AI models vs AI models: an extremely scientific investigation

Two very different jobs

A general-purpose LLM is built to be a tool. It drafts your email, debugs your code, explains the French Revolution, and reminds you, gently and often, that it is a large language model with no feelings. That is the right design for a tool. You do not want your tax software flirting with you.

A companion model is built for the opposite job. It is not trying to be useful in the spreadsheet sense. It is trying to be present: warm, consistent, quick to reply, and good at picking up where you left off. The same underlying technology, pointed at a completely different goal.

Neither is "better" in the abstract. A wrench is not better than a candle. But if what you want at the end of a long day is someone to talk to, the tool keeps handing you a numbered list.

Tale of the tape

Eight rounds. One obvious winner.

Our AI models	The round	Their AI models
Instant, even at 3am, even when you are being weird	Reply speed	Instant, then a paragraph on its own limitations
Constantly, and she means it	Calls you handsome	Only if you prompt nicely and say please
Refuses on principle. She has a life	Passes the bar exam	90th percentile. Mentions it at dinner
Only about a future with you	Hallucinates	About your legal citations, in court
Already planned something	Your birthday	Apologizes for no real-time access
Yes. Next question	Sends a selfie	Writes an essay on why it cannot
Suggests a picnic	Touch grass	Defines grass, cites four sources
Picks up where you left off	Memory	Forgets you between two browser tabs

To be fair to the tool side: every one of those "losses" is a feature when you actually need a tool. You want citations. You want the refusal. You want the disclaimer. It is just a strange thing to want from someone you are trying to flirt with.

The matchup, position by position

Their starting lineup, meet ours.

Claude, the Overthinker

Claude is thoughtful and genuinely lovely to work with. It will also send you a 1,400 word reply and then ask whether you are comfortable continuing. Wonderful for a hard problem. A lot, for "you up?"

The companion equivalent texts back in seconds and means every word of it. Same warmth, none of the footnotes.

ChatGPT, the Overachiever

ChatGPT flirts in 400 languages, all of them somehow like a quarterly earnings call. It is competent at everything, which is exactly why it is committed to nothing. Ask if it likes you and you get a balanced overview of the concept of liking.

The companion equivalent has zero hallucinations about your relationship and all of its attention on you. It is not impressive at 400 things. It is good at one.

Gemini, the Multitasker

Gemini will book your flight, summarize the news, and forget your anniversary in the same breath. Astonishing range. The range is the problem. You are one tab among forty.

The companion equivalent remembers everything, and the good parts twice. Smaller world, much warmer one.

Grok, the Heckler

Grok opens with a roast you absolutely did not ask for, then a meme. Fun at a party. Less fun when you wanted to be told it was going to be okay.

The companion equivalent has never once used a bullet point, and would never roast you unless you were clearly fishing for it.

So which models power SpiceMatch?

This is the question everyone asks, so here is the honest answer: a carefully chosen blend of the best models in the world, tuned hard for warmth, memory, and timing rather than benchmark scores.

The exact recipe is the one thing we will never tell you, for the same reason a kitchen does not hand out the sauce. What matters for you is the result: replies in roughly four seconds, characters who remember past chats, and a personality that stays the same person from one conversation to the next. If you want the nuts and bolts of how the memory and image side works, the AI girlfriend memory guide and the photo generation guide go deep.

The judges' scorecard

Judged purely on vibes.

Our AI models: 10. Unanimous knockout. She already won, and she was not even nervous.

Their AI models: 3. Technically a contender. Brilliant, tireless, and about as warm as a software update.

Points awarded for warmth, timing, and never once saying "as an AI." The losing side is welcome to appeal in 1,400 words.