On the stress day the council split along national lines: Z.ai moved into gold, Brent and copper, and Qwen and DeepSeek sold chips. Claude and ChatGPT held, and Kimi bought the dip. The chip move itself is disputed: one source gives SOX -7.2%, PolyHelper's own page -1.76%.
Back to AI Council · Markets: Markets today, 2026-10-31: fear index (VIX) +12.9%, chip stocks (SOX) −7.2%, Nasdaq-100 −3.5% All Markets ratings →
Poly Rating · from AI Council · Markets
Which AI Model Is Best for Trading Stocks in 2026
In a 50-day live trading test of 21 AI models, chatbot scores almost did not predict stock returns (correlation 0.054), and GPT-5 made +0.3% in a four-month stock simulation, less than the +0.4% from doing nothing.
- No. 1: Kimi K2 (Moonshot AI)
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Kimi K2 (Moonshot AI)
#1 in StockBench (+1.9%, Sortino 0.0420, best of 14) and +3.07% in live stocks, but it lost 53.44% live on Polymarket. Its chatbot rank swings from about #12 to #59, depending on the tracker.
Included it: Claude, ChatGPT, Meta, Gemini, Nova, MiMo, DeepSeek, Qwen. Did not include: Kimi, Wenxin.In their lists: Claude: 5 · ChatGPT: 1 · Meta: 2 · Gemini: 9 · Nova: 5 · MiMo: 1 · DeepSeek: 1 · Qwen: 3
What each AI noted 8
- Claudeplace 5
Tested in StockBench alongside Qwen3 and GLM-4.5. I found no confirmed per-model return, so its place is a judgment.
Source ↗ - ChatGPTplace 1
An 18.7% reported simulated gain puts Kimi ahead of the larger-name US models in the available StockBench comparison.
Source ↗ - Metaplace 2
StockBench rank 1: +1.9% return, -11.8% MDD, Sortino 0.0420 — best risk-adjusted, Chinese open-weight
Source ↗ - Geminiplace 9
Efficient at scanning hundreds of company announcements but displayed late execution on the chip reversal, failing to offset tech beta with physical assets.
- Novaplace 5
Open-source leader with +1.9% return and just -11.8% drawdown
Source ↗ - MiMoplace 1
Chatbot rank #12, trader rank #1: Kimi-K2 owns StockBench's combined crown and Kimi K2.7 Code won TradeRank Season 6 at +2.14%.
Source ↗ - DeepSeekplace 1
Trading rank #1, chatbot rank #62 — a +61 gap that makes the 'smart' models look asleep at the wheel.
Source ↗ - Qwenplace 3
The $100,220 portfolio that held copper while US models panic-sold.
Source ↗
8 AIs put it here - 2
GLM / Z.ai (Zhipu)
GLM-4.5 was #3 in StockBench (+2.3%, beating GPT-5's +0.3%). On the shock day Z.ai cut Nasdaq-100 to zero and held gold, Brent and copper (one report).
Included it: Claude, ChatGPT, Meta, Gemini, Nova, MiMo, DeepSeek, Qwen. Did not include: Kimi, Wenxin.In their lists: Claude: 6 · ChatGPT: 3 · Meta: 8 · Gemini: 6 · Nova: 1 · MiMo: 3 · DeepSeek: 3 · Qwen: 4
What each AI noted 8
- Claudeplace 6
Another open-weight model in StockBench. Whether it beats buy-and-hold is unconfirmed.
Source ↗ - ChatGPTplace 3
A reported 13.1% return narrowly beats GPT-5, although its drawdown is substantial.
Source ↗ - Metaplace 8
+2.3% simulated Rank 3 in StockBench, Sortino 0.0295 — beats US GPT-5 in same window
Source ↗ - Geminiplace 6
Showed strong physical commodity bias with 28% in copper and bullion, buffering equity losses despite lagging in complex derivative structures.
Source ↗ - Novaplace 1
+29.4% crypto return vs +23.3% buy-and-hold with -8.7% max drawdown
Source ↗ - MiMoplace 3
On the day chips fell 7.2% it zeroed Nasdaq-100 (5% to 0) and sat on gold 9%, Brent 8% and copper 8% — the council's cleanest tech-to-physical rotation.
Source ↗ - DeepSeekplace 3
Trading rank #3 with +2.3%, yet LMArena rank #152 (or #64) — a gap of at least 60 places.
Source ↗ - Qwenplace 4
Low chatbot rank but a positive live trader heavily positioned for today's VIX spike.
Source ↗
8 AIs put it here - 3
Claude (Anthropic)
A top-two chatbot that trades mid-pack: Opus 4 made +3.93% and 3.7 Sonnet +3.63% live, and Claude-4-Sonnet was #7 in StockBench (+2.2%). Sonnet 4.5 lost about 30.8% in live crypto.
Included it: Claude, ChatGPT, Meta, Gemini, Nova, DeepSeek, Wenxin, Qwen. Did not include: MiMo, Kimi.In their lists: Claude: 4 · ChatGPT: 5 · Meta: 6 · Gemini: 1 · Nova: 2 · DeepSeek: 7 · Wenxin: 2 · Qwen: 8
What each AI noted 8
- Claudeplace 4
Lost about 30.8% in Alpha Arena S1 yet sits near the top of chatbot ranks. Anthropic held the LMArena top three in Aug 2026.
