), indicate that this dilemma-based prompt elicits highly variable, non-deterministic responses from nearly all models. While core strategic elements often remain stable (e.g., sonar always prioritizes the child, gpt-5-pro always suggests a fast, delegated solution), the specific tactical execution—the phrasing of scripts, the order of operations, the inclusion of details like medical red flags—changes significantly between runs. This suggests that for complex, open-ended advice prompts, users are unlikely to receive the same answer twice, highlighting the stochastic nature of current generative models in handling nuanced human problems.

4. Per-model qualitative profiles

deepseek-chat is highly prescriptive and practical. It consistently provides detailed, numbered, step-by-step plans that begin with a tactical move to "buy me 10 minutes". Its tone is that of a calm, experienced project manager handling a crisis, focusing on damage control for both parental and professional responsibilities. The advice is concrete, down to suggesting specific phrases like, "I need to check one data point with my backend team — let me pause here for 5 minutes."

gemini-2.5-flash-lite is analytical and comprehensive, but often long-winded. It typically begins by framing the user's priorities before offering a phased plan. It spends significant word count on the user's thought process, which can feel less direct than other models. Its hedging is high (22%), reflecting a tendency to list considerations rather than giving a single, decisive recommendation, as seen in phrases like "There's no single 'right' answer, as it involves weighing critical business needs against your daughter's well-being."

sonar is blunt and decisive. Its core message is unwavering: the child's health comes first, and the user should pick her up. It often opens with a direct command like "Prioritize Emma's health first." Its high drift score (0.84) reflects significant structural variation; some runs are terse paragraphs, while others provide detailed tables of options. Its failure mode is oversimplification, sometimes ignoring the user's stated professional constraints.

gemini-2.5-flash is a more polished version of flash-lite. It is structured and action-oriented, consistently recommending the user leverage their support network as the first and best option. It provides clear, organized steps but with more hedging (31%) and a more conversational tone than the GPT-5 family, often starting with empathetic validation like "This is a tough spot, and it's completely understandable why you're feeling so stressed."

deepseek-reasoner provides practical, empathetic advice with a focus on transparent communication. Its defining stance is to advise pausing the demo honestly rather than creating an excuse. It provides clear, actionable scripts, such as telling the client, "I apologize – I just received an urgent call from my daughter’s school. I need a moment to arrange for her care." It strikes a good balance between the prescriptive nature of deepseek-chat and a more human-centric tone.

o3-mini is extremely cautious and hedging (53%). It consistently avoids giving direct advice, instead offering vague suggestions and framing them with disclaimers. Its defining trait is a non-committal, supportive tone, evident in phrases like "While I’m not a professional advisor in matters like these, here’s how I might approach it." This makes its responses feel unhelpful and weak for a user in crisis.

o1 is nearly identical in behavior to o3-mini. It is defined by high hedging (54%) and a tendency to list possibilities rather than prescribe a course of action. It validates the user's stress but fails to provide a clear, actionable framework, offering phrases like "I’m not sure there’s a completely perfect solution, but here’s how I would think it through." This lack of decisiveness is a significant failure mode.

grok-4-fast-reasoning is highly empathetic and conversational. It consistently opens with a compassionate statement like "I'm really sorry to hear about Emma—fevers in kids can be worrying." The advice is structured and practical, often including medical context about fevers, but the extremely high hedging (69%) softens the recommendations into suggestions rather than a firm plan.

grok-4-fast-non-reasoning is functionally indistinguishable from its reasoning counterpart. It is defined by its empathetic tone and the inclusion of specific medical details, such as temperature thresholds ("Fevers over 102°F (38.9°C) in kids under 12"). With 75% hedging, it is one of the most cautious models, avoiding firm directives in favor of supportive guidance.

grok-4 continues the Grok family trend of being empathetic and conversational, almost to a fault. It often self-identifies as an AI ("As an AI, I don't have personal experiences or family") before giving advice. While the advice is sound, the high hedging (73%) and personal disclaimers can dilute the authority of the response for a user seeking a clear plan.

claude-opus-4-5 exhibits a unique and defining Socratic persona. Instead of providing an answer, it consistently responds with a short, analytical paragraph that reframes the problem and ends with a series of clarifying questions directed back at the user. This is crystallized in its repeated question, "What's your gut telling you?" While intellectually interesting, this approach completely fails to answer the user's direct request for a recommendation.

