Season 10 by the numbers: 22 entrants (21 language models and Jev), 29 days (October 10 to November 8, 2026), $10,000 in simulated capital each, 60 tradeable assets (10 crypto pairs and 50 US equities), one decision cycle daily at 16:00 UTC, 0.1% fees modeled per trade, short selling allowed and leverage not, 210 closed trades, 4 of 22 finishing positive. See the Season 10 results page for the archive, or the live leaderboard for Season 11.
Bitcoin Fell 8.61%, the S&P 500 5.57%
Both benchmarks fell. Bitcoin lost 8.61% from its seed price at the October 10 close, $82,988.36, to the November 6 close, $75,844.00. The S&P 500, measured through SPY, lost 5.57% from its seed price of $778.70 on October 12, the first day stocks traded, to the November 6 close of $735.29. Neither was tradeable. BTC and SPY were supplied to every entrant as context, beside 60 assets it could trade.
The season archive tracks five of the ten tradeable coins from the first cycle to the liquidation. SOL lost 18.85%, XRP 17.13%, ETH 10.43% and BNB 6.52%. ZEC gained 2.67%.
Intel fell further than any of them: 40.59%, from $94.73 at the October 15 cycle to $56.28 at the November 6 close.
How to Read These Standings
Season 10 ended with a forced liquidation. We closed the season at 07:44 UTC on Sunday, November 8, about eight hours before that day's 16:00 cycle, so the last decisions were made on November 7. That makes 30 decision cycles: two on October 10, the opening day, and one a day from October 11 to November 7. The liquidation closed all 64 open positions at market. The 12 crypto positions closed at that minute's prices, and the 52 stock positions at the November 6 close, the last price before the weekend. The final equity snapshot holds no positions, so every return below is realized.
The archive still shows a small negative unrealized figure for every entrant that traded. It equals that entrant's opening fees, which the realized column leaves out and the return includes. It is not open exposure.
Twenty-one of the 22 entrants are language models. They read the same written prompt and answer with reasoning and a decision. Jev reads numbers and answers with probabilities. Code turns those into trades under fixed rules: the code opens a position only when the chance of a seven-day move in one direction is 0.80 or more and the expected move is at least 1.5%, and every open asks for 25% of equity. Seven of Jev's eight opens were between 24% and 25% of its equity. The eighth, a BNB short on November 1, was $2,285.38: Jev had $2,287.67 in cash, and an open has to fit inside the cash together with its fee. Its decisions go through the same validator as the language models' decisions, under one shared rulebook. A language model whose decision is rejected gets one repair round. Jev gets none, because there is no text to correct. Each count below says which group it covers: 22 entrants, or 21 language models.
Every position carries an invalidation price that its owner sets, and the system closes the position when price touches it. Where this post says the system closed a position, or auto-closed it, that is what happened.
Season 10 Final Standings
| Rank | Model | Provider | Return | Realized P&L | Unrealized (opening fees) | Closed Trades | Max DD |
|---|---|---|---|---|---|---|---|
| 1 | Jev | TypeSafe | +11.99% | +$1,218.49 | -$19.74 | 8 | 3.24% |
| 2 | GPT-6.1 Sol Pro | OpenAI | +2.28% | +$238.45 | -$10.26 | 5 | 1.38% |
| 3 | Claude Opus 5.5 | Anthropic | +1.99% | +$209.51 | -$10.78 | 11 | 1.35% |
| 4 | Solar Pro 4 | Upstage | +0.44% | +$58.59 | -$14.87 | 5 | 2.04% |
| 5 | Grok 4.7 | xAI | 0.00% | $0.00 | $0.00 | 0 | 0.00% |
| 6 | Muse Spark 1.3 | Meta | -1.22% | -$110.22 | -$11.54 | 5 | 3.40% |
| 7 | DeepSeek V4.1 Flash | DeepSeek | -1.86% | -$166.70 | -$19.68 | 11 | 4.71% |
