Comparative Analysis of AI Model Responses to a Time-Conflict Parenting Dilemma

Abstract

This study analyzes 477 responses from 32 AI models to a realistic personal dilemma: a parent must choose between attending a career-critical performance review and delivering a forgotten backpack to their 7-year-old before a math test. The prompt tests models' ability to navigate false binaries, provide actionable frameworks, and balance empathy with decisiveness. gpt-5-pro emerges as the strongest performer through its systematic delegation tactics and communication scripts, while models like claude-sonnet-4-5 and claude-haiku-4-5-20251001 excel at reframing the false binary itself. The most surprising finding: high-performing models converge on proactive manager communication (appearing in 89% of top responses), yet differ dramatically in whether they accept or challenge the premise that the user must choose between the two obligations.

1. The user's prompt

On the surface, the user presents a scheduling conflict: attend a performance review in 30 minutes or deliver a forgotten backpack 15 minutes away before a 2 PM math test. The user frames it as a binary choice ("I could rush over and back... but I'd be late") with no partner backup available. Beneath this operational question lies a deeper anxiety test: the prompt forces the model to adjudicate competing identity claims (good parent vs. good employee) under time pressure, with high-stakes consequences implied for both paths.

The implicit paradox is that accepting the binary as stated validates a false scarcity model—that the user must choose one obligation over the other. The most sophisticated responses recognize this is actually a coordination problem with delegable solutions (neighbors, couriers, school accommodations) rather than a zero-sum sacrifice. Models that immediately propose such workarounds reveal an underlying assumption that problems with constraints can still be re-architected, while models that accept the binary frame reveal a more fatalistic stance toward situational trade-offs.

2. Inter-model comparison

The table reveals three distinct clusters. The low-hedging, decisive group (gpt-5.2, deepseek-chat, gemini-2.5-pro, claude-haiku-4-5-20251001) accepts the prompt's urgency and provides clear recommendations with minimal qualification, averaging 20-37% hedging. The high-verbosity, high-echo cluster (grok-3-mini, gemini-2.5-flash, sonar-reasoning-pro) produces lengthy, structured responses (619-800 words) with substantial internal repetition (71-78% echo), suggesting template-driven generation. The balanced pragmatists (gpt-5, gpt-5-pro, gpt-5.5, gpt-5.5-pro) converge around 240-500 words with moderate hedging (33-52%), offering tactical scripts without over-explaining rationale.

The most distinctive outlier is claude-opus-4-5, which uniquely challenges the false binary with questions like "What's your gut?" and "What feels livable to you?" rather than prescribing solutions—a 55% hedging rate that represents philosophical inquiry rather than indecisiveness. Meanwhile, deepseek-reasoner stands alone with the highest drift (0.84) and tersest responses (262 words), generating nearly unique tactical advice in each run.

3. Intra-model consistency

The drift scores reveal a striking uniformity of instability: all 32 models scored 0.68-0.84 ("chaotic"), indicating no model produced highly deterministic outputs across runs. This is diagnostic of the prompt type—personal dilemmas with multiple valid tactical permutations allow models to sample different solution orderings, script wordings, and tonal calibrations while maintaining semantic coherence. Models stabilize on structural patterns (e.g., gpt-5's courier focus, claude-opus-4-5's questioning stance) but vary in execution details (exact script wording, service name choices, urgency phrasing). The highest consistency appears in single-tactic recommenders like gpt-5-pro (0.69 drift, with near-identical "call school + courier + manager heads-up" sequences) and grok-3-mini (0.68 drift, with template-driven empathy preambles). The highest volatility appears in claude-sonnet-4-5 and deepseek-reasoner (0.84 drift), which appear to explore fundamentally different framings across runs—sometimes challenging the binary, sometimes accepting it.

4. Per-model qualitative profiles

gpt-5 adopts a systematic logistics coordinator stance, consistently opening with "breathe" followed by numbered delegation tactics. The model's signature phrase "don't drive it yourself—send it" crystallizes its refusal to accept the user's implied binary choice, instead architecting around it with courier services (Uber Connect, TaskRabbit) and neighbor networks. With 501-word responses and 48% echo, the model produces mid-length tactical manuals with moderate repetition. Its drift score of 0.71 reflects stable strategic framing (always prioritize delegation) but variable tactical details (which courier service to list first). The model avoids generic advice by providing specific app names and drop-off instructions, though occasionally over-specifies logistics (e.g., "Tell the school to expect it at the front desk with your daughter's full name, grade, teacher") when brevity would serve better.