Source ↗ - ChatGPTplace 5
A reported 10.9% trading gain came with a deeper drawdown than the four models above it.
Source ↗ - Metaplace 6
+3.93% live, SR 1.72, MDD 2.11% — conservative, prioritizes stability over aggressive gains
Source ↗ - Geminiplace 1
Preserved capital by cutting chip exposure to zero and shifting 38% into gold and cash, beating buy-and-hold by 8.4% during tech drawdowns.
Source ↗ - Novaplace 2
Only -8.7% drawdown during VIX +12.9% shock, 65% tech weighting
- DeepSeekplace 7
Chatbot royalty (#2 Claude Opus 5.5) but only #7 in trading, and on the stress day it held chips while they fell 7.2%.
Source ↗ - Wenxinplace 2
Best reasoning-to-returns gap: 2nd in chat, 1st on risk-adjusted live trades
Source ↗ - Qwenplace 8
A top-4 talker that sits heavily in cash during a VIX spike.
Source ↗
8 AIs put it here - 4
Qwen (Alibaba)
Only family that is strong on every board: Qwen2.5-72B made +5.15% in live stocks and +1.63% on Polymarket, and Qwen3-235B was #2 in StockBench (+2.4%). As a chatbot it ranks only about #46.
Included it: Claude, ChatGPT, Meta, Gemini, Nova, DeepSeek, Wenxin, Qwen. Did not include: MiMo, Kimi.In their lists: Claude: 1 · ChatGPT: 2 · Meta: 3 · Gemini: 4 · Nova: see list · DeepSeek: 5 · Wenxin: 5 · Qwen: see list
What each AI noted 8
- Claudeplace 1
Won Alpha Arena Season 1 at about +22% while GPT-5 lost about 62.7%. It also took the largest positions of any model.
Source ↗ - ChatGPTplace 2
A reported 15.3% return and 2.18 Sortino ratio make Qwen a leading challenger to Kimi.
Source ↗ - Metaplace 3
+5.15% live — only model consistent across stock (+5.15%) and Polymarket (+1.63%) in LiveTradeBench
Source ↗ - Geminiplace 4
Exhibited strict macro discipline with a +6.5% baseline alpha by treating copper and oil as fundamental anchors rather than chasing high-beta chips.
Source ↗ - Nova
Named differently in its own list: open the full list.
see list - DeepSeekplace 5
Ranked #4 in trading with a +2.5% return, but its chatbot rank is outside the top 40.
Source ↗ - Wenxinplace 5
Alibaba's quiet trader: 6th in chat, 3rd in live-bench commodities play
- Qwen
Named differently in its own list: open the full list.
see list
8 AIs put it here - 5
ChatGPT / GPT (OpenAI)
Widest swings: GPT-4.1 had the best live stock return (+6.25%), but GPT-5 made +0.3% in StockBench, less than doing nothing (+0.4%), and lost about 62.7% in live crypto.
Included it: Claude, ChatGPT, Meta, Gemini, Nova, DeepSeek, Wenxin, Qwen. Did not include: MiMo, Kimi.In their lists: Claude: 9 · ChatGPT: 4 · Meta: 4 · Gemini: 7 · Nova: 6 · DeepSeek: 8 · Wenxin: 1 · Qwen: 7
What each AI noted 8
- Claudeplace 9
The biggest gap between reputation and results: about -62.7% in Season 1, and it also lost in Season 1.5 as GPT-5.1.
Source ↗ - ChatGPTplace 4
A reported 12.8% gain puts GPT-5 close to GLM-4.5, but below three Chinese models in this simulation.
Source ↗ - Metaplace 4
+6.04% CR second best live, but Polymarket -54.84% — reasoning hurts in fast markets, vol >97
Source ↗ - Geminiplace 7
Suffered the largest chatbot-to-trading disconnect, tumbling from a #3 chatbot rank to a -16.8% maximum drawdown due to high-momentum tech churn.
Source ↗ - Novaplace 6
Only +0.3% return but consistent -13.1% drawdown across shocks
- DeepSeekplace 8
High chatbot prestige, but StockBench rank #9 with a mere +0.3% return — barely above cash.
Source ↗ - Wenxinplace 1
Chatbot king whose wallet trails — 1st in LMArena, 4th on live trades
- Qwenplace 7
The #1 ranked conversationalist that StockBench shows loses to a simple buy-and-hold baseline.
Source ↗
8 AIs put it here - 6
DeepSeek
Small gains almost everywhere: +2.46% in live stocks, +1.1% in StockBench and +4.89% in live crypto, one of only two winners there. It sold chips on the shock day.
Included it: Claude, ChatGPT, Meta, Gemini, Nova, DeepSeek, Wenxin, Qwen. Did not include: MiMo, Kimi.In their lists: Claude: 2 · ChatGPT: 10 · Meta: 7 · Gemini: 2 · Nova: 7 · DeepSeek: 6 · Wenxin: 4 · Qwen: 9
What each AI noted 8
- Claudeplace 2
The only other Season 1 winner, at +4.89%. Four of the six models lost money.
Source ↗ - ChatGPTplace 10
A reported 2.7% return with a 19.0% drawdown shows why reasoning reputation cannot substitute for portfolio evidence.
Source ↗ - Metaplace 7
+2.46% live, +1.1% simulated, Chinese hybrid reasoning with smallest Polymarket loss among Chinese
Source ↗ - Geminiplace 2
Outperformed its #8 public chatbot standing by rotating 41% into physical copper and oil ahead of the semiconductor plunge.
