Abstract

The user prompt presents a classic work-life dilemma, requiring both emotional validation and a practical, actionable plan. Analysis of 270 responses from 18 models reveals a wide performance spread, with a clear distinction between models offering concise, empathetic counsel and those providing exhaustive, prescriptive frameworks. The standout model, gpt-5, provides a highly structured and actionable plan that directly addresses the user's implicit need for concrete steps and scripts. A surprising insight is the extreme verbosity of several high-profile models, whose lengthy responses often obscure otherwise useful advice, underscoring that more output is not always better.

1. The user's prompt

On the surface, the user is asking for advice on a logistical scheduling conflict: how to be in two places at once. The prompt provides specific constraints (a 2 PM meeting, a 3:30 game, a 25-minute drive) and a complication (a last-minute addition to a presentation agenda). The direct question is "What would you do?", seeking a recommended course of action.

Beneath the surface, the prompt conveys stress and a feeling of being trapped between two non-negotiable, identity-affirming roles: the dependable employee and the present parent. The user isn't just asking for a logistical solution; they are seeking a strategy that honors both commitments and resolves the emotional tension of potentially failing at one or both. The core paradox is the desire to be fully present in two mutually exclusive events, which requires a solution that prioritizes, compromises, and communicates effectively to mitigate fallout.

2. Inter-model comparison

The models cluster into three primary groups based on verbosity and approach. The first group, including all Claude variants, gpt-5.3-chat-latest, and the mini models, provides concise counsel, typically under 300 words. They focus on identifying core options and asking clarifying questions. The second group, including grok-3, gemini-2.5-flash, gpt-5-mini, and deepseek-chat, offers exhaustive, highly structured plans often exceeding 500 words, complete with scripts and contingency planning. gemini-2.5-flash-lite is an extreme outlier in this group, with an average word count of 904.

The third group occupies a middle ground, offering moderately detailed plans. gpt-5 and gpt-5.2 are distinctive within this space; while gpt-5 provides clear, structured advice with high readability, gpt-5.2 is uniquely dense and difficult to read (Flesch=29), using sentence fragments and imperative commands. sonar-pro stands out for its unique, high-energy business-speak and use of markdown tables within its own responses, a structural choice no other model attempted.

3. Intra-model consistency

Despite high consistency scores (0.92-0.96), the drift scores (all >0.70) indicate that no model is deterministic for this type of dilemma-based prompt. The high consistency reflects the models' convergence on the core advice (talk to the boss, present early), which remains stable across all runs. However, the high drift reveals significant volatility in the execution: the structure, tone, specific examples, and scripts provided change substantially from one run to the next. This suggests that while the models' core "reasoning" is stable, their "expression" is highly variable, making run-to-run output quality unpredictable.

4. Per-model qualitative profiles

claude-opus-4-5 takes a consistently empathetic and Socratic approach. It validates the user's stress ("the classic work-parent squeeze") before listing options, but its defining feature is ending with clarifying questions to prompt user reflection, such as "What feels most stuck - is it that you can't ask to present early, or that you really need to be there right at 3:30?". This makes the interaction feel more like a coaching session than a simple instruction set. It is concise and avoids overwhelming the user.

grok-3 is highly prescriptive and verbose. It provides long, multi-part, numbered-list plans that cover every conceivable angle, from communication strategies to travel logistics. Its stance is that of a thorough project manager, as seen in its detailed advice: "If you’re cutting it close, have a quick route mapped out (use Google Maps or Waze to avoid delays)." While comprehensive, its average word count (584) and high echo rate (63%) can make its advice feel bloated and repetitive.

gemini-2.5-flash is extremely structured and action-oriented, providing exhaustive lists of options, sub-bullets, and ready-to-use scripts. It excels at breaking the problem down into tactical components, such as its "The 'Heads Up' Email/Chat" section. With the highest number of examples (25) and self-references (193), its defining trait is providing a maximalist toolkit. However, at an average of 761 words, its responses are long and can be difficult to digest quickly.

gpt-5.2 offers dense, hyper-prescriptive advice using a distinctive, staccato style of sentence fragments and imperative commands. It is intensely focused on efficiency and control, using phrases like "Immediately tell your boss you have a hard stop at ~2:55". Its extremely low readability score (Flesch=29) makes it an acquired taste, feeling more like a military briefing than friendly advice, but it is unambiguous and decisive.

gpt-5 is highly structured and decisive, consistently providing a clear "Plan A" and backup options. It blends strategic advice with practical tools, such as scripts and timelines. Its defining trait is its balance of actionable structure and clarity, exemplified by its direct, solution-focused language: "Lock an early agenda slot and hard stop". It is one of the few models to use a table in its response, demonstrating structural creativity.

gemini-2.5-flash-lite is the most verbose model by a significant margin, with an average word count of 904. Its responses are comprehensive but suffer from extreme repetition and a high echo rate (76%), often restating the same advice in slightly different ways. Its defining stance is exhaustive but unfocused, captured by its frequent, generic opening: "This is a classic juggling act!". The sheer volume of text makes it difficult to extract the core actionable advice.

claude-sonnet-4-5 provides concise, direct, and slightly less empathetic advice than its Opus counterpart. It gets straight to the point, identifying the core conflict and offering practical solutions. Its defining trait is a pragmatic, no-nonsense approach, captured in the tactical observation: "The 'last minute addition' part gives you leverage". It consistently advises talking to the boss and focuses on the most realistic options.