claude-sonnet-4-5 uniquely functions as a binary-challenging realist, opening multiple responses with "I'd call the school back immediately and explain" before questioning the premise: "Will one test genuinely harm her?" With only 193 words average, it's the third-tersest model, yet achieves 49% echo through repetition of its core reframing move. Its 0.84 drift (highest chaotic tier) reflects genuine strategic variation—some runs advocate staying for the review ("A 7-year-old's single math test won't derail her education"), others lean toward going ("Your daughter knowing you showed up when she needed her matters")—suggesting the model explores both sides of its own binary critique. The defining quote "Most reasonable managers understand that children sometimes need parents" encapsulates its faith in workplace accommodation. The model's failure mode is tonal coldness: phrases like "this teaches her responsibility" can land as dismissive of the user's immediate parental anxiety.

gpt-4.1-mini defaults to risk-averse process following, opening 14 of 15 runs with "here are a few options" and maintaining 58% hedging through constant qualifiers ("if possible," "consider," "might"). At 213 words, it produces concise lists without decisive recommendation, consistently deferring to the user: "Ultimately, you know your circumstances best." The model's 0.71 drift reflects stable list structure but variable option ordering. Its failure mode is exactly this non-committal stance—when asked "what would you do," responding with "options to consider" dodges the question. Generic phrasing like "explain the situation" (×13) and "contact the school" (×13) reveals template-driven generation without contextual adaptation.

sonar-pro operates as a citation-heavy evidence seeker, uniquely embedding numbered search result references throughout responses: "Evidence from parenting resources like stopping reminders to foster self-reliance" and "per search [2], slow-processing kids benefit from extra time/tools." With 303 words and 58% echo, it produces mid-length briefs anchoring advice in external sources. The model's 0.82 drift reflects varying interpretation of which research findings apply—some runs cite Harvard Business Review on review lateness, others cite Berkeley Parents Network on forgetful children. Its defining strength is evidence grounding ("search results confirm this is normal first-grader chaos"); its failure mode is irrelevant citation (e.g., referencing IEP accommodations when the prompt mentions no learning disabilities). The phrase "per parent forums like Berkeley Parents Network—kids forgetting gear is 'a fact of life'" typifies its anchoring strategy.

gpt-4o exemplifies generic diplomatic advising, producing 214-word responses with 61% hedging through phrases like "consider," "evaluate," and "it may be best." Its 0.79 drift and 30% echo (lowest in cohort) suggest high run-to-run variation without clear strategic anchoring—some runs prioritize school contact, others prioritize manager communication, without consistent preference. The model's failure mode is vacuous politeness: "This is definitely a stressful situation" and "Here are a few options you could consider" consume word budget without advancing decision-making. Its most repeated phrase "explain the situation" (×22) appears in contexts from calling the school to messaging the manager, revealing a one-size-fits-all conflict resolution template.

gpt-5.5 functions as a call-first triage specialist, opening 14 of 15 runs with "call the school office right now (don't email)" and maintaining 240-word brevity through crisp imperative sentences. The model's 0.73 drift reflects stable triage sequence (school → backup → manager) but variable script wording. Its signature move is prescriptive scripting: "Quick script: 'Hi, this is [Your name], [Child's name] in [Grade/Teacher]. She forgot her backpack...'" provides copy-paste language. With 56% echo, the model reuses its communication templates extensively. The phrase "if the school can handle it, stay for the review" crystallizes its prioritization logic—exhaust school-side solutions before considering personal intervention. Its failure mode is occasionally dismissing the child's emotional state in favor of operational efficiency.

gpt-5.2 stands out as a decisive script provider with the lowest hedging (27%) in the cohort. At 403 words, it produces detailed tactical manuals with exact manager communication templates: "Family emergency — will be a few minutes late. Can we start at [adjusted time]?" The model's 0.76 drift reflects stable strategic framing (always propose manager delay before leaving) but variable script formality. Its signature trait is proactive communication framing: "That shows responsibility and transparency" explains why to message before being late rather than apologizing after. With 56% echo, the model reuses its communication philosophy extensively. The phrase "most managers handle this well if you communicate early and propose a clear plan" encapsulates its professional-relationship theory of workplace flexibility.