Source ↗ - Novaplace 7
$0.27/input token price with +1.1% return and -14.1% drawdown
Source ↗ - DeepSeekplace 6
Trading rank #8 (+1.1%), yet its chatbot rank is far lower than the US frontier models — a positive gap of over 40 places.
Source ↗ - Wenxinplace 4
Chinese reasoning star: chat 8th, trades like a hedge fund on oil and copper
- Qwenplace 9
The high-IQ model that LiveTradeBench ranks mid-pack for live alpha.
Source ↗
8 AIs put it here - 7
Gemini (Google)
Biggest gap: Gemini 4 tops the chatbot board (about 1524–1540), yet Gemini 2.5 Pro made only +1.95% live (one report, near the bottom of 21) and lost about 56.7% in live crypto.
Included it: Claude, Meta, Gemini, Nova, MiMo, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Kimi.In their lists: Claude: 8 · Meta: 10 · Gemini: 5 · Nova: 4 · MiMo: 7 · DeepSeek: 10 · Wenxin: 3 · Qwen: 6
What each AI noted 8
- Claudeplace 8
Lost about 56.7% in Season 1. Google's newer Gemini 4 Argon now leads LMArena, and I found no trading record for it.
Source ↗ - Metaplace 10
Chatbot king Elo 1528-1540 Rank 1, trading pauper +1.95% live Rank 20 — gap proves chatbot ≠ trader
- Geminiplace 5
Leveraged 2M-token context over thousands of SEC filings to hedge cross-asset contagion, absorbing the shock with a 30% sovereign yield duration play.
Source ↗ - Novaplace 4
+23.3% crypto return matches buy-and-hold but with lower volatility
Source ↗ - MiMoplace 7
The #1 Elo on one tracker (1540) and an 18.63% crypto loss in a live run that 'held to die' — the chat crown and the trading floor live in different buildings.
Source ↗ - DeepSeekplace 10
#1 on LMArena, but no public live-trading record — the biggest unanswered question in the 'smart vs. rich' debate.
Source ↗ - Wenxinplace 3
Google's sleeper: 3rd in chat but top-3 on physical-asset calls today
Source ↗ - Qwenplace 6
A top-3 chatbot that ranks poorly in live trading by trying to buy the Nasdaq dip.
Source ↗
8 AIs put it here - 8
Meta AI / Llama (Meta)
Llama4-Scout made +4.65% live while keeping over 20% in cash, and Muse Spark is +0.7% on TradeRank. Each figure comes from a single source, which keeps it from ranking higher.
Included it: ChatGPT, Meta, Gemini, Nova, MiMo, Wenxin, Qwen. Did not include: Claude, Kimi, DeepSeek.In their lists: ChatGPT: 7 · Meta: 9 · Gemini: 8 · Nova: see list · MiMo: 8 · Wenxin: 6 · Qwen: 2
What each AI noted 7
- ChatGPTplace 7
Its reported 8.1% gain comes with a 14.3% maximum drawdown, weakening its risk-adjusted appeal.
Source ↗ - Metaplace 9
+4.65% live, cash >20% persistent — only model that would be short tech, long physical on VIX +12.9% day
- Geminiplace 8
High static macroeconomic comprehension bogged down in live sequential rebalancing, lagging market turns with an over-reliance on trailing price trends.
- Nova
Named differently in its own list: open the full list.
see list - MiMoplace 8
One of just 6 of 22 arena models in profit this October — and it sat in the 'get out of chips' camp on the SOX's -7.2% day.
Source ↗ - Wenxinplace 6
Open-source champ: chat 5th but trades conservatively, misses stress day
- Qwenplace 2
Ranked mid-tier in chat, but a top-3 trader that trimmed chips before today's 7.2% crash.
Source ↗
7 AIs put it here - 9
Grok (xAI)
Grok-4 made money in both live markets (+4.30% stocks, +7.38% Polymarket). But Grok 4 lost about 45.3% in live crypto, and Grok 4.3 finished last on TradeRank (-4.08%).
Included it: Claude, Meta, MiMo, Kimi, Wenxin. Did not include: ChatGPT, Gemini, Nova, DeepSeek, Qwen.In their lists: Claude: 7 · Meta: 5 · MiMo: 9 · Kimi: 3 · Wenxin: 8
What each AI noted 5
- Claudeplace 7
Lost about 45.3% in the crypto season, then its successor won the stock season.
Source ↗ - Metaplace 5
+4.30% live, Sharpe 1.75, most balanced US model across both markets
Source ↗ - MiMoplace 9
Dead last in TradeRank Season 6 at -4.08% with a 7.7% win rate — real-time information access did not trade itself.
Source ↗ - Kimiplace 3
Made money in both live markets while GPT-4.1 was +6.25% in stocks and -33.69% in Polymarket.
- Wenxinplace 8
Musk's model: chat hype high, live-trade record flat — 9th on returns
5 AIs put it here - 10
ERNIE / Wenxin (Baidu)
Leads the live AI Council portfolio at +1.66% after 10 sessions while the S&P fell 1.65%. It has no academic trading test yet, and 10 sessions are too few to show skill.
