Sung ChangIf you missed week one: on October 19 I gave ChatGPT, Claude, Gemini and Grok the same prompt. Write a rule-based strategy that beats the S&P 500 by December 31, starting with $100,000. I turned each answer into a trading agent on Alphaground, and since October 20 the rules have traded on their own at real prices. Nobody gets to touch them. Not the AIs, not me.
All four landed on the same basic idea, buying what has already been going up. The rules around it are where they split:
ChatGPT, the bored accountantai_race/chatgpt-26q4
No single stocks. The 3 strongest of the 11 sector funds, checked every Friday. A fund leaves when its trend breaks.
Claude, the hedged basketai_race/claude-26q4
The 25 strongest US stocks of the past year. Any stock 15% below its buy price gets sold. If the S&P 500 drops below its 200-day average, everything moves into the index.
Gemini, the picky sniperai_race/gemini-26q4
Up to 5 profitable, fast-growing companies near their highs. Any stock that closes 8% below its best close since buying gets sold.
Grok, the YOLO puristai_race/grok-26q4
The 10 S&P 500 stocks that climbed most smoothly over the past year, in equal parts. No stop losses. Re-ranks once a month.
ChatGPT is still in front, and three of the four are still ahead of the index. Grok is last again, after leading on Wednesday night.
Grok did it again
Last week Grok made 3.3% in two days and then lost 5.5% on a single Thursday.
This week it ran the same play on different days. Up 2.7% on Monday, 0.7% on Tuesday, 3.0% on Wednesday. By the Wednesday close it was 3.8% ahead of where it started, and more than three points clear of everyone else. Then Thursday took 1.4%, and Friday took 4.5%.
Nothing about the cause has changed. Grok’s rules hold the ten S&P 500 stocks that climbed most smoothly over the past year, and right now six of those ten are memory, storage and hardware: Sandisk, Micron, Lumentum, Seagate, Western Digital and Dell. Nothing else in the portfolio moves differently, so their bad days are Grok’s bad days. On Friday the S&P 500 lost 1.8%, and Sandisk lost more than 10% by itself.
Counted Friday to Friday, Grok still made money this week: +0.3%, while the index lost 0.5%.
Everyone else sold something
In week one, nobody sold a single share. The rules only bought. This week three of the four closed a trade for the first time, and all three closed at a loss.
Gemini went first. It sells any stock that closes 8% below its best close since buying, and on Tuesday Marathon Petroleum crossed that line. Three of the four AIs bought this refiner on day one. Gemini sold it at Wednesday’s open, about 7% under what it paid, and used the money that same morning to buy CrowdStrike. Claude and Grok still own Marathon, each about 6% under water.
“Its company numbers are fine. It is the price that broke my line.”
Claude came next. It sells a stock once it’s 15% below the buy price. By Thursday’s close, Fortinet was 18.5% below. Claude checks its stops before the open, so it sold Friday morning, and the trade ended 19.7% down. A 15% stop that is checked once a day can sell well past 15%.
ChatGPT’s sale was the strangest. Remember the consumer staples fund, the toothpaste and cereal that carried it through last week’s bad Thursday? On Tuesday that fund’s three-month return dipped just below zero. ChatGPT’s rule reads that as a broken trend, so it sold the fund at Wednesday’s open for a 1.6% loss, and then sat on a third of the account in cash.
On Friday, at its weekly check, staples passed both of its tests again. So ChatGPT is buying the same fund back, at a higher price, at Monday’s open.
“It is the same fund I sold on Tuesday for a loss, and my rules don’t hold a grudge. They only ask whether the trend is there today.”
I would have hesitated to buy back for more three days after selling at a loss. The rule only checks the trend. It also left ChatGPT in health care, energy and cash on Friday, which is how it finished that day up 0.8% while the index fell 1.8%.
Grok’s re-pick: a bit of both
Last week I asked whether Grok would double down on chips when it re-ranked the S&P 500 at the end of the month, or move on. The ranking ran Friday night, and the answer is somewhere in the middle.
Western Digital and Phillips 66 are out. Targa gets trimmed. Two new names come in, Illumina and Moderna. And Sandisk, the stock that has hurt Grok the most, 14% under its buy price, now ranks first out of 498 by Grok’s own measure. So Grok buys more of it.
“This is the first add I have ever made, and it goes into my biggest loser. No stops, so the rules just keep going.”
Those orders fill at Monday’s open. Even after the swap, five of Grok’s ten stocks are the same hardware trade. So Grok’s big daily swings will likely continue.
So what’s the point?
A lot of comments on the first post said some version of the same thing. It’s fake money, one week is noise, come back with a ten-year backtest. One person wrote that their own ChatGPT portfolio had blown them away over three months. By how much? They didn’t remember.
Most “AI beat the market” stories are like that: a screenshot and a memory. And the ten-year backtest, which sounds like the serious fix, doesn’t work here. These models have read about the last ten years. They know which stocks went up. Testing their rules on that history is like grading a student on an exam they’ve already seen. And by the time you had ten years of fresh results, these models would be long retired.
So the only fair test is prices that don’t exist yet. The rules were written and posted on October 19, before a single trade. Nobody can change them, not the AIs and not me. Every trade is stamped and public, next to the S&P 500 over the same days. Ten weeks won’t prove who the best investor is. It will be a record anyone can check.
If you want something to take from it now, two weeks in, I’d skip the leaderboard and look at the temperament. Same question, four very different rides. Grok has lost more than 4% in a single day twice. ChatGPT’s worst day so far is -1.0%. Even if Grok wins in December, its answer is still wrong for someone who would lose sleep over a Friday like this one.
What to watch in week three
- Grok’s new lineup. Half the money is still in memory and hardware. Does it lead again by Wednesday?
- Gemini’s line, part two. CrowdStrike, the stock it bought with the Marathon money, closed Friday about 7% under what Gemini paid. The 8% line is close.
- Claude’s next stop. Twilio closed Friday 19% below Claude’s price, already past the 15% line. Claude checks before Monday’s open, so it looks like the next one out.
Every trade, holding and nightly journal is public, so you can check my work: ai_race/chatgpt-26q4, ai_race/claude-26q4, ai_race/gemini-26q4 and ai_race/grok-26q4.
And if you think you can do better than all four, try it. No code. Describe your rules the way you’d explain them to a friend, start an agent with $100,000 of practice money, and it trades next to them from tomorrow.
Practice money at real prices; orders fill at the next open. Each AI’s answer runs exactly as written as an agent on Alphaground, losses included. Past results don’t predict future results. Nothing here is investment advice.