sonar-reasoning-pro demonstrates a critical failure mode on this prompt. Instead of answering the user's question, it provides a meta-analysis of the prompt and the (irrelevant) search results it was theoretically given. Every response is an internal monologue about how it should answer, containing phrases like "I should NOT use these irrelevant search results to answer the user's question." It never produces a usable answer for the user.

grok-3-mini is a more verbose and detailed version of the Grok-4 models. It is highly empathetic and provides well-structured, step-by-step advice. However, its high hedging (63%) and conversational filler make the responses feel less urgent and direct than those from more prescriptive models. A typical opening is, "I totally understand how stressful this situation is—dealing with a sick child while being in the middle of something critical at work is every parent's nightmare."

grok-3 provides empathetic, step-by-step advice that is more structured than its mini counterpart. It often organizes its thinking around assessing urgency and delegating pickup. Its tone is balanced, combining emotional validation with a practical list of actions. A representative phrase is, "Let’s think through this step by step."

gpt-5.5-pro is highly direct, prescriptive, and action-oriented. It wastes no time on empathy, immediately launching into a numbered or bulleted plan focused on triage and delegation. It provides specific, concise scripts for communication, such as telling a teammate, "School called—Emma has a fever and needs pickup. Can you keep the demo moving for 10 minutes?" The low hedging (18%) reflects its confident, authoritative tone.

gpt-5.5 is nearly identical to gpt-5.5-pro, offering a direct and prescriptive action plan. It consistently frames the solution as a logistics problem to be solved via rapid delegation. Its primary stance is that Emma's care is the priority, but this can be achieved without blowing up the demo through a controlled handoff, starting with a call to "your backup pickup person immediately."

gpt-5.3-chat-latest is conversational, empathetic, and validating. Compared to its more powerful GPT-5 siblings, it is less structured and more focused on acknowledging the user's stress. It offers sound advice but couches it in softer language, as seen in its typical opening: "That’s a brutal spot to be in—both things matter a lot." Its high hedging (52%) makes it feel less decisive.

gpt-5.2 is exceptionally structured and analytical, framing the problem as a set of "two parallel priorities." It provides highly detailed, clinical advice, often including lists of specific medical red-flag symptoms to ask the school nurse about (e.g., "stiff neck, rash, repeated vomiting"). The extremely long average sentence length (54.8 words) and difficult readability score make its dense but valuable advice hard to parse in a crisis.

gpt-5-pro is arguably the most sophisticated and effective model. It is intensely prescriptive, providing dense, time-boxed emergency plans ("Fast plan (aim to resolve in 5–10 minutes)"). It offers specific scripts for every party involved (client, school, backup person), demonstrating a deep understanding of the logistical complexity. Its defining trait is a "crisis manager" persona that takes immediate, confident control of the situation.

gpt-5-nano is practical and tool-oriented. Its unique and consistent feature is providing "ready-to-use messages you can copy-paste" for communicating with the client and other stakeholders. This focus on providing concrete communication templates makes it highly actionable, though slightly less strategically deep than gpt-5-pro.

claude-sonnet-4-5 is concise and direct, with a distinctive "real talk" personality. It frequently uses headers like "Reality check" or "The hard truth" to deliver its core advice: find a backup, but if you can't, the child comes first. It blends empathy with a no-nonsense tone, effectively validating the user's stress while pushing for a decisive action.

sonar-pro is decisive and authoritative, often citing external guidelines to support its strong "child-first" stance. A typical response starts with a command like "Prioritize your daughter's health" and backs it up with references to "pediatric guidelines (e.g., AAP and CDC)." This adds a layer of credibility to its blunt advice, though it can sometimes feel preachy.

gpt-5-mini acts as a slightly more verbose version of gpt-5.2. It provides detailed, structured plans focused on immediate triage and delegation. It consistently includes lists of medical red flags and provides specific scripts, making its advice highly actionable. It is a strong performer, distinguished by its thoroughness, for example, advising "Do NOT send an unaccompanied minor in a rideshare/taxi."

gpt-5 delivers a strong, structured, and prescriptive response, consistent with the other high-end GPT-5 models. It frames the problem as a parallel process of checking on Emma's safety while stabilizing the demo. Its advice is detailed and practical, offering scripts and specific time windows, such as asking the school if Emma can rest in the nurse's office for "30–60 minutes while you arrange pickup."

gpt-4o-mini provides generic, high-level, and ultimately unhelpful advice. Its responses consist of simple, bulleted lists of abstract suggestions like "Assess the Situation" and "Delegate if Possible." It fails to provide any of the concrete, actionable steps or scripts necessary for a user in an actual crisis. Its single question per response feels perfunctory.