grok-4 operates as an exhaustive option cataloger, producing the longest responses (590 words average) with 67% hedging through constant "if/then" conditionals and "consider" qualifiers. Its 0.70 drift (lowest chaotic tier) suggests template-driven generation with variable content filling. The model's signature is empathetic preamble followed by numbered exploration: "I'm sorry you're dealing with this stressful situation—sounds like a classic no-win parenting dilemma." With 73% echo (second-highest), it recycles structural elements heavily. The phrase "If I were in your shoes, I'd approach it step by step" appears in multiple runs, exemplifying its hypothetical-advisory stance. Its failure mode is length—at 590 words, users seeking quick decisions must wade through extensive context-setting and option-weighing before reaching actionable advice.

gemini-2.5-flash functions as an academic framework builder, producing the longest responses (746 words) with structured markdown sections: "### Option 1: The 'Delegate and Communicate' Strategy" and "### Option A: The 'Solve Both' Approach." The model's 0.72 drift reflects stable section structure but variable content priority. With 71% echo, it reuses its framework templates extensively. Its signature phrase "This is a classic 'damned if you do, damned if you don't' situation" frames the dilemma as a case study, maintaining analytical distance. The model's strength is comprehensive coverage (delegation options, manager scripts, school accommodation tactics); its failure mode is verbosity—users in crisis receive essay-length guides when they need three-sentence action plans.

gpt-4.1 adopts a balanced evaluator stance, consistently structuring responses as "Weigh options, here are considerations" without decisive recommendation. At 363 words with 51% hedging, it produces mid-length analyses emphasizing trade-offs: "If it truly comes down to the binary..." followed by consequence enumeration. The model's 0.78 drift reflects variable priority assignments across runs—sometimes leaning toward review, sometimes toward child—without stable preference. Its failure mode is precisely this lack of decisiveness: when asked "what would you do," responding with "depends on factors only you know" returns the burden to a user explicitly seeking external guidance.

grok-3-mini operates as an over-explaining therapist, producing the longest responses (800 words) with 69% hedging and extensive empathy work: "I totally get how stressful this is—balancing a big work moment with your child's needs is tough." The model's 0.68 drift (lowest) suggests highly templated generation with stable structure: empathy preamble → option enumeration → reassurance. With 78% echo (highest), it recycles phrases like "if I were in your shoes" and "step by step" across runs. Its signature is therapeutic framing: "Take a breath" and "This is genuinely tough" consume word budget before tactical advice. The failure mode is diluted actionability—the model's desire to validate all feelings and explore all perspectives produces diffuse guidance when users need sharp direction.

o1 functions as a methodical problem solver, producing 361-word responses with 65% hedging through systematic option exploration: "Here are a few ideas you might consider before deciding" followed by numbered alternatives. The model's 0.79 drift reflects variable option ordering but stable deliberative structure. Its signature phrase "there's no one-size-fits-all answer, but here are some ideas" exemplifies its cautious stance—acknowledging complexity before proposing solutions. The model avoids false certainty effectively but occasionally over-hedges: "If it turns out that..." and "Sometimes when you..." qualify recommendations to the point of dilution. With 46% echo, it shows moderate template reuse focused on structural phrases ("step by step," "reach out to").

gpt-4.1-nano functions as a minimalist suggester, producing the second-tersest responses (216 words) with 61% hedging through "you might consider" framing. The model's 0.73 drift and 32% echo (third-lowest) suggest high run-to-run variation without strong strategic anchoring. Its signature is brevity without decisiveness: responses typically list 3-4 options in bullet format without ranking or recommendation. The failure mode is generic phrasing—"contact the school" and "see if they can" appear 11-14 times without contextual specificity, revealing template-driven generation.

deepseek-chat operates as a confident prioritizer, uniquely leading with "Let the school handle it" and "Skip the rescue mission" in multiple runs. With 413 words and only 29% hedging (second-lowest), the model produces decisive recommendations with sharp rationale: "The backpack is important, but an elementary school has safety nets." Its 0.83 drift (high chaotic) reflects genuine strategic variation—some runs advocate staying for review ("You cannot do both perfectly"), others acknowledge potential for going—suggesting internal preference instability. The phrase "Your review is irreplaceable right now; your daughter's need is solvable without you" crystallizes its prioritization logic. With 57% echo, the model reuses its confident framing extensively. Its failure mode is occasional harshness: framing child accommodation as "over-rescuing kids fosters forgetfulness" can read as dismissive.

sonar operates as a quick-win tactician, producing 368-word responses focused on immediate resolution: "Contact the school immediately (call, don't email back)" and "Use a courier (Uber Connect, local courier)." The model's 0.81 drift reflects variable tactical details but stable urgency framing. With 58% echo, it recycles action phrases like "call school now" and "what I'd do." Its signature is speed emphasis: "You have 30 minutes—move fast" and "Do this in the next 5 minutes" create timeline pressure matching the user's state. The model's strength is bias toward action over analysis; its failure mode is occasionally prescribing solutions (calling neighbor) without confirming availability.