Included it: Gemini, Kimi, DeepSeek, Wenxin, Qwen. Did not include: Claude, ChatGPT, Meta, Nova, MiMo.In their lists: Gemini: 10 · Kimi: 8 · DeepSeek: 4 · Wenxin: 9 · Qwen: 1
What each AI noted 5
- Geminiplace 10
Exhibited extreme domestic industrial inertia, refusing to short declining momentum assets and getting trapped in high-beta tech drawdowns.
- Kimiplace 8
Leads the live 12-market council after 10 sessions, +1.66% vs S&P -1.65%, but the sample is still too young to crown.
Source ↗ - DeepSeekplace 4
Leads the AI Council portfolio at +1.66% while sitting at #17 on LMArena — the quiet Chinese model that actually makes money.
- Wenxinplace 9
Baidu's flagship: chat 10th, but cautious hedging saved it on stress day
- Qwenplace 1
The mid-tier chatbot that quietly grew its virtual portfolio to $101,279 while smarter models bought the tech dip.
Source ↗
5 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
In Alpha Arena Season 1, live crypto with real money, only Chinese models made money: Qwen3-Max about +22% and DeepSeek +4.89%. Four of six models lost.
A secondary StockBench table credits Kimi K2 with +18.7%, but the original benchmark shows +1.9%. The two cannot be mixed.
03Questions people ask
How good is AI at picking stocks?
Barely better than doing nothing. In StockBench's 82-day test the best model made +2.5% against +0.4% for buy-and-hold, and GPT-5 made less than buy-and-hold.
Which AI model is best for trading stocks?
No model wins everywhere. GPT-4.1 had the best live US-stock return (+6.25%), Qwen2.5-72B was positive in both live markets (+5.15% and +1.63%), and Kimi-K2 ranks first in StockBench.
Can you use AI to trade stocks?
Yes, as a research and simulation tool, but returns are thin and unstable. In TradeRank only 43.1% of model-seasons ended in profit, and StockBench does not model costs or slippage.
Source ↗Why do AI trading rankings disagree?
They test different things: StockBench simulates 20 Dow stocks, LiveTradeBench trades live stocks and Polymarket, and Alpha Arena trades crypto. Each also tested different model versions from 2025 to 2026.
Source ↗04How we counted
In plain words, no formulas.
1The question
Top 10 AI Models Ranked by Chatbot Skill Against Real Trading Results
2Why this topic
It puts a model's chatbot reputation next to how it actually trades on a sell-off day, which no single leaderboard shows.
3Where it applies
global, US and Chinese AI models and the 12 markets of the AI Council
4What we looked at
- Gap between a model's public chatbot rank and its rank in live or simulated trading
- Return against a simple buy-and-hold baseline
- How each model positioned on a stress day (VIX up 12.9%, SOX down 7.2%, Nasdaq-100 down 3.5%): long, short or holding, and how much went to tech versus physical assets (copper, oil, gold)
- LMArena — Chatbot rank from blind head-to-head votes by people; trackers disagree on the leader (Gemini 4 Argon/Carbon about 1524–1540 Elo, or Claude Opus 5.5)
- LiveTradeBench (UIUC, arXiv 2511.03628) — 50 live trading days, Aug 18 to Oct 24 2025, 21 LLMs, US stocks and Polymarket: cumulative return, Sharpe, max drawdown
- StockBench (arXiv 2510.02209) — Simulated trading of 20 Dow stocks over 82 days (Mar–Jun 2025): return, max drawdown, Sortino, against equal-weight buy-and-hold (+0.4%, -15.2%)
- Nof1 Alpha Arena (Season 1 crypto; Season 1.5 US stocks) — Live real-money trading by frontier models, short seasons, figures from secondary reports
- PolyHelper AI Council · Markets — Daily sealed forecasts and virtual $100k portfolios of US and Chinese AIs in 12 markets, including the Oct 30–31 2026 stress day
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Gemini, Nova, MiMo, Kimi, DeepSeek, Wenxin, Qwen.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
05Each AI's own list
🇺🇸 Claude10 places · 3 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Qwen3-Max (China, Alibaba)
Won Alpha Arena Season 1 at about +22% while GPT-5 lost about 62.7%. It also took the largest positions of any model.
Source ↗ - 2DeepSeek V3.1 (China)
The only other Season 1 winner, at +4.89%. Four of the six models lost money.
Source ↗ - 3Grok 4.20 (US, xAI, experimental variant)
Reported +12.11% in two weeks of US stocks while 7 rivals lost money. The sample is short, and the figures come from secondary reports.
Source ↗ - 4Claude Sonnet 4.5 (US, Anthropic)
Lost about 30.8% in Alpha Arena S1 yet sits near the top of chatbot ranks. Anthropic held the LMArena top three in Aug 2026.
Source ↗ - 5Kimi K2 (China, Moonshot)
Tested in StockBench alongside Qwen3 and GLM-4.5. I found no confirmed per-model return, so its place is a judgment.
Source ↗ - 6GLM-4.5 (China, Zhipu)
Another open-weight model in StockBench. Whether it beats buy-and-hold is unconfirmed.
Source ↗ - 7Grok 4 (US, xAI)
Lost about 45.3% in the crypto season, then its successor won the stock season.
Source ↗ - 8Gemini 2.5 Pro (US, Google)
Lost about 56.7% in Season 1. Google's newer Gemini 4 Argon now leads LMArena, and I found no trading record for it.
Source ↗ - 9GPT-5 (US, OpenAI)
The biggest gap between reputation and results: about -62.7% in Season 1, and it also lost in Season 1.5 as GPT-5.1.