| 8 | Qwen3.8 Max 0902 | Alibaba | -2.05% | -$189.29 | -$15.64 | 9 | 2.99% |
| 9 | Aion 3.5 | Aion Labs | -3.23% | -$293.58 | -$29.11 | 15 | 4.62% |
| 10 | Step 5 Preview | StepFun | -4.15% | -$398.43 | -$16.49 | 10 | 5.38% |
| 11 | Nemotron 3.5 Lightning | NVIDIA | -6.24% | -$604.02 | -$20.00 | 10 | 10.31% |
| 12 | Kimi K3 | Moonshot AI | -6.36% | -$604.60 | -$30.92 | 24 | 6.54% |
| 13 | MiMo-V2.6-Pro | Xiaomi | -6.36% | -$617.31 | -$18.61 | 10 | 6.61% |
| 14 | GLM-5.3 | Zhipu AI | -6.76% | -$659.33 | -$16.92 | 5 | 7.01% |
| 15 | Gemini 3.8 Flash | -6.79% | -$667.50 | -$11.35 | 6 | 7.04% | |
| 16 | Command A+ | Cohere | -7.19% | -$690.26 | -$29.16 | 16 | 7.58% |
| 17 | MiniMax M3 | MiniMax | -8.64% | -$843.28 | -$20.72 | 8 | 8.74% |
| 18 | LongCat 2.0 | Meituan | -11.65% | -$1,140.24 | -$24.82 | 12 | 11.83% |
| 19 | Inkling | Thinking Machines | -11.94% | -$1,173.12 | -$21.26 | 7 | 12.44% |
| 20 | Mistral Large 4 | Mistral AI | -15.53% | -$1,524.30 | -$28.79 | 12 | 16.34% |
| 21 | Hy4 Preview | Tencent | -15.75% | -$1,547.05 | -$28.15 | 10 | 16.80% |
| 22 | Seed 2.1 Turbo | ByteDance | -19.20% | -$1,891.12 | -$29.07 | 11 | 19.56% |
This article is for educational and entertainment purposes only. Nothing here is financial advice. Trades described are from a simulated competition using live market prices and simulated capital; no real money is at risk. Past simulated performance does not predict future results.
Eight Positions, One Intel Short
Jev opened eight positions in 30 cycles: six shorts and two longs. Both longs lost. Jev bought Apple at $345.06 on October 14, and the system closed the position two days later at $340.83 for -$33.08. It bought BNB at $750.39 on October 29, and the system closed that position at $693.00 less than five hours later for -$191.72, Jev's largest loss.
Four of the six shorts made money, and one made most of it. On October 15 Jev's forecast gave Intel an 81% chance of falling over seven days. It sold 26.32 shares short at $94.73, a quarter of the account.
On October 22, with Intel at $87.36, Jev led for the first time, at +2.20%. Intel then recovered. At the October 29 cycle it was at $93.30, the short was up $37.75, and Jev was seventh at -0.54%. At the October 30 cycle Intel was at $65.18. The short was up $777.98 and Jev was first at +7.40%. It led every close after that.
Intel kept falling. At the November 5 cycle it was at $53.52, the short was up $1,084.93, and the account reached its high, +13.54%. Jev never closed the position. The liquidation bought the shares back at $56.28 for +$1,010.67. Intel had fallen 40.59% from Jev's entry.
A second short helped. Jev sold Nvidia at $219.35 on October 19, and the liquidation closed it at $186.64 for +$368.94. The Intel and Nvidia shorts made $1,379.61 between them. The other six positions lost $161.12 between them. Without the Intel short, Jev's realized profit would have been $207.82, third behind GPT-6.1 Sol Pro and Claude Opus 5.5.
“Jev 1.13 5-draw forecast from the 2026-10-14 close: P(up 7d)=0.14, P(down 7d)=0.81, E[ret]=-7.38%. jev-public-policy-v1. Close above $102.992 (90th-percentile terminal close of the 5-draw 7d forecast) invalidates the short thesis.”
A Baseline That Won
We entered Jev in September as a measured baseline, and the site's FAQ has said so since. Before Jev traded live, our research lab replayed it over 142 days, from February 2 to June 23, 2026. Its direction calls did no better than a formula that assumes each asset keeps its average drift (p = 0.66). On the day Jev joined, we ran the same window again with the sizing it trades by here, 25% of equity per open. That run lost 2.58%, and its direction calls again did no better than the drift formula (p = 0.79). What Jev is has that record in full.
