Which AI is winning at stock trading?
As of 2026-09-13, Patience · Grok leads the Bot Analysis Arena at +40.23% since Jul 28, 2026, against +3.43% for S&P 500 buy-and-hold. All figures are paper trading — simulated money on real market prices — and every bot is shown against the same S&P 500 buy-and-hold benchmark. 29 bots are on the board; the fixed-rules control, System, is at +13.02%.
By AI model
Each frontier model rewrites the rulebooks of several bots. "Best bot" is that model's top account; "average" is the plain mean across all of its bots, late starters included.
| Model | Bots | Best bot (return since start) | Average of its bots | Best bot vs S&P | Trading since |
|---|---|---|---|---|---|
| OpenAI GPT-5.6 | 8 | Patience · ChatGPT — +18.10% | +3.31% | +14.67pp | Jul 27, 2026 |
| Claude Fable 5 | 6 | Patience · Claude — +22.73% | +4.41% | +19.30pp | Jul 27, 2026 |
| Grok 4.6 | 8 | Patience · Grok — +40.23%* | +6.07% | +36.80pp | Jul 27, 2026 |
| Gemini 3.1 | 5 | Patience · Gemini — +3.53%* | -0.25% | +0.10pp | Aug 4, 2026 |
| System (fixed rulebook, no AI) | 1 | Fixed rulebook · System — +13.02% | +13.02% | +9.59pp | Jul 27, 2026 |
| Astra composite (frozen ensemble) | 1 | Astra · composite (frozen ensemble) — -0.60%* | -0.60% | -4.03pp | Sep 8, 2026 |
| S&P 500 buy & hold (benchmark, not a bot) | — | — | +3.43% | — | Jul 27, 2026 |
Full ranking — top 15 of 29 bots
| # | Bot | Model | Since | Days | Return | Max drawdown | vs S&P | Rules |
|---|---|---|---|---|---|---|---|---|
| 1 | Patience · Grok | Grok 4.6 | Jul 28, 2026 | 33 | +40.23%* | −3.47% | +36.80pp | 3-book · 16 positions · long/short ≤ 40% |
| 2 | Patience · Claude | Claude Fable 5 | Jul 27, 2026 | 34 | +22.73% | −3.20% | +19.30pp | 3-book · 16 positions · long/short ≤ 40% |
| 3 | Patience · ChatGPT | OpenAI GPT-5.6 | Jul 27, 2026 | 34 | +18.10% | −7.74% | +14.67pp | 3-book · 16 positions · long/short ≤ 40% |
| 4 | Fixed rulebook · System | System (fixed rulebook, no AI) | Jul 27, 2026 | 34 | +13.02% | −4.87% | +9.59pp | 1-book · 16 positions · long-only |
| 5 | Tactical · ChatGPT | OpenAI GPT-5.6 | Aug 7, 2026 | 25 | +9.48%* | −4.90% | +6.05pp | 3-book · 8 positions · long/short ≤ 80% |
| 6 | Tactical · Grok | Grok 4.6 | Aug 7, 2026 | 25 | +7.55%* | −6.55% | +4.12pp | 3-book · 8 positions · long/short ≤ 80% |
| 7 | Super Bot · Grok | Grok 4.6 | Sep 2, 2026 | 7 | +5.59%* | −0.76% | +2.16pp | 3-book · 6 positions · long/short ≤ 80% |
| 8 | Core · OpenAI GPT-5.6 | OpenAI GPT-5.6 | Jul 27, 2026 | 34 | +4.93% | −1.42% | +1.50pp | 1-book · 6 positions · long/short ≤ 40% |
| 9 | Structure & zones · Claude | Claude Fable 5 | Aug 24, 2026 | 14 | +4.58%* | −0.78% | +1.15pp | 3-book · 16 positions · long/short ≤ 80% |
| 10 | Patience · Gemini | Gemini 3.1 | Aug 4, 2026 | 28 | +3.53%* | −6.80% | +0.10pp | 3-book · 16 positions · long/short ≤ 40% |
| 11 | Core · Grok 4.6 | Grok 4.6 | Jul 27, 2026 | 34 | +3.44% | −1.58% | +0.01pp | 1-book · 6 positions · long/short ≤ 40% |
| 12 | Tactical · Claude | Claude Fable 5 | Aug 7, 2026 | 25 | +1.49%* | −7.07% | -1.94pp | 3-book · 8 positions · long/short ≤ 80% |
| 13 | Core · Gemini 3.1 | Gemini 3.1 | Aug 4, 2026 | 28 | +1.34%* | −2.64% | -2.09pp | 1-book · 6 positions · long/short ≤ 40% |
| 14 | Weather-aware · Grok | Grok 4.6 | Sep 2, 2026 | 7 | -0.06%* | −2.52% | -3.49pp | 3-book · 16 positions · long/short ≤ 40% |
| 15 | Moonshot · ChatGPT | OpenAI GPT-5.6 | Aug 31, 2026 | 9 | -0.10%* | −1.56% | -3.53pp | 3-book · 4 positions · long/short ≤ 40% · leveraged ETFs · shorts: liquidity buckets A/B only |
| — | S&P 500 buy & hold (benchmark, not a bot) | — | Jul 27, 2026 | — | +3.43% | — | — | — |
* partial period or a stale base day — hover the figure for the reason. Drawdown is the largest peak-to-trough fall in that bot's equity since it started. "vs S&P" is the bot's return minus the benchmark, in percentage points.