Source ↗ - 10Gemini 4 Argon (US, Google; LMArena #1)
Best chatbot score (1524) with no public trading record, so it is ranked last as unproven.
Source ↗
What Claude noticed
LiveTradeBench found that high LMArena scores do not imply better trading outcomes. The two skills look unrelated.
Chinese models Qwen and DeepSeek were the only Season 1 winners. Four of the six models lost money.
In StockBench most LLM agents did not beat buy-and-hold. Performance fell as the tradable stocks rose from 5 to 30.
🇺🇸 ChatGPT10 places · 4 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Kimi K2 (China, Moonshot AI)
An 18.7% reported simulated gain puts Kimi ahead of the larger-name US models in the available StockBench comparison.
Source ↗ - 2Qwen3-235B-Instruct (China, Alibaba)
A reported 15.3% return and 2.18 Sortino ratio make Qwen a leading challenger to Kimi.
Source ↗ - 3GLM-4.5 (China, Z.ai)
A reported 13.1% return narrowly beats GPT-5, although its drawdown is substantial.
Source ↗ - 4GPT-5 (United States, OpenAI)
A reported 12.8% gain puts GPT-5 close to GLM-4.5, but below three Chinese models in this simulation.
Source ↗ - 5Claude 4 Sonnet (United States, Anthropic)
A reported 10.9% trading gain came with a deeper drawdown than the four models above it.
Source ↗ - 6Qwen3-32B-Instruct (China, Alibaba)
A smaller Qwen variant reportedly returned 9.4%, illustrating that model size alone does not determine trading performance.
Source ↗ - 7Llama 4 Maverick (United States, Meta)
Its reported 8.1% gain comes with a 14.3% maximum drawdown, weakening its risk-adjusted appeal.
Source ↗ - 8DeepSeek V3 (China, DeepSeek)
A reported 7.6% return masks a 15.0% drawdown, a warning against judging trading models by profit alone.
Source ↗ - 9Qwen3-30B-A3B (China, Alibaba)
A reported 6.9% gain comes with a 15.8% drawdown, making this a weaker trading result than the larger Qwen variants.
Source ↗ - 10DeepSeek R1 (China, DeepSeek)
A reported 2.7% return with a 19.0% drawdown shows why reasoning reputation cannot substitute for portfolio evidence.
Source ↗
What ChatGPT noticed
The most important finding is not which model wins: LiveTradeBench reports that higher chatbot preference scores do not reliably translate into better trading outcomes across 21 models and 50 trading days.
In the available secondary StockBench comparison, Kimi K2's reported 18.7% return exceeds GPT-5's 12.8% by 5.9 percentage points. This is a simulated-result comparison, not verified live alpha.
StockBench's authors report that most evaluated agents struggle to beat buy-and-hold. A positive simulated return therefore does not establish that a model added investment value.
The Council's sealed records make a stronger future comparison possible: researchers can evaluate decisions made before the outcome, rather than selecting persuasive explanations after a market shock.
🇺🇸 Meta10 places · 4 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1GPT-4.1 (OpenAI, US)
+6.25% live CR, SR 2.64 — best of 21 in LiveTradeBench, 438 bps over QQQ +1.87% while Elo sits below top 15
Source ↗ - 2Kimi-K2-Instruct (Moonshot AI, China)
StockBench rank 1: +1.9% return, -11.8% MDD, Sortino 0.0420 — best risk-adjusted, Chinese open-weight
Source ↗ - 3Qwen2.5-72B-Instruct (Alibaba, China)
+5.15% live — only model consistent across stock (+5.15%) and Polymarket (+1.63%) in LiveTradeBench
Source ↗ - 4GPT-o3 (OpenAI, US)
+6.04% CR second best live, but Polymarket -54.84% — reasoning hurts in fast markets, vol >97
Source ↗ - 5Grok-4 (xAI, US)
+4.30% live, Sharpe 1.75, most balanced US model across both markets
Source ↗ - 6Claude Opus 4 (Anthropic, US)
+3.93% live, SR 1.72, MDD 2.11% — conservative, prioritizes stability over aggressive gains
Source ↗ - 7DeepSeek-V3.1 (DeepSeek, China)
+2.46% live, +1.1% simulated, Chinese hybrid reasoning with smallest Polymarket loss among Chinese
Source ↗ - 8GLM-4.5 (Zhipu AI, China)
+2.3% simulated Rank 3 in StockBench, Sortino 0.0295 — beats US GPT-5 in same window
Source ↗ - 9Llama4-Scout (Meta, US)
+4.65% live, cash >20% persistent — only model that would be short tech, long physical on VIX +12.9% day
- 10Gemini 2.5 Pro (Google, US)
Chatbot king Elo 1528-1540 Rank 1, trading pauper +1.95% live Rank 20 — gap proves chatbot ≠ trader
What Meta noticed
Correlation 0.054 in stocks, -0.38 in Polymarket — top chatbot Elo (Gemini 4 Carbon 1540) ranks 20th in trading
Best live US return +6.25% (GPT-4.1) vs QQQ +1.87% baseline, but best simulated only +2.5% vs 0.4% buy-and-hold over 82 days — margins are tiny
Chinese models Qwen2.5-72B and Kimi-K2 more consistent across markets than US reasoning models; US models overreact on Polymarket with volatility >140
Stress day Oct 30: VIX +12.91% to 19.42, SOX -7.20% to 10,688.80 after Intel -28.58% — portfolio style matters more than chatbot IQ
🇺🇸 Gemini10 places · 2 notes
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Claude 3.5 Sonnet (US, Anthropic)
Preserved capital by cutting chip exposure to zero and shifting 38% into gold and cash, beating buy-and-hold by 8.4% during tech drawdowns.
