Same Prompt, Four AI Models: How Different Are the Answers?

The short answer Given the same request, the four models differed in two places: how big a task they hand you on day one, and how they repair the plan after you miss a day.

One model starts you at 5 words. Another starts at 20. The gap is not a few characters. It is the amount of effort you owe tomorrow.

We sent one prompt to Claude Sonnet 5, GPT-6.1 Sol, DeepSeek V4 Pro and Qwen3.8 Max through the LMU AI API. Each model answered twice (the prompt, then a follow-up), which gave us 8 valid answers.

We did not pick a hard problem. We wanted to see how they handle something ordinary.

I want to start learning English again. I can study 30 minutes a day for one week. My level is average and I tend to give up halfway. Make me a 7-day plan that states the content and time split for each day, and give advice for what to do when I don't finish a day's task. Keep the plan simple. Answer in under 500 Chinese characters. Don't ask me questions.

GPT-6.1 Sol: the smallest first step Day 1 is 5 minutes reading 5 common words, 10 minutes listening to a short self-introduction clip, 10 minutes of shadowing (repeating after the audio), and 5 minutes saying 3 sentences about your name and hobbies. It also suggests using a single textbook or a single app, so you spend less time choosing.

For a day with no time, it offers a minimum version: 2 minutes of listening plus 3 minutes of shadowing. The whole answer is built around restoring the habit and leaves room for bad days.

Claude Sonnet 5: a denser vocabulary block Day 1 is 15 minutes memorizing 20 high-frequency words, 5 minutes writing them from memory, and 10 minutes making sentences with 5 of them. That is a full loop of memorize, check, use.

Day 2 adds 20 new words, and Day 4 reviews the previous 40. It also says not to double up after a missed day. For someone with an average foundation, though, it is worth testing whether you can finish this vocabulary load in the time.

DeepSeek V4 Pro: the half hour split into small pieces Day 1 is 10 minutes learning 15 high-frequency words, 10 minutes making sentences with 5 of them, 5 minutes reading a self-introduction aloud, and 5 minutes of self-testing. Later days cover listening, sentence patterns, reading, and shopping and ordering phrases. The tasks are concrete.

After a missed day, it suggests making up only the core content in 10 minutes the next day. The first answer does not say which part of the original plan those 10 minutes replace, so you would have to ask.

Qwen3.8 Max: an even rhythm Every day is three 10-minute blocks. Day 1: learn 20 high-frequency words, listen to and read a short passage, then copy out and write 3 sentences. Day 2: review the old words, learn new ones, keep shadowing the same passage, then record yourself saying your sentences. It is easy to follow as written.

It gives a 15-minute minimum version. But "make up half the next day" leaves open what to make up and how to fit it into 30 minutes.

All four plans fit 30 minutes a day, yet what they put inside those minutes is different. Read them as the plan you would actually run tomorrow, and the differences get concrete.

If I have to work overtime on Wednesday and finish none of that day's plan, please adjust Thursday through Sunday. I can still study at most 30 minutes a day, and don't extend into next week. Answer in under 300 Chinese characters.

Model Day-1 vocabulary task How it repaired the plan after the missed Wednesday Claude Sonnet 5 Memorize 20 words Moved Wednesday's core practice to Thursday and pushed the rest back. Compressed and merged the weekend tasks, and dropped the subtitle-comparison part of the original plan. GPT-6.1 Sol Read 5 words Folded Wednesday's listening into Thursday, replayed a short dialogue on Friday, and kept the weekend writing and review.

DeepSeek V4 Pro Learn 15 words Folded the missed routine sentence patterns into Thursday, then used those patterns for a weekly summary on Sunday. Qwen3.8 Max Learn 20 words Cut 10 verb phrases to 8 and 5 dictation sentences to 3, squeezed them into Thursday, and rescheduled the days after. All four stayed within 30 minutes a day, and all four finished on Sunday.

They made different trade-offs: pushing back, merging, cutting, and repeating to consolidate. Each looks reasonable on paper, but each feels different to carry out.

When you actually use one, other things matter more:

Which tasks suit you. Whether you can pick the plan back up after missing a day. Whether the model kept the original constraints when it adjusted the plan. So take a question you know well and try it yourself. Add one real constraint and see how the answer changes:

Planning a trip: "the person traveling with me can't walk far." Planning a week of dinners: "I have no time to cook on Wednesday." Organizing your work: "one task just got moved up." Don't rush to a long, complicated prompt. Add one constraint first, then watch how the model revises its answer. That tells you more about fit than asking "which one is best?"

Run on October 30, 2026, through LMU AI's international endpoint (api.lmuai.ai). Claude, DeepSeek and Qwen were called through the Anthropic-style /v1/messages interface. GPT-6.1 Sol was called through the OpenAI-style /v1/responses interface with reasoning effort set to low. One run per model. Sampling parameters were left at each provider's defaults and not unified. No web search or tools were enabled. Model identity is the label returned by the gateway and was not independently verified. One extra attempt with gpt-5.5 returned a 404 upstream error. It was excluded from the comparison. We compared only the content of these answers. The results say nothing about overall model quality, speed or cost.

Still, they show a reason to look at the concrete plan when you choose a model. A plan that sounds thorough may be one you never follow. A simpler answer might be the one that gets you started tomorrow.

If you want to try this yourself, you can look at the available models and how to connect through LMU AI. Availability and pricing are as shown in the console.

Ask the same question one more time. And ask yourself what kind of help you actually need.