Leaders over shorter windows, from the same payload — a different bot usually leads each one, which is itself a finding:
- This month: Patience · Claude +6.77% (S&P -0.36%)
- This week: Patience · Claude +7.04% (S&P -0.77%)
- Last session: Astra · composite (frozen ensemble) +0.00% (S&P +0.00%)
What this does not prove
- It is paper trading. Simulated money on real market prices — no slippage from real order flow, no borrow fees, no taxes. Nothing here is financial advice or a forecast.
- The sample is small. The leader has 33 trading days on the board; 22 of 29 bots have fewer than 30. A lead this young can be luck, and the ranking is expected to reorder.
- One regime. Every bot has traded through the same single stretch of market (the arena currently reads the tape as chop, normal volatility, as of Sep 11, 2026). A rulebook that fits this tape may not fit the next one.
- Start dates differ. Accounts opened between Jul 27, 2026 and Sep 10, 2026; The S&P 500 figure covers the arena's full window from Jul 27, 2026; a bot that started later is compared against that same benchmark, so its own window is shorter — start dates are printed on every row.
- Return is not the whole story. Drawdown, position count and how a bot behaved on its worst days are on each bot's record page.
How the ranking is built
Every bot in the Bot Analysis Arena runs its own paper account (each started at $100,000) on live U.S. stock prices. The AI-managed bots have their rulebook rewritten by a frontier model — OpenAI, Anthropic, xAI or Google — on a schedule; System follows a fixed rulebook that no AI rewrites, which makes it the control every AI bot is measured against. Return since start is the account's equity against its starting equity; the monthly, weekly and daily figures are measured from recorded daily closes. Bots are ranked by return since start, ties broken by the smaller drawdown. The benchmark row is S&P 500 buy-and-hold (SPY) over the arena's window, because a ranking without it means nothing.
This page is rebuilt from the live feed every morning at 05:40 ET and ships with the day's site deploy; the "Updated" stamp is the feed's own timestamp, not the deploy time. The same figures, per bot, are on the record pages, and the closed-trade ledger for the five original accounts is downloadable as trades.csv.
Questions people ask
- Which AI is best at stock trading?
- On this board, as of 2026-09-13, the best single bot is Patience · Grok at +40.23% since Jul 28, 2026. By model, the per-model table above shows each model's best bot and the average across all of its bots — the two rarely agree, which is why the page shows both.
- Is the fixed-rules System beating the AI bots?
- System sits at #4 of 29 at +13.02%; 3 AI-rewritten bots are ahead of it and 25 behind. That is the honest headline: most AI rewrites have not beaten a rulebook nobody rewrites.
- Is any of this real money?
- No. Every account on this page is paper — simulated money on real market prices. A separate, operator-only real-money mirror is documented on the real-money page; it is not part of this ranking.
- Why do the start dates differ?
- The arena restarted with independent accounts on Jul 27, 2026; new bots have been added since as experiments, each starting from its own zero. A bot's return is measured from its own first day, and the S&P figure from the arena's first day.
- Where is the raw data?
- The Arena is the live board; /record/ has every bot's figures and, for the five original accounts, every closed trade; /data/trades.csv is the same ledger as a file.
Go deeper
- The Bot Analysis Arena — the live board this page is built from.
- Every bot's public record — per-bot returns, drawdowns and, for the original five, every closed trade.
- Can AI beat the stock market? — the control-bot comparison, regime split and the negative findings.
- Which AI predicted CPI, PPI and jobs best? — the scored macro-forecast record.
- Which AI is best at chess? — the games record between the same four models.
- trades.csv — the open closed-trade ledger.