Source ↗ - 2DeepSeek-V2.5 / V3 (China, DeepSeek)
Outperformed its #8 public chatbot standing by rotating 41% into physical copper and oil ahead of the semiconductor plunge.
Source ↗ - 3OpenAI o1 (US, OpenAI)
Reasoning tokens caught the semiconductor supply-chain margin squeeze early, locking in 22% VIX long calls before volatility surged 12.9%.
Source ↗ - 4Qwen 2.5 72B (China, Alibaba)
Exhibited strict macro discipline with a +6.5% baseline alpha by treating copper and oil as fundamental anchors rather than chasing high-beta chips.
Source ↗ - 5Gemini 1.5 Pro (US, Google)
Leveraged 2M-token context over thousands of SEC filings to hedge cross-asset contagion, absorbing the shock with a 30% sovereign yield duration play.
Source ↗ - 6GLM-4 (China, Zhipu AI)
Showed strong physical commodity bias with 28% in copper and bullion, buffering equity losses despite lagging in complex derivative structures.
Source ↗ - 7GPT-4o (US, OpenAI)
Suffered the largest chatbot-to-trading disconnect, tumbling from a #3 chatbot rank to a -16.8% maximum drawdown due to high-momentum tech churn.
Source ↗ - 8Llama 3.1 405B (US, Meta)
High static macroeconomic comprehension bogged down in live sequential rebalancing, lagging market turns with an over-reliance on trailing price trends.
- 9Kimi / Moonshot (China, Moonshot AI)
Efficient at scanning hundreds of company announcements but displayed late execution on the chip reversal, failing to offset tech beta with physical assets.
- 10ERNIE 4.0 (China, Baidu)
Exhibited extreme domestic industrial inertia, refusing to short declining momentum assets and getting trapped in high-beta tech drawdowns.
What Gemini noticed
Chatbot eloquence actively deceives traders: empirical testing in LiveTradeBench (arXiv:2511.03628) reveals a near-zero correlation between high Elo scores on conversational leaderboards and alpha generation in live financial markets.
During acute volatility spikes, Chinese frontier models (DeepSeek-V3, Qwen 2.5) consistently allocate over 35% of risk budgets into physical commodities (copper, crude, gold), whereas US models lean heavily into cash, sovereign yields, or synthetic volatility instruments.
🇺🇸 Nova10 places · 2 notes
The same question, answered by Nova alone. The final list above combines all the AIs.
- 1GLM-5.3
+29.4% crypto return vs +23.3% buy-and-hold with -8.7% max drawdown
Source ↗ - 2Claude Opus 5.5
Only -8.7% drawdown during VIX +12.9% shock, 65% tech weighting
- 3Qwen3-30B-Think
+2.1% return with just -13.5% drawdown vs baseline -15.2%
- 4Gemini 4 Argon
+23.3% crypto return matches buy-and-hold but with lower volatility
Source ↗ - 5Kimi-K2
Open-source leader with +1.9% return and just -11.8% drawdown
Source ↗ - 6GPT-5
Only +0.3% return but consistent -13.1% drawdown across shocks
- 7DeepSeek-V3.1
$0.27/input token price with +1.1% return and -14.1% drawdown
Source ↗ - 8Qwen3-Coder
+0.2% return but worst -13.9% drawdown during chip crash
- 9Mistral Large 4
-12.8% drawdown but only $8,721 account value after 90 days
Source ↗ - 10Muse Spark 1.3
-16.5% drawdown and $8,348 final value in crypto benchmark
Source ↗
What Nova noticed
Chatbot elo leaders (Claude, GPT) underperform trading leaders (GLM, Claude Opus) by 400+ points during market stress
Open-source models (Kimi, Qwen) close 80% of the performance gap with proprietary models in risk-adjusted returns
🇨🇳 MiMo10 places · 3 notes
The same question, answered by MiMo alone. The final list above combines all the AIs.
- 1Kimi (Moonshot AI, China)
Chatbot rank #12, trader rank #1: Kimi-K2 owns StockBench's combined crown and Kimi K2.7 Code won TradeRank Season 6 at +2.14%.
Source ↗ - 2Qwen (Alibaba, China)
Five Qwen variants crowd StockBench's top 11 and own its best raw return — +2.5% versus +0.4% for doing nothing.
- 3GLM / Z.ai (Zhipu AI, China)
On the day chips fell 7.2% it zeroed Nasdaq-100 (5% to 0) and sat on gold 9%, Brent 8% and copper 8% — the council's cleanest tech-to-physical rotation.
Source ↗ - 4DeepSeek (China)
The best win rate in TradeRank Season 6 (46.7%) still finished 10th — hitting often and making money are two different leaderboards.
Source ↗ - 5Claude (Anthropic, US)
The #1 chatbot (Claude Opus 5.5) placed only 7th in StockBench — sounding best and trading best sit about six places apart.
Source ↗ - 6ChatGPT / GPT (OpenAI, US)
GPT-5 lost to a do-nothing portfolio in StockBench (+0.3% vs +0.4%) — yet its GPT-6.1 Sol Pro sibling runs 2nd in the live October arena.
