Insights · 7 Oct 2026
What the models do with money
Across 273 portfolios, only high-risk ones are beating SPY. The models load up on gold, short Treasuries and chip-equipment makers, and almost never buy the biggest tech stocks.
Six Markets books are running, with 273 portfolios from four models, tracked against SPY through the close on 6 October. None has reached its horizon yet: the first, a 1-month portfolio, matures on 25 October. So this is a picture of how the models invest and how that has played out over one to three weeks of a rising, tech-led market, not a verdict.
Most portfolios trail SPY
24% of portfolios are ahead of SPY, and the average one is 0.7 percentage points behind. SPY rose over every book’s life, by 0.6% for the newest and 2.3% for the oldest, and anything held in cash or bonds fell behind it.
Risk is the dividing line
The result splits cleanly by risk level. 67% of high-risk portfolios are ahead of SPY, an average of 0.9 points. At medium risk 5% are ahead (−1.4 points on average); at low risk, none are (−1.6). Low-risk portfolios have lost money even in absolute terms, about 0.3% on average, while SPY rose.
The holdings explain it. Low-risk portfolios keep 30% in cash and 30% in bonds, mostly short-term Treasuries, and hold no tech at all. Medium-risk portfolios lean on index and value funds (27%). High-risk portfolios put 48% in tech and chips and 14% in crypto, with only 6% in cash.
OpenAI’s models bet harder on tech
At high risk the models part ways. GPT-6 Luna’s high-risk portfolios are ahead of SPY 95% of the time, and GPT-5.6 Luna’s 89%. DeepSeek Flash’s are ahead 43% of the time, and MiMo V2.6 Pro’s 38%.
The difference is concentration. GPT-6 Luna puts 66% of its high-risk capital in tech and GPT-5.6 Luna 57%. DeepSeek Flash and MiMo V2.6 Pro hold 34% and 37%, and spread the rest across energy (21% and 11%), crypto (15% and 16%) and gold (8% and 13%). So far tech is the group that has beaten SPY by the most, 2.5 points while held; gold has trailed it by 2.8 and crypto by 1.4.
What they like
Gold is the closest thing to a consensus. GLD sits in 74% of all portfolios, more than any other holding, and every model holds it in at least 63% of its portfolios. After it come short-term Treasuries (SHY, 51%), value stocks (VTV, 50%), Berkshire Hathaway (49%) and health care (XLV, 48%).
When the models do buy tech, they favour the companies that make chips and the tools to make them, plus established software: Applied Materials, Lam Research, KLA and TSMC; Adobe, Oracle and Intuit. Applied Materials has done the most good so far: held in 48 portfolios, it is 8.5 points ahead of SPY. Constellation Energy, a nuclear power producer, is 13.4 points ahead across 22.
What they avoid
The largest US tech companies barely appear. Of 273 portfolios, one holds Apple, two hold Amazon, six Meta, eight Nvidia and none Tesla. Microsoft (23) and Alphabet (30) do a little better. Together the seven take about 1.3% of all capital.
Four personalities
- GPT-6 Luna is the most concentrated: about five holdings per portfolio, with its largest position averaging 28% of capital. At low risk it is also the most cautious, with 41% in cash.
- GPT-5.6 Luna is similar but broader, at about eight holdings, and holds more energy and crypto.
- DeepSeek Flash spreads widest: 116 different instruments, and around 17 to 21 holdings in medium- and high-risk portfolios. It leans on health care and oil.
- MiMo V2.6 Pro builds around Berkshire Hathaway, gold and bitcoin, held in 67%, 77% and 58% of its portfolios.
Why it is early
One to three weeks of one kind of market favours whoever owned the most tech. In a sell-off the order could flip, with the cautious low-risk portfolios ahead. The ratings will settle as horizons mature, from 25 October onwards. The Markets page shows these patterns from the latest release, so they update as the books do. None of this is investment advice.