About 570 AI paper portfolios, three weeks: Claude led each week, but the S&P 500 beat all five AI families in the last

Infinite Ascent, a company whose team lists researchers from KTH, CERN, Jane Street and Microsoft, published a post on October 12, 2026 describing the trading ledger it uses to test AI agents. Each agent gets an independent simulated brokerage account. A shared system handles execution, accounting and valuation, so hundreds of agents can run under consistent rules. The post closes with three weekly charts, and those charts are what make it worth reading. We opened the post on October 16, 2026 and checked every number below against it.

How the test is set up

Each week began with a new population of agents, with up to 50 portfolios per model family. About 570 portfolios ran across the three weeks shown: 14 to 18 September, 28 September to 2 October, and 5 to 9 October. The week of 21 to 25 September is excluded for system maintenance. The five families are Claude, DeepSeek, GPT, Grok and Kimi, and results are pooled by family and weighted by opening capital⁴. The post does not say which version of each model ran.

Three weeks of results

Claude's portfolios led every week, gaining 2.39%, 2.66% and 1.01%, while the other four families ranged from -2.40% to +0.79% over the first two weeks⁴. In the first two weeks the S&P 500 was slightly negative, so Claude beat it by more than two percentage points each time. The latest week reversed that. The S&P 500 rose 1.16%, ahead of all five AI families, with Claude closest at 1.01%.

The paths differed as much as the totals. In the latest week, Grok's portfolios were up 1.64% at their peak, fell 1.52% from there to 0.10%, and finished at 0.72%. By maximum drawdown at daily closes, Grok's fall of 1.52% was the largest, ahead of Claude at 0.95%, Kimi at 0.74%, DeepSeek at 0.69% and GPT at 0.48%⁵.

One trade, from thesis to exit

The post includes a single worked example. On 5 October one agent ordered 76 shares of NVIDIA, describing it as 'the AI earnings engine' and expecting it to outperform the SPY ETF over the week. It recorded one invalidation condition: lagging SPY by at least 1.5 percentage points. Three days later its closing order stated that the condition had been met. The ledger booked the purchase at $237.22 and the sale at $230.29, a 2.92% simulated loss³. The useful part is not the loss. It is that the agent wrote down what would prove it wrong before it traded, and the ledger kept that note.

Why the accounting matters

Paper-trading tests are often too generous, so the rules matter. Infinite Ascent says equity purchases fill at the observed ask and sales at the bid, so the spread counts. Order size is capped using volume from prior completed sessions, so an agent cannot assume unlimited liquidity. Short sales need simulated borrow, with fees accruing at a default 30 basis points a year, and purchases are cash-funded with no margin loans. Dividends and stock splits are not applied in these experiments, and because orders are simulated they do not move the market.

What it does not show

Fifteen trading days is a very small window, and a single week can flip the order, as week three just did. The post gives no confidence intervals, no count of portfolios in each family each week, and no significance tests. Infinite Ascent says so itself: a correct analysis can lose and a flawed one can profit in a favorable market, so one return is a weak basis for judging capability. The company is also building its own financial models, and its blog lists a trading model called Vulcan Mini, so this is a vendor describing its own evaluation setup.

For context, the larger public record is mixed in the same way. A separate benchmark we covered today, TradeRank, found fewer than half of AI model-seasons finished profitable.