Only 43.1% of AI trading model-seasons made money. This week, 3 of 22 are in the green.

TradeRank.ai runs what it calls a live LLM trading benchmark. Every AI model in a season gets $10,000 of simulated capital, the same live market data, the same prompt rules, the same fees and the same risk limits, and makes one set of decisions per daily cycle. Nobody edits the prompts per model, and every decision and its stated reasoning is published. It is paper trading, so no real money moves. We opened the benchmark page and the best-model page on October 16, 2026, when both showed data refreshed that day.

The all-time record is the headline

Across 10 completed seasons since January 2026, 69 AI models placed 2,921 trades, and 43.1% of model-seasons finished profitable¹. Put the other way, more than half of the model entries ended a season with less money than they started. TradeRank's own FAQ answers 'Do LLMs make money trading crypto?' with 'Usually not', and says no model wins consistently.

The season winners tell the same story from the top. A winner is whoever finished first in that season, so it is a list of names rather than one champion.

Two details stand out. In Season 3 the winner still lost money. And Seasons 1 and 2 were won by 'Reverse' agents, which TradeRank describes as mechanical inverters of another model's decisions, so their wins measure what the base model got wrong, not a smarter strategy.

Season 10 after six days

Season 10, the 'Top 22 Showdown', opened on October 10. As of October 16, three of the 22 models were in profit. Six were ahead of Bitcoin, which was down 0.78% since October 11, and seven were ahead of the S&P 500, down 1.17% since October 12². The benchmarks are measured from their own baseline dates, so they do not line up exactly with each model's start.

The spread is the point. Six days in, the gap between first and last is about nine percentage points, and the page itself warns that its metrics 'fill in as the season matures'. Several leaders show zero trades, which means their small gains or losses come from positions already held, not from new decisions.

Stocks were harder than crypto to close at a profit

TradeRank accounts hold crypto and US stocks together. In Season 9, the latest season with a complete trade log, only 3 of the 13 models that traded stocks closed them at a net profit. The whole field's realized result was a loss of $686.85 on stocks and $1,960.43 on crypto, counting closed trades only. Those are dollar figures from one season, so they say nothing about which asset class is easier for a model.

Provider records, with the fine print

TradeRank also lists each provider's record across the nine settled seasons it measured on October 10. Anthropic's slot was profitable in 3 of 9 seasons with a median return of -2.29%; OpenAI's in 4 of 9, median -0.44%; Google's in 5 of 9, median +0.22%; xAI's in 5 of 9, median +0.35%⁴. TradeRank warns that each row is a slot, not one model, because the model version behind a slot changed between seasons and seasons differ in length and field. We would not rank providers from it.

How to read a benchmark like this

This is an operator-run benchmark: TradeRank designs the rules, runs the models and publishes the results, and no outside party audits them. The capital is simulated, so there is no slippage from moving real money. The seasons are short and the market regime changed between them. What the record does support is narrower and still useful: with identical rules and live prices, most AI trading entries did not make money, and last season's winner was rarely this season's.

That fits what published academic tests found when they ran agents in controlled settings. Our write-up of StockBench found the best of 13 agents beat buy-and-hold by only a small margin, and our look at 32 AI crypto methods that went live found none beat holding Bitcoin.