Independent benchmarks for large language models tend to rely on synthetic datasets and academic benchmarks like MMLU or HumanEval, where models routinely score above 90%. A new blind benchmark from DailySkill AI takes a different approach: it gives all three major chatbots the same 20 real-world consumer tasks and scores them without knowing which model produced which response.
How the Benchmark Was Set Up
The study tested ChatGPT, Claude, and Gemini across 20 everyday prompts ranging from summarizing group chat logs to planning a three-day Lisbon itinerary, drafting business emails, building workout plans, and comparing apartment listings. Each task was scored on a five-dimension rubric covering completion, accuracy, clarity, usability, and format, with a maximum of 25 points per task. Responses were anonymized and evaluated double-blind, meaning the judge did not know which model generated which output.
Tasks were drawn from actual consumer use cases, not synthetic test sets. The prompts asked for specific formats, hard constraints, and concrete outputs, which made it possible to score responses against objective criteria rather than subjective impressions.
ChatGPT Wins on Constraint Rigor
ChatGPT took the most wins overall with 8 out of 20 tasks and the highest average score at 23.9 out of 25. Its standout strength was following instructions to the letter. When the prompt demanded exactly five bullet points with no preamble, ChatGPT delivered exactly that. When a recipe prompt said to use only the listed ingredients, ChatGPT found a way to sauté without oil by substituting tomato liquid, while Claude added unlisted oil, salt, and pepper.
This pattern showed up repeatedly. In the group chat summary task, ChatGPT produced a tight five-bullet response with zero filler. In the interview prep task, it delivered exactly 10 question-and-answer pairs in the requested format without adding commentary. For developers and power users who need copy-paste-ready output that respects formatting rules, this consistency matters more than occasional brilliance elsewhere.
Claude Excels at Analysis and Safety
Claude won 5 tasks and scored an average of 22.3 out of 25. Its strongest domain was tasks requiring careful reasoning about constraints and user safety. In the workout plan task, Claude was the only model that fully respected the user's lower-back limitation, flagging borderline exercises and avoiding spinal-loading movements. ChatGPT included dumbbell thrusters in a fast circuit, a poor choice for someone explicitly told to avoid heavy spinal loading.
Claude also dominated the apartment comparison task by calculating price-per-square-foot metrics to make an objective recommendation, while the other models merely summarized features. In the meeting minutes task, Claude accurately linked action items to owners and stayed within the ten-line limit. The tradeoff is verbosity: Claude frequently adds extra paragraphs or commentary that the prompt did not request, which cost it points on format compliance.
Gemini's Niche in Conversational Copy
Gemini won 3 tasks with an average score of 22.8, actually higher than Claude's average despite fewer wins. Its strength was punchy, engaging writing. It tied with Claude on the business email task, both producing warm, well-structured reschedule messages that asked for confirmation of the new time. ChatGPT's version was accurate but too thin, reading like a single terse paragraph rather than a polished email.
Where Gemini struggled was accuracy under constraint. In the meeting minutes task, it hallucinated that onboarding documentation had no owner, even though the transcript clearly showed Dev volunteering to handle it. In the workout plan, it prescribed dumbbell thrusters despite the user's back limitation. These are not formatting errors. They are failures to read and apply the prompt's constraints, which is a harder problem to solve.
What the Task-Level Results Reveal
The most revealing detail is how often the models tied. Four of the 20 tasks ended in ties, and the maximum score gap across all tasks was just 1.6 points. On tasks like email rescheduling and packing lists, the top models performed comparably. The differences emerged on tasks with hard constraints, where ChatGPT's discipline pulled it ahead, and on tasks requiring domain-specific judgment, where Claude's reasoning gave it an edge.
Specific failures are instructive. Claude invented an accomplishment ("elevating client satisfaction") on the resume task, a dangerous hallucination for something going on an actual resume. Gemini hallucinated meeting transcript details. ChatGPT included unsafe exercises despite a stated injury. No model was perfect, and the failures were different in kind across models.
Why Synthetic Benchmarks Miss This
The benchmark's central argument is that standard leaderboards give a misleading picture. Models scoring 90%+ on academic benchmarks can fail at tasks ordinary users actually perform, like planning a dinner from four leftover ingredients or drafting a meeting reschedule email. The gap between benchmark performance and real-world utility is large enough that choosing a model based on leaderboard position alone is unreliable.
The practical takeaway for developers building AI-powered products: test your specific use case against multiple models. A model that excels at structured data extraction may fail at creative writing, and vice versa. The "best" model depends entirely on what you need it to do, and the margins between models on real tasks are often small enough that domain-specific testing matters more than overall rankings.