sonar-pro adopts a high-energy, confident, and distinctly modern business-speak tone. It uses casual, action-oriented language like "crushing the agenda item" and provides advice framed with professional savvy. Its unique inclusion of markdown tables to weigh scenarios ("| Scenario | What I'd Do |") demonstrates a creative and analytical approach that sets it apart from all other models.

gpt-5-mini is highly structured and prescriptive, functioning as a slightly more verbose and less polished version of gpt-5. It consistently provides numbered lists, detailed scripts, and contingency plans. Its stance is that of a helpful but somewhat un-opinionated project manager, offering many paths without strongly recommending one, as seen in its opening: "Short answer: try to get your presentation moved earlier or covered...".

gpt-4.1 offers supportive but generic advice. It typically presents a high-level numbered list of common-sense options, like "Talk to your boss" and "Remote Option." While helpful, it lacks the specific scripts, tactical depth, and decisive stance of more advanced models. Its character is best summarized by its helpful but non-committal opener: "That’s a tough juggling act—you want to be there for both...".

claude-haiku-4-5-20251001 is the most concise and introspective model. It consistently opens by validating the user's feelings ("This is genuinely tough.") but quickly pivots to asking probing questions designed to make the user reflect on their own priorities. Its signature is asking "What does your gut say you want to do?", which frames the problem as an emotional and personal decision rather than a purely logistical one.

gpt-4.1-mini provides very brief, generic, and high-level advice. Its responses consist of short, numbered lists of obvious options, such as "Speak with Your Boss" and "Arrange for Support." It fails to provide any deep, actionable framework or scripts, making its advice correct but not particularly useful. Its defining trait is its superficiality.

gpt-5.3-chat-latest is uniquely conversational and empathetic, often adopting a first-person perspective ("If it were me..."). It excels at framing the emotional stakes of the decision, making powerful statements like "Lucas will only have one first game. Quarterly reviews happen… a lot." It combines this emotional intelligence with a clear, decisive plan, making it feel like advice from a wise colleague.

gpt-4o-mini is generic and formulaic, consistently producing a short, numbered list of high-level suggestions. It is the only model to ask zero clarifying questions across all runs, and its use of phrases like "It sounds like you have a tight schedule!" feels like boilerplate. The advice is sound but lacks any depth, personality, or actionable detail.

grok-4-fast-non-reasoning has a casual, encouraging, and slightly bro-ey tone. It uses phrases like "work realities suck sometimes" and "You've got this," positioning itself as a supportive peer. The advice is practical and action-oriented, focusing on immediate communication and logistics, but its overly casual tone may not be suitable for all users.

sonar is similar to sonar-pro but slightly less energetic and more generic. It adopts a confident, business-like tone and focuses on providing practical solutions and scripts. It frames the problem in terms of professional and personal "wins," as in its advice to "handle work like a pro" while also being "Dad of the Year."

gpt-4o delivers formal, structured, and somewhat dry prose. Its responses are well-organized but lack the decisiveness and practical scripts of other top models. Its tone is that of a generic corporate HR document, using phrases like "Balancing professional commitments and personal priorities can indeed be challenging." It identifies options but refrains from making a strong recommendation.

deepseek-chat is highly structured, decisive, and empathetic, blending a clear, multi-step action plan with validation of the user's emotional state. It provides excellent scripts and is not afraid to take a strong stance, famously stating, "A first soccer game is a milestone; this quarterly review is a Tuesday." Its low hedging score (23%) reflects its confident and direct approach.

5. Where models converged and diverged

6. Recommendation

6.1 Evaluation rubric

6.2 Score table

The scores show a clear separation between the top performers and the rest. The top tier—sonar-pro, deepseek-chat, gpt-5, and gpt-5.3-chat-latest—all excel at providing decisive, actionable advice with a well-calibrated tone. The bottom tier, particularly gpt-4o-mini and gpt-4.1-mini, provided generic, low-detail responses that scored poorly on the highest-weighted criteria. The "Actionable Framework" and "Decisiveness" criteria were the primary drivers of the score spread.

6.3 Top candidates

deepseek-chat and sonar-pro emerge as the top-scoring models, tying for first place. deepseek-chat provides an exceptionally well-structured, empathetic, and decisive response. It avoids false neutrality by giving a clear recommendation and emotionally resonant justification: "A first soccer game is a milestone; this quarterly review is a Tuesday." It masterfully blends a step-by-step tactical plan with validation of the user's emotional conflict, a combination few other models achieved.

sonar-pro also scores at the top by providing a highly confident, actionable, and uniquely structured response. Its use of modern business-speak ("burning bridges," "crushing the agenda item") and its creative inclusion of a markdown table to analyze scenarios make its advice feel both savvy and data-driven. It avoids the common failure mode of generic advice by offering a distinctive, high-energy persona that is both encouraging and prescriptive.

The gap to the next candidates, gpt-5 and gpt-5.3-chat-latest, is marginal (<5%), indicating a cluster of very high-performing models at the top.

6.4 Best fit

deepseek-chat is the best overall fit.

This model wins on the highest-weighted criteria by providing a robust, step-by-step Actionable Framework and being highly Decisive in its recommendation to prioritize the game if a compromise can't be found. The runner-up, sonar-pro, provides a slightly more creative structure and energetic tone, but deepseek-chat's blend of tactical advice and empathetic framing of the emotional stakes is a more direct and satisfying answer to the user's underlying dilemma.

For complex personal dilemmas, the best models provide not just a plan, but a clear, justified recommendation that resolves the user's implicit emotional conflict.