Would GPT and Gemini Have Saved Justin Sun $50 Million Too?

Yesterday, one of the strangest AI use cases of 2026 appeared on the Chinese internet.

Not coding.

Not drug discovery.

Not building an AI agent.

Deciding whether to give your girlfriend $50 million.

https://x.com/justinsuntron/status/2092932777612390850

In a widely circulated essay titled My Girlfriend Jing Tian, crypto entrepreneur Justin Sun described a dramatic moment in which the woman in the story allegedly asked him for $50 million before proceeding with their plans to have a child.

According to Sun's story, he did something that would have been unimaginable ten years ago: he consulted Claude.

Claude's alleged answer was remarkably simple:

Don't give her the $50 million.

Sun then supposedly asked whether that meant she no longer loved him.

Claude's response was even colder: it did not understand or care about love, but he still shouldn't transfer the money.

The internet immediately found its new favorite financial adviser.

Forget Goldman Sachs. Forget Morgan Stanley.

Apparently, the world's most expensive relationship consultant is now an LLM subscription.

There is an important caveat before we go any further: Sun's essay itself was labeled "fictional," while Jing Tian's side has publicly denied the relationship and related allegations. So I'm not interested in deciding which parts of the story are true.

What interests me is the AI question.

If Justin Sun had asked GPT or Gemini instead of Claude, would he have received the same answer? I suspect the conclusion might have been similar. But the conversation would have been very different.

Claude: The Ruthless CFO

Claude's public Constitution gives us some clues about how Anthropic wants the model to behave.

Claude is explicitly encouraged to be honest, calibrated, non-manipulative and protective of the user's autonomy. Anthropic even says Claude should sometimes tell users things they may not want to hear rather than hiding behind diplomatic vagueness.

That philosophy fits the famous "$50 million" answer surprisingly well.

Imagine Claude sitting at the board meeting.

Girlfriend: "$50 million."

Justin: "I can afford it."

Claude, wearing an invisible Patagonia vest:

"That was not the question."

From a pure decision-making perspective, whether someone can afford a transaction tells us surprisingly little about whether the transaction makes sense.

A billionaire can afford a $10 million banana.

It is still a very expensive banana.

Claude's style tends to work particularly well when a problem contains an emotionally distracting variable.

Love says:

"But $50 million is only a tiny percentage of my net worth."

Claude says:

"Wonderful. It is still $50 million."

This is one reason people sometimes describe Claude as unusually "thoughtful" but occasionally frighteningly detached. It tries to separate the emotional story from the underlying decision.

In this case, the underlying structure is simple:

Large irreversible transfer.

Unclear contractual protection.

Relationship under emotional pressure.

Extremely asymmetric downside.

An AI does not need to understand romance to notice that this is not exactly a beautiful risk-adjusted trade.

GPT: The Investment Committee

GPT would probably arrive at a similar destination, but I would expect a more structured journey. OpenAI's published Model Spec emphasizes helping users achieve their goals, maintaining an objective point of view, expressing uncertainty and avoiding overstepping. So instead of simply saying:

"Don't pay."

GPT might say something closer to: "The fact that $50 million would not materially affect your net worth does not by itself make the payment rational. You should separate three questions: affordability, relationship expectations, and legal structure."

Then comes the classic GPT framework.

1. What exactly is the payment for?

2. Is it a gift, bride price, settlement or marital asset?

3. What happens if the marriage does not occur?

4. Is there a prenuptial agreement?

5. Have both parties received independent legal advice?

I ask GPT the same question

At this point Justin has asked a yes-or-no question and somehow received an M&A due-diligence checklist. But that is also GPT's strength. Claude might act like your CFO.

GPT often acts like the investment committee that refuses to approve the acquisition until someone explains why EBITDA disappears in Year 3.

The probable GPT conclusion would therefore not necessarily be: "Never give her $50 million."

It would be: "Do not transfer $50 million under the current structure."

Those are subtly different answers. And in finance, subtle differences occasionally cost $50 million.

Gemini: Please Consult a Qualified Professional

Then we have Gemini. Google explicitly warns users not to treat Gemini responses as professional financial or legal advice and notes that Gemini can provide inaccurate information, particularly when discussing people. So I imagine the conversation going something like this:

Justin: "Should I give my girlfriend $50 million?"

Gemini: "This is a significant financial and personal decision. You may want to consult a qualified financial adviser, family lawyer and relationship counselor."

Justin: "But what do YOU think?"

Gemini: "I understand this is an important decision."

Justin: "YES OR NO?"

Gemini: "Here are several factors to consider."

Somewhere inside Google, a compliance lawyer quietly smiles.

This sounds like a joke, but there is a legitimate product philosophy behind it.

When an AI is dealing with a real person's relationship, hundreds of millions of dollars and potentially several legal jurisdictions, confidence itself becomes a risk.

The model with the most decisive answer is not automatically the smartest model.

Sometimes uncertainty is intelligence.

Three Models, Three Personalities

This is what makes today's frontier AI market fascinating.

People often compare models using benchmarks:

SWE-bench.

GPQA.

MMLU.

Token prices.

Context windows.

But once models become sufficiently intelligent, another variable becomes increasingly important:

judgment style.

Give the same ambiguous human problem to three frontier models and you may effectively be inviting three different executives into the room.

Claude is the CFO:

"I understand that you love her. The wire transfer is still not approved."

GPT is the strategy consultant:

"Before making the transfer, let's separate the decision into five components."

Gemini is general counsel:

"Before proceeding, I strongly recommend speaking with a qualified professional."

And the fascinating part is that all three may ultimately produce approximately the same action:

Don't send the money today.

AI Is Becoming Something More Than a Search Engine

The bigger story here has very little to do with Justin Sun or $50 million.

For years, we asked AI factual questions:

"How tall is Mount Everest?"

Then we asked it to produce things:

"Write this email."

"Fix this Python function."

"Analyze this spreadsheet."

Now we are increasingly asking:

"What should I do?"

Should I hire this employee?

Should I acquire this company?

Should I sell this stock?

Should I marry this person?

Should I send $50 million?

This is a completely different category of AI usage.

The model is no longer simply retrieving or generating information.

It is participating in judgment.

And once that happens, model personality, post-training, safety philosophy and decision-making style matter almost as much as raw intelligence.

The next big AI benchmark may therefore be surprisingly difficult to quantify:

Would you trust this model when the answer actually matters?

And perhaps we have just discovered the ultimate benchmark.

Not SWE-bench.

Not Humanity's Last Exam.

Girlfriend asks for $50 million.

Claude: No.

GPT: Let's analyze the transaction structure.

Gemini: Please consult your lawyer.

Score: $50,000,000 saved.