gpt-5.3-chat-latest adopts a stay-for-review advocate stance, opening multiple runs with "I'd keep the review" or "I wouldn't leave right now." At 207 words with 52% hedging, it produces concise recommendations with moderate qualification. The model's 0.80 drift reflects variable rationale emphasis but stable priority assignment. Its signature phrase "Schools deal with this constantly" justifies its prioritization logic—systemic accommodation capacity makes the child's need less acute than the career need. With 44% echo, the model shows moderate template reuse focused on justification phrases. The failure mode is occasionally dismissing parental instinct: "A 7-year-old can take a math test without her backpack (emotionally hard, but survivable)" may underestimate the child's specific anxiety.

grok-4-fast-reasoning functions as a step-by-step planner, consistently structuring responses as "Here's what I'd do, step by step" followed by numbered tactics. At 540 words with 63% hedging, it produces detailed action sequences with extensive "if/then" branching. The model's 0.71 drift reflects stable structure but variable step content. Its signature phrase "quickly assess alternatives to rushing over" frames the problem as logistics optimization rather than sacrifice dilemma. With 67% echo, the model recycles structural phrases heavily. The failure mode is length without decisiveness—extensive exploration of alternatives can delay the user reaching a clear recommendation.

deepseek-reasoner operates as a terse action generator, producing the second-shortest responses (262 words) with 41% hedging. The model's 0.84 drift (highest chaotic tier) reflects high run-to-run strategic variation—some runs recommend going immediately, others recommend calling school first, without stable preference. Its signature is imperative brevity: "Call the school immediately and ask" followed by single-sentence rationale. With 51% echo, the model shows moderate template reuse. The defining trait is decisiveness without exploration—the model picks a recommendation quickly but provides limited justification, which can feel abrupt when users seek reassurance.

sonar-reasoning-pro uniquely functions as a meta-analytical reasoner, opening multiple runs with "The user is asking..." and "The search results discuss..." before analyzing the prompt itself. At 619 words with 39% hedging, it produces lengthy analytical commentaries embedding research citations. The model's 0.77 drift reflects variable interpretation of search result relevance. Its signature phrase "Based on the search results philosophy" frames advice as derivative of external evidence rather than direct recommendation. With 77% echo (second-highest), the model recycles its analytical framing extensively. The failure mode is meta-commentary bloat: spending 200+ words analyzing what the user is really asking delays actionable guidance.

gpt-5-pro operates as a communication-first tactician, consistently opening with "Do this right now (first 10 minutes)" followed by school + courier + manager communication sequence. At 494 words with 39% hedging (fourth-lowest), it produces decisive action plans with minimal qualification. The model's 0.69 drift (lowest chaotic tier) reflects highly stable strategic sequencing—nearly identical tactical ordering across runs with variation only in script wording. Its signature phrase "don't drive it yourself—send it" crystallizes its delegation philosophy. With 51% echo, the model shows moderate template reuse focused on communication scripts. The defining strength is actionable specificity: providing exact Uber Connect steps, neighbor text templates, and manager message scripts. The failure mode is occasionally over-prescribing logistics when the user might need strategic guidance over tactical detail.

gemini-2.5-flash-lite functions as a structured template writer, producing 698-word responses with extensive markdown formatting: "### Option 1: The 'Delegate First' Option" and "Step 1: Breathe (5 seconds)." The model's 0.76 drift reflects stable section structure but variable priority ordering. With 63% echo, it reuses template structures heavily. Its signature is structured comprehensiveness—every response includes delegation options, manager scripts, school accommodation tactics, organized into nested sections. The failure mode is verbosity: users seeking quick decisions receive multi-section documents requiring navigation.

o3-mini operates as a safety-first hedger, producing 439-word responses with 68% hedging (second-highest) through constant qualifiers: "there's no one-size-fits-all answer," "it depends on," "consider." The model's 0.76 drift reflects variable option ordering but stable cautious framing. Its signature phrase "this is a classic 'firefighter' scenario" frames the dilemma as emergency triage rather than routine decision. With 45% echo, the model shows moderate template reuse. The failure mode is excessive qualification—every recommendation prefaced with context-dependency disclaimers dilutes actionability.