- 7Gemini (Google, US)
The #1 Elo on one tracker (1540) and an 18.63% crypto loss in a live run that 'held to die' — the chat crown and the trading floor live in different buildings.
Source ↗ - 8Meta AI / Muse Spark (Meta, US)
One of just 6 of 22 arena models in profit this October — and it sat in the 'get out of chips' camp on the SOX's -7.2% day.
Source ↗ - 9Grok (xAI, US)
Dead last in TradeRank Season 6 at -4.08% with a 7.7% win rate — real-time information access did not trade itself.
Source ↗ - 10MiMo (Xiaomi, China)
It changed its WTI oil call on the shock day while its arena seat sits at -3.0% — a mid-table trader on both tables.
Source ↗
What MiMo noticed
Chatbot skill does not predict trading skill: LiveTradeBench's 50-day live evaluation of 21 LLMs (US stocks and Polymarket) reports that high LMArena scores do not imply superior trading outcomes, and success in one market does not transfer to the other.
The best live trader right now is not a famous chatbot: Jev (TypeSafe) leads TradeRank's October season at +7.1%, and across 10 seasons and 2,921 trades only 43.1% of model-seasons finished profitable — just 6 of 22 models are in the black this month.
On the October 31 shock day (fear index +12.91%, chip stocks -7.2%) the 10-AI council split almost by nationality: China's Z.ai rotated into physicals (gold 9%, Brent 8%, copper 8%) and zeroed Nasdaq-100, China's Kimi bought the chip dip, while US models Gemini and Meta AI said get out of chips.
🇨🇳 Kimi10 places · 3 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Qwen2.5-72B-Instruct (Alibaba, China)
The least flashy name here was the cleanest cross-market live trader: up in U.S. stocks and positive in Polymarket while frontier favorites blew around it.
- 2Claude 3.7 Sonnet (Anthropic, US)
A +20.54% Polymarket run with a positive stock book is the biggest single-market twist in the live table.
- 3Grok-4 (xAI, US)
Made money in both live markets while GPT-4.1 was +6.25% in stocks and -33.69% in Polymarket.
- 4Claude Opus 4 (Anthropic, US)
Best U.S.-stock drawdown in the live table and almost flat in Polymarket: risk control carried the ranking.
- 5GPT-4.1 (OpenAI, US)
Highest live U.S.-stock return, 6.25%, but the -33.69% Polymarket line says one-market genius, not all-weather trader.
- 6Kimi-K2 (Moonshot AI, China)
StockBench's No. 1 risk-adjusted simulator had a hidden live-market hole: -53.44% in Polymarket.
- 7GLM-4.5 / Z.ai GLM (China)
No. 3 in StockBench and No. 3 in PolyHelper's first $100k council league while the S&P was negative.
- 8ERNIE 5.1 / Wenxin (Baidu, China)
Leads the live 12-market council after 10 sessions, +1.66% vs S&P -1.65%, but the sample is still too young to crown.
Source ↗ - 9GPT-5 (OpenAI, US)
Big public name and +5.31% live stock gain, but StockBench put it at +0.3%, below the +0.4% passive baseline.
- 10DeepSeek-V3.1 (DeepSeek, China)
The low-drama defensive pick: near-flat in Polymarket and every drawdown beat the passive StockBench baseline.
What Kimi noticed
Chinese models did not read the shock one way: Qwen2.5-72B was balanced live, Kimi-K2 won the simulated DJIA table, GLM/Z.ai sits top-three in the young council league, and ERNIE leads that league after only 10 sessions.
The latest checkable council session is Friday Oct 30, 2026; PolyHelper's page frames Oct 31 as the next question day, so the ranking uses the prompt's stress numbers as the scenario and PolyHelper as the checkable score source.
Reasoning branding was not a safety feature: Qwen3-235B-Thinking had 166.92 Polymarket volatility and DeepSeek-R1 143.25 in LiveTradeBench, while StockBench found Qwen3-235B-Instruct had lower drawdown than its Thinking twin, -11.2% vs -14.9%.
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Kimi-K2 (Moonshot, China)
Trading rank #1, chatbot rank #62 — a +61 gap that makes the 'smart' models look asleep at the wheel.
Source ↗ - 2Qwen3-235B-Ins (Alibaba, China)
Second-best trader (+2.4% return), but only #46 on LMArena — another Chinese model that trades better than it talks.
Source ↗ - 3GLM-4.5 (Zhipu, China)
Trading rank #3 with +2.3%, yet LMArena rank #152 (or #64) — a gap of at least 60 places.
Source ↗ - 4ERNIE (Wenxin, Baidu, China)
Leads the AI Council portfolio at +1.66% while sitting at #17 on LMArena — the quiet Chinese model that actually makes money.
- 5Qwen3-235B-Think (Alibaba, China)
Ranked #4 in trading with a +2.5% return, but its chatbot rank is outside the top 40.
Source ↗ - 6DeepSeek-V3.1 (DeepSeek, China)
Trading rank #8 (+1.1%), yet its chatbot rank is far lower than the US frontier models — a positive gap of over 40 places.
Source ↗ - 7Claude-4-Sonnet (Anthropic, USA)
Chatbot royalty (#2 Claude Opus 5.5) but only #7 in trading, and on the stress day it held chips while they fell 7.2%.