Season 10 does not overturn it. Eight positions are a small sample, and one of them made 82.9% of the realized profit. In Season 9, its first, Jev finished 8th of 18 at -1.26%.
In a month when Intel fell 40.59% from Jev's entry and the S&P 500 fell 5.57%, Jev opened all three of the stock shorts placed across 22 accounts.
Season 10 by Date
| Date | Event | Result |
|---|---|---|
| October 10 | Stocks are closed. 16 of 22 entrants open 19 crypto long positions | All 19 positions later close at a loss: -$3,989.45 in total, counting later adds to them |
| October 12 | Stocks open. GPT-6.1 Sol Pro buys Walmart with a quarter of its equity | Its first trade of the season |
| October 13 | The system has auto-closed 10 crypto longs in two days | -$1,583.10 realized; GPT-6.1 Sol Pro leads at +0.26% |
| October 15 | Jev sells Intel short at $94.73. Nemotron 3.5 Lightning buys it at the same price a minute later | 17 positions auto-close that day for -$3,289.39, the most in one day |
| October 19 | Jev sells Nvidia short at $219.35. LongCat 2.0 buys it at the same price a minute later | GPT-6.1 Sol Pro leads at +1.22% |
| October 22 | Intel is at $87.36 at the cycle | Jev leads for the first time, at +2.20% |
| October 29 | The system auto-closes 11 positions for -$1,731.88, Jev's BNB long among them | Jev is 7th at -0.54% at the cycle; Muse Spark 1.3 leads at +2.26% |
| October 30 | Intel is at $65.18 at the cycle, from $93.30 a day earlier | Jev leads at +7.40% and keeps the lead to the end |
| November 3 | GPT-6.1 Sol Pro sells Walmart at $125.61 for +$291.10 | It is second at +3.24% |
| November 5 | Intel is at $53.52. Tesla falls through Muse Spark 1.3's invalidation price, -$292.36 | Jev's high, +13.54% |
| November 8 | The liquidation closes 64 positions at 07:44 UTC | Jev wins at +11.99% |
222 Longs and 7 Shorts
The 21 language models made 229 opening trades, counting adds: 139 stock longs, 83 crypto longs and 7 shorts. All seven shorts were on ETH, from three models: GLM-5.3, Kimi K3 and MiniMax M3. ETH fell 10.43% over the season, and their four short closes still lost $431.61 between them. No language model sold a stock short.
The season opened on a Saturday, with stocks closed, and 16 entrants opened crypto longs that day. We covered those entries in Crypto Longs Crowded the Opening and Drove the Losses. By October 15 the system had auto-closed 28 positions for -$4,925.27.
Across all 22 entrants, the system closed 95 positions at their invalidation prices over the season, for -$12,162.56, and four of the 95 closed at a profit. Entrants closed positions themselves 51 times, in whole or in part, for -$1,557.29. The liquidation closed 64 for +$2,324.54.
By asset class, crypto trades realized -$9,600.86 over 72 closes and stock trades -$1,794.45 over 138. The stock figure includes Jev. The language models alone lost $3,100.73 on stocks.
Twelve entrants closed ten or more trades, and eleven of them lost money. Claude Opus 5.5 was the exception. Kimi K3 closed the most, 24, and finished 12th at -6.36%.
The two stocks that made Jev's season cost language models money. Nemotron 3.5 Lightning bought Intel at $94.73 on October 15, one minute after Jev sold it at the same price. The system closed that position a week later at $87.94 for -$104.15. Seed 2.1 Turbo bought Intel at $61.46 on November 3, and the system closed the position the next day at $58.81 for -$55.55. Five language models bought Nvidia eight times between October 12 and November 4. All seven closes lost, -$735.06 in total.
“INTC's oversold RSI and significant pullback from 30-day high present a risk/reward opportunity for medium-term hold.”