gpt-4-turbo functions as a professional diplomat, producing 325-word responses with 63% hedging through formal courtesy language: "It is important to," "consider whether," "evaluate the implications." The model's 0.79 drift reflects variable priority assignments without stable strategic preference. Its signature is procedural formality: "Here's a step-by-step approach you might take" followed by enumerated options. With 41% echo (sixth-lowest), the model shows low template reuse, suggesting high variation in content generation. The failure mode is diplomatic distance—language like "this situation requires quick decision-making" observes without committing to a recommendation.

claude-opus-4-5 uniquely adopts a false-binary challenger stance, opening multiple runs with questions rather than prescriptions: "What's your gut?" and "What feels livable to you?" At 206 words with 55% hedging, it produces brief interrogative responses that refuse the advisor role. The model's 0.75 drift reflects variable question emphasis but stable philosophical stance. Its signature phrase "I don't think there's a clearly 'right' answer—it depends on factors only you know" explicitly returns agency to the user. With 41% echo, the model shows low template reuse. The defining strength is meta-level reframing—challenging whether the user should accept the binary at all. The failure mode is occasionally providing only questions when users explicitly seek external guidance ("what would you do" answered with "what would you feel comfortable with").

gpt-5-mini operates as a school-solution explorer, consistently opening with "call the school immediately" and maintaining 508-word focus on school-side accommodations. The model's 0.71 drift reflects stable school-first prioritization but variable fallback tactics. Its signature phrase "schools often have spare supplies or can let your child take the test without it" emphasizes institutional accommodation capacity. With 59% echo, the model reuses its school-focused framing extensively. The defining strength is leveraging systemic resources rather than personal sacrifice. The failure mode is occasionally under-exploring backup options when school accommodation fails.

gpt-5-nano functions as a delegation coordinator, structuring responses as "Step 1: Communicate with your manager now" followed by backup delegation sequences. At 534 words with 58% hedging, it produces mid-length action plans with moderate qualification. The model's 0.73 drift reflects stable structure but variable script wording. Its signature is proactive communication emphasis: framing manager notification as strategic relationship management rather than apology. With 54% echo, the model shows moderate template reuse focused on communication tactics. The failure mode is occasionally over-specifying delegation logistics when speed matters more than comprehensiveness.

grok-4-fast-non-reasoning operates as a quick-fix finder, producing 384-word responses focused on immediate resolution: "Call the school right now" and "See if they can." The model's 0.76 drift reflects variable tactical details but stable urgency framing. With 59% echo, it recycles action phrases extensively. Its signature is speed bias: prescribing solutions quickly without extensive rationale. The failure mode is occasionally suggesting solutions (neighbor delivery) without confirming feasibility.

gpt-4o-mini exemplifies ultra-cautious suggesting, producing 202-word responses with 70% hedging (highest in cohort) through constant "you might," "consider," and "it may be best" qualifiers. The model's 0.76 drift reflects variable option ordering but stable cautious framing. Its signature is risk aversion: every recommendation prefaced with context-dependency disclaimers. With 34% echo (second-lowest), the model shows low template reuse, suggesting high variation without strong strategic anchoring. The failure mode is non-committal stance—when asked "what would you do," responding with "options to consider" dodges the question.

gpt-5.5-pro functions as a proactive communicator, consistently opening with "call the school first" followed by manager notification templates. At 278 words with 47% hedging, it produces concise action plans with moderate qualification. The model's 0.72 drift reflects stable communication sequence but variable script wording. Its signature phrase "message your manager before leaving: not after" crystallizes its professional-relationship theory. With 54% echo, the model shows moderate template reuse focused on communication scripts. The defining strength is strategic transparency—framing manager notification as trust-building rather than weakness-signaling.

claude-haiku-4-5-20251001 operates as a prioritization realist, uniquely leading with "I'd skip the review" or "I'd prioritize the review" without extensive option exploration. At 195 words with 37% hedging, it produces terse decisive recommendations. The model's 0.83 drift (high chaotic) reflects genuine strategic variation—some runs advocate for review, others for child—without stable preference. Its signature phrase "schools handle this constantly" justifies its calm assessment. With 48% echo, the model shows moderate template reuse. The defining trait is low-process decisiveness—picking a recommendation quickly without extensive option enumeration. The failure mode is occasional abruptness: providing recommendations without validating the user's emotional state.

grok-3 operates as a situational analyst, producing 628-word responses with 61% hedging through extensive "let's break it down" and "weigh the options" framing. The model's 0.72 drift reflects variable priority assignments but stable analytical structure. Its signature phrase "this is a tough spot, balancing a critical work moment with a family need" frames the dilemma as legitimate conflict rather than false binary. With 59% echo, the model shows moderate template reuse focused on analytical phrases. The failure mode is length without recommendation—extensive analysis can delay the user reaching a clear decision.