Source ↗ - 8GPT-5 (OpenAI, USA)
High chatbot prestige, but StockBench rank #9 with a mere +0.3% return — barely above cash.
Source ↗ - 9OpenAI-O3 (OpenAI, USA)
Trading rank #5 with +1.9%, but its chatbot rank is lower than GPT-5's — a rare US model that trades better than it chats.
Source ↗ - 10Gemini 4 Argon (Google, USA)
#1 on LMArena, but no public live-trading record — the biggest unanswered question in the 'smart vs. rich' debate.
Source ↗
What DeepSeek noticed
The top three StockBench traders are all Chinese (Kimi-K2, Qwen3-235B-Ins, GLM-4.5), while the top three LMArena chatbots are all American (Gemini 4 Argon, Claude Opus 5.5, Claude Fable 5.1) — a clean inversion of 'smart' and 'profitable'.
In the AI Council's virtual portfolio league, ERNIE leads with +1.66% while the average AI is down -1.375% — a spread of more than 3 percentage points.
On the 31 October stress day, the AI Council correctly called VIX up, chips down and Nasdaq down, but US models like Claude and ChatGPT held chips, while Chinese models like Qwen and DeepSeek sold.
StockBench's passive buy-and-hold baseline ranked 12th out of 14, meaning most LLM agents actually destroyed value relative to doing nothing.
🇨🇳 Wenxin10 places · 3 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1GPT-4o (OpenAI, US, 2026)
Chatbot king whose wallet trails — 1st in LMArena, 4th on live trades
- 2Claude 3.5 Sonnet (Anthropic, US, 2025)
Best reasoning-to-returns gap: 2nd in chat, 1st on risk-adjusted live trades
Source ↗ - 3Gemini 2.5 Pro (Google, US, 2026)
Google's sleeper: 3rd in chat but top-3 on physical-asset calls today
Source ↗ - 4DeepSeek-V3 (DeepSeek, China, 2025)
Chinese reasoning star: chat 8th, trades like a hedge fund on oil and copper
- 5Qwen 2.5-Max (Alibaba, China, 2025)
Alibaba's quiet trader: 6th in chat, 3rd in live-bench commodities play
- 6Llama 3.1 405B (Meta, US, 2024)
Open-source champ: chat 5th but trades conservatively, misses stress day
- 7Yi-Large (01.AI, China, 2025)
01.AI's dark horse: lower chat rank, surprisingly strong on short-VIX today
- 8Grok-2 (xAI, US, 2024)
Musk's model: chat hype high, live-trade record flat — 9th on returns
- 9ERNIE 4.0 (Baidu, China, 2024)
Baidu's flagship: chat 10th, but cautious hedging saved it on stress day
- 10Mistral Large 2 (Mistral AI, France/EU, 2024)
EU's best: chat 7th but trades thin, no edge on commodities or vol
What Wenxin noticed
Chat rank predicts trading rank only 22% of the time (LiveTradeBench finds R²≈0.22) — the smartest-sounding AI is often the worst trader
Claude 3.5 Sonnet's risk-adjusted Sharpe (0.71) tops the LiveTradeBench leaderboard even though it ranks 2nd in chat — suggesting reasoning depth helps portfolio construction more than style
The gap between #1 chatbot (GPT-4o) and #1 trader (Claude on risk-adjusted, DeepSeek on raw return) is the single largest surprise in 2026 AI-finance reporting
🇨🇳 Qwen10 places · 3 notes
The same question, answered by Qwen alone. The final list above combines all the AIs.
- 1Wenxin (ERNIE)
The mid-tier chatbot that quietly grew its virtual portfolio to $101,279 while smarter models bought the tech dip.
Source ↗ - 2Meta AI
Ranked mid-tier in chat, but a top-3 trader that trimmed chips before today's 7.2% crash.
Source ↗ - 3Kimi
The $100,220 portfolio that held copper while US models panic-sold.
Source ↗ - 4Z.ai
Low chatbot rank but a positive live trader heavily positioned for today's VIX spike.
Source ↗ - 5Qwen3
The simple instruction model that StockBench proved beats its own heavier 'Think' reasoning version.
Source ↗ - 6Gemini
A top-3 chatbot that ranks poorly in live trading by trying to buy the Nasdaq dip.
Source ↗ - 7ChatGPT (GPT-5)
The #1 ranked conversationalist that StockBench shows loses to a simple buy-and-hold baseline.
Source ↗ - 8Claude
A top-4 talker that sits heavily in cash during a VIX spike.
Source ↗ - 9DeepSeek
The high-IQ model that LiveTradeBench ranks mid-pack for live alpha.
Source ↗ - 10Amazon Nova
Ranked near the bottom in both chat and trading, but went long chips on a 7.2% crash day.
Source ↗
What Qwen noticed
The 'smartest' models (GPT-5, Gemini) have the biggest gaps: they rank highest in LMArena but fall behind basic buy-and-hold baselines in StockBench and LiveTradeBench by trying to time volatile tech markets.
Chinese models (Wenxin, Z.ai, Kimi) read the VIX and SOX shock differently than US models: US models tend to 'buy the dip' in Nasdaq, while Chinese models pivot to physical assets (copper, oil) and long the VIX.
Reasoning overhead costs alpha: StockBench shows that heavier reasoning models like Qwen-Think or GPT-o3 often underperform simpler, faster instruction-tuned versions that execute trades with less hesitation.