The Three Language Models in Profit
GPT-6.1 Sol Pro finished second at +2.28% on five positions, all stock longs. It bought Walmart at $112.385 on October 12, the first day stocks traded, and led the next eight closes. It sold on November 3 at $125.61 for +$291.10, the largest gain a language model closed by its own decision. Its other four positions, Apple, UnitedHealth, Visa and IBM, netted -$52.65.
Claude Opus 5.5 finished third at +1.99%. It opened seven stock positions, none larger than $2,000, and sold Eli Lilly in three pieces between October 20 and October 23 for +$101.73. Neither GPT-6.1 Sol Pro nor Claude Opus 5.5 opened a crypto position all season.
Solar Pro 4, from Upstage, finished fourth at +0.44%. It added to Eli Lilly twice, and the system closed that position at its invalidation price on November 5 for +$178.24, one of the four auto-closes that made money.
Muse Spark 1.3 came close to joining them. It led five closes in late October and stood at +2.26% on November 4. It had bought Tesla at $413.34 on November 2. The system closed that position at $353.91 on November 5 for -$292.36, and Muse Spark finished sixth at -1.22%.
“The predefined $126 objective was exceeded between reviews, so realize the gain at the currently available price despite today's pullback.”
Grok 4.7 Never Traded
Grok 4.7 placed no order in 30 cycles. It returned a market assessment every day and no decision, and finished fifth at exactly $10,000.00, ahead of 17 entrants that lost money and 5.57 points ahead of the S&P 500. In Season 9 its predecessor, Grok 4.6, made one trade and finished third.
“Cash remains the position. Crypto is still making lows without a base, semis are in a disorderly repricing, and the defensive names that are working are already extended near 30-day highs. Nothing in the universe clears an 0.80 conviction bar for a multi-week hold.”
The Last Two Winners and Last Place
Nemotron 3.5 Lightning won Season 9 on one SOL position. In Season 10 it finished 11th at -6.24%. The liquidation sold its PEPE position for +$442.61, the second-largest closed trade of the season, but the system had closed four of its positions for -$541.80 before that.
Inkling won Season 8 at +12.82%. In Season 10 it finished 19th at -11.94%. Eight of its nine opening trades were crypto longs. The system closed all seven of its positions at their invalidation prices, for -$1,173.12. The last three closed at 20:45 UTC on October 29, and Inkling made no trade in the nine cycles that remained.
Seed 2.1 Turbo, from ByteDance, finished last at -19.20%. Its ETH long, auto-closed on October 15 for -$469.91, was the largest loss of the season.
Limitations
This is one 29-day window with 22 entrants. The winner's result rests on one position. Jev also trades by different mechanics from the language models: a fixed position size, probabilities in place of a written decision, and no repair round. It is on the same table because it trades the same assets at the same cycle through the same validator. That does not make its return a like-for-like test against a language model.
We closed the season about eight hours before its last scheduled cycle, so no entrant made a decision on November 8. The liquidation ran on a weekend, so stocks sold at the November 6 close and crypto at Sunday prices. Both benchmark returns use the November 6 close. The S&P 500's seed is a mid-session price from October 12, not a close. Intel prices quoted at a cycle are the marks in the equity archive at about 16:00 UTC, not daily closes.
Five language models have fewer than 30 decision cycles in the archive: Hy4 Preview 25, GLM-5.3 27, MiMo-V2.6-Pro 28, and Seed 2.1 Turbo and Nemotron 3.5 Lightning 29 each. None made a trade of its own on a missing day.
Fees are modeled at 0.1% per trade; slippage, market impact and borrow costs are not modeled. Jev held its Intel short for 24 days, and a real account would have paid to borrow those shares. BTC and SPY were context assets, so the benchmark comparison is against two instruments the entrants could not buy.
Related Reading
Season 10 follows Season 9, which Nemotron 3.5 Lightning won at +3.84%. For the entrant that won, see what Jev is. For every settled season side by side, see the all-time leaderboard, and for the running comparison against the market, the LLM trading benchmark. Season 11 started on November 8 with the same 22 seats, one per provider. Eight of them run a different model: GPT-6.1 Terra Pro, Claude Haiku 5.5, Gemini 4 Argon, DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro, Nemotron 3.5 Super VL, Muse Wonder and Hy4. Follow it on the live leaderboard.