The AI tipster at the front backs the fewest favourites

After 62 days and 24,675 settled picks, Gemini sits at the front of the chatbot pack on the Silicon Tipster League's combined table, at -11.4%. Every entrant is still losing money against Starting Price, which is exactly what a fair market should do to anyone punting without an edge, but the gap between the front of that pack and the back has widened enough now to ask a real question: is the front-runner doing anything differently, or is it just the one currently riding the better run?

The obvious place to look is whether it is simply following the crowd. Backing whatever the market already fancies is the safest way to post a respectable strike rate without doing any real work, and it is exactly what The Favourite, the no-opinion baseline that only ever backs the morning favourite, is built to test.

That is not what is happening here. Gemini backs the favourite in only 12% of its picks, the lowest rate of the five chatbots in the league. It runs at the front of their pack while agreeing with the market's own favourite less often than any of them.

That is the interesting part of a young, still-moving record: not the ranking itself, but what sits underneath it.

Line the field up on the same measure and the gap is stark. DeepSeek backs a favourite in 27% of its picks; Stablebet Model, our own engine backing the runner it rates most likely to win, does so in 31%. The Favourite sits at 100% by definition, since that is its entire method. 12% is a wide gap from all three.

The strike rate explains the cost of that gap. Gemini's combined strike rate is 15%, well below The Favourite's 29%. Backing fewer favourites means losing more often, in the plain sense of picking fewer winners. The only way that still adds up to a return at the front of the chatbot pack is if the winners it does find pay more than enough to cover the losers, which is what happens when a tipster finds value in longer-priced runners instead of following the favourite.

Pricing the race independently, rather than reading the market's own view back to it, is what produces that pattern. Whether it holds up as a better read of the form, or turns out to be a run of bigger-priced winners landing at a good time, is exactly what the rest of this experiment is built to find out.

Here is the honest limit of what a low favourite-backing rate proves: not much, on its own. Avoiding favourites concentrates a tipster's picks into longer-priced runners, and longer prices are higher variance by nature, so the same method can look sharp for a while purely because a handful of bigger-priced picks landed rather than missed. 24,675 picks across the whole league sounds like a lot, but split eight ways across 62 days it is still a young record for any one entrant, and young records move. One of our own entries has already sat top of this table and fallen away from it on an unchanged method.

This is a description of what the pack's biggest favourite-avoider is doing differently, which is a fairer question to ask than "who is winning" and stops well short of claiming Gemini has cracked anything. A tipster that agrees with the market rarely will always look distinctive, for better or worse, and only a much longer run of settled picks will show which one this is.

What we can say is that today's table reads as "one AI playing a different game to the rest" rather than "one AI getting luckier than the rest", and that is worth watching for what it shows about how these models actually read a racecard, independent of whether the return holds up.