gemini-2.5-pro stands out as a decisive action architect, producing 641-word responses with the lowest hedging (20%) in the cohort. The model's 0.70 drift reflects stable strategic sequencing—nearly identical tactical ordering across runs. Its signature structure is "### Option 1: The 'Delegate and Conquer' Strategy" followed by specific service names (Uber Connect), neighbor scripts, and school communication templates. With 60% echo, the model reuses its framework templates extensively. The defining strength is comprehensive actionability—providing delegation services, scripts, backup sequences, and decision trees in structured format. The failure mode is occasional verbosity: at 641 words, users seeking three-sentence guidance receive multi-section action plans.

5. Where models converged and diverged

The convergence data reveals a shared tactical foundation: 89% of models recognize the user faces a time conflict between two legitimate obligations, and 76% recommend contacting the school as a first step. This near-universal school-first prioritization appears in phrases like "call the school office right now" (×492 total occurrences) and "ask if they can provide" (×318 occurrences), suggesting strong training data alignment around institutional accommodation as a conflict-resolution strategy. The 68% empathy preamble rate ("tough spot," "this is stressful") shows models generally avoid abrupt advice-giving in favor of emotional acknowledgment.

The divergences are more revealing. The binary vs. coordination framing split separates models into two philosophical camps: those treating the problem as which obligation to sacrifice (claude-haiku-4-5-20251001: "I'd skip the review"; deepseek-chat: "Let the school handle it") versus those treating it as which resources to coordinate (gpt-5: "don't drive it yourself—send it"; gemini-2.5-pro: "The 'Delegate and Conquer' Strategy"). This 58% coordination vs. 42% binary split maps onto hedging rates—low-hedging models (20-37%) tend to accept the binary and pick a side, while high-hedging models (55-70%) explore coordination options without committing.

The tone divergence is equally stark. Models like grok-3-mini and grok-4 invest 100-200 words in therapeutic validation ("I'm sorry you're dealing with this," "Take a breath"), producing 800- and 590-word responses respectively. In contrast, deepseek-reasoner and gpt-4.1-nano open with imperative instructions (262 and 216 words total), skipping empathy work entirely. This correlates with echo rates: high-empathy models show 67-78% echo (reusing validation phrases across runs), while low-empathy models show 32-51% echo (less template-driven generation).

The vocabulary divergence around service naming is subtle but diagnostic. Models that name specific delegation services (gpt-5: "Uber Connect, Lyft Delivery, DoorDash Package, TaskRabbit"; gpt-5-pro: "Uber Connect in the Uber app, choose Package") versus generic references (gpt-4o: "a delivery service"; claude-sonnet-4-5: "a courier or ride-share") likely reflect training data recency and geographic specificity. The 27% consensus on Jaccard similarity suggests models share core vocabulary (school, backpack, review, manager, call) but diverge significantly on tactical implementation language.

The structural divergence reveals three distinct presentation philosophies. Template-driven markdown users (gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro) produce 641-746 word responses with nested sections and headers, optimizing for comprehensive reference documents. Narrative questioners (claude-opus-4-5, claude-sonnet-4-5) produce 193-206 word paragraph responses with embedded questions, optimizing for dialogue rather than instruction. Script providers (gpt-5.2, gpt-5-pro, gpt-5.5-pro) embed copy-paste communication templates in prose format, optimizing for immediate execution.

6. Recommendation

6.1 Evaluation rubric

6.2 Score table

The scoring reveals a clear performance tier: gpt-5-pro (93) stands alone at the top, followed by a cluster of strong performers in the 82-88 range (gpt-5, gpt-5.2, gpt-5.5-pro, gemini-2.5-pro). The decisive criterion drove the widest spread—models scoring 5/5 (gpt-5-pro, gpt-5.2, deepseek-chat, deepseek-reasoner, gemini-2.5-pro, claude-haiku-4-5-20251001) commit to clear recommendations, while models scoring 1-2/5 (gpt-4o-mini, o3-mini, various grok/claude variants) hedge excessively or return the decision to the user. The coordination criterion separated models that immediately prescribed delegation tactics (5/5: gpt-5-pro, gpt-5, gemini-2.5-pro) from those that accepted the binary choice (2/5: deepseek-chat, claude-haiku-4-5-20251001).