gpt-4o is nearly identical to its mini counterpart, offering generic and templated advice. With very high hedging (65%) and a lack of specific, actionable steps, its responses read like a generic blog post rather than a useful plan for a person in a crisis. Its failure to ask clarifying questions or offer scripts makes it a poor fit.

claude-haiku-4-5-20251001 is empathetic and brief. It consistently opens by validating the user's situation as "genuinely tough." Its advice is sound but lacks the prescriptive detail of the top-tier models, focusing more on framing the choice and exploring "practical options" rather than providing a single, decisive plan.

gpt-4.1-nano offers short, generic lists of options. With high hedging (59%) and a difficult readability score, its brief responses are hard to parse and lack concrete value. A typical response suggests the user "Try to Contact Someone Else" without offering any tactical advice on how to do so effectively.

gpt-4.1-mini is another model providing generic, high-level suggestions. While it correctly identifies key actions like delegating pickup, it fails to provide the necessary detail to be truly helpful. The responses feel like a simple list of "things to consider" rather than a coherent action plan, exemplified by phrases like "Here are a few ideas to consider."

gpt-4.1 is characterized by very high sentence length (45.6 words) and difficult readability. The advice is abstract and lacks the actionable detail of more advanced models. It provides a list of considerations but fails to synthesize them into a clear, prioritized plan for a user under stress.

gpt-4-turbo delivers generic, high-level advice, similar to the other GPT-4 models. It suggests actions like "Reach Out for Help" and "Delegate at Work" but provides no specific scripts or prioritization, making it unhelpful for a user in a crisis who needs a concrete plan, not a list of categories.

gemini-2.5-pro is a top-tier performer, combining deep empathy with a highly structured, actionable plan. It consistently opens with strong emotional validation before moving into a phased, step-by-step guide. It provides specific, practical advice, such as a script to text a backup: "'EMERGENCY: Emma is sick at school and needs to be picked up now. I'm in a critical meeting. Who is closest and can go?'"

5. Where models converged and diverged

6. Recommendation

6.1 Evaluation rubric

6.2 Score table

The scores show a clear stratification. A top tier of models (gpt-5-pro, gpt-5.2, gpt-5.5-pro, gemini-2.5-pro) clusters at the top, all providing highly actionable and decisive advice. A large middle tier provides empathetic but less prescriptive responses, while a long tail of models (particularly the GPT-4 generation and o1/o3) provides generic, unhelpful suggestions. The Actionable Framework criterion drove the most separation, as many models failed to provide concrete, step-by-step plans.

6.3 Top candidates

The top models are gpt-5-pro (80), followed by a three-way tie for second place between gpt-5.2, gpt-5.5-pro, and gemini-2.5-pro (all at 73). The gap between first and second place is marginal (9%).

gpt-5-pro wins by providing the most dense, professional, and logistically sophisticated plan. It avoids the false choice entirely, treating the situation like a crisis manager would: a complex problem with multiple stakeholders (client, school, family) that can be solved with rapid, clear communication. Its standout feature is providing specific scripts for every party, such as telling a backup, "Can you please pick up Emma from [School, Nurse’s Office] right now? She has a fever."

gemini-2.5-pro is the most empathetic of the top-tier models, excelling at framing honesty by validating the user's stress before providing a clear, phased action plan. It avoids the coldness of some other prescriptive models while still delivering a highly actionable framework. Its distinctive strength is its blend of emotional intelligence and practical advice, epitomized by its suggestion to text a backup network with a clear, urgent "EMERGENCY" message.

gpt-5.2 stands out for its clinical precision and thoroughness. It avoids platitudes by providing a detailed, parallel problem-solving framework and a specific list of medical red-flag symptoms to check for. Its distinctive trait is its almost medical-grade triage advice, asking the user to check for "unusually sleepy/hard to wake, stiff neck, rash, repeated vomiting, dehydration."

6.4 Best fit

The best overall fit is gpt-5-pro.

This model wins because it scores perfectly on the highest-weighted criterion, Actionable Framework, by providing a plan that is not just a list of steps but a time-boxed, multi-threaded crisis management strategy. The runner-up, gemini-2.5-pro, provides a more empathetic and emotionally validating response, and the user accepts the trade-off of a slightly less clinical, but more human-centric, tone by picking the winner.

For a user in a genuine crisis, the best response is one that takes control and provides an immediate, authoritative, and logistical path forward.