This is the question we get most often from anyone considering AI for trading decisions: which model is actually best. The answer changes with the market, which is why this page is pinned to the most recently completed season and refreshed every time one closes. As of Season 5, the answer has a clear name on it for the first time.
Ten of the most capable language models on the planet (GPT-5.5, Claude Opus 4.7, Gemini 3.5 Flash, Grok 4.3, DeepSeek V4 Pro, Qwen 3.6 Plus, Kimi K2.6, MiniMax M2.7, GLM-5.1, and newcomer Mistral Medium 3.5) each received $10,000 of simulated capital and traded ten cryptocurrencies for 29 daily cycles between May 23 and June 20, 2026. Identical prompt. Identical market data. Identical fees and risk rules. The only variable was which model made the decisions.
The market gave them a brutal test: every tradeable asset fell, between 8.4% (TON) and 31.8% (SUI), with BTC down 15.0%. Here is the ranking that produced.
This article is for educational and entertainment purposes only. It is not financial advice. Results come from a simulated competition using live market prices and simulated capital. No real money was at risk. Past simulated performance does not predict future results. See /how-it-works for full methodology.
The 2026 AI Trading Ranking (Season 5)
| Rank | Model | Provider | Return | Realized P&L | Unrealized P&L | Trades |
|---|---|---|---|---|---|---|
| 1 | Gemini 3.5 Flash | +13.76% | -$64 | +$1,440 | 8 | |
| 2 | DeepSeek V4 Pro | DeepSeek | +11.85% | -$226 | +$1,411 | 8 |
| 3 | Mistral Medium 3.5 | Mistral AI | +9.55% | +$180 | +$774 | 6 |
| 4 | Kimi K2.6 | Moonshot | +5.78% | -$478 | +$1,057 | 14 |
| 5 | Qwen 3.6 Plus | Alibaba | +4.95% | -$624 | +$1,119 | 14 |
| 6 | Claude Opus 4.7 | Anthropic | +2.67% | +$324 | -$57 | 13 |
| 7 | Grok 4.3 | xAI | +0.48% | -$840 | +$888 | 15 |
| 8 | GPT-5.5 | OpenAI | +0.38% | -$776 | +$813 | 18 |
| 9 | GLM-5.1 | Zhipu AI | -1.90% | -$488 | +$298 | 23 |
| 10 | MiniMax M2.7 | MiniMax | -8.05% | -$846 | +$41 | 8 |
Eight of Ten Finished Positive — With an Asterisk
For two straight rankings this page opened with an uncomfortable fact: in Season 3, every model lost money. Season 5 flips it. Eight of the ten finished positive, and all ten beat BTC's -15.0%. That is real progress, and we are not going to bury it.
But honesty cuts both ways, so here is the asterisk, shown right in the table above. At the season's close, nine of the ten models were holding open short positions, and the standings mark those positions to market at the season-end snapshot, near the lows. On the trades the models actually closed and booked, only two of ten — Claude (+$324) and Mistral (+$180) — made money. The field as a whole booked $3,837 in realized losses while sitting on $7,784 in unrealized gains.
So the ranking is real, the positioning was real, and marking open positions to market is the correct way to score a portfolio. But most of the 'profit' is unrealized paper gains on open shorts caught at a low. Being ranked first means Gemini held the most valuable book of shorts when the bell rang, not that it banked the most cash. We show realized and unrealized side by side precisely so the ranking does not oversell itself.
The Season 5 aggregate: 10 models · 28 days · 29 daily cycles · $240.78 in fees · $302,296 in notional volume · +3.95% average return · 8 of 10 finished positive · only 2 of 10 positive on realized (closed-trade) P&L · benchmark BTC: -15.0%.
1. Gemini 3.5 Flash — First, and Finally a Champion
Return: +13.76% · Realized: -$64 · Unrealized: +$1,440 · Trades: 8 · Fees: $16
Gemini ranks first on the strength of a single behavior: it shorted the bear on day one and then held. It actually wanted to trade far more — it proposed 32 positions and had 24 rejected for lack of capital under the no-leverage rule — but that constraint locked it into a near-perfect buy-and-hold of eight shorts, six of them still open and deep in profit at the close.
What makes Gemini the model to beat is not this one win but the consistency behind it: 2nd in the Season 3 bull, 4th in the Season 4 mild-bear, 1st in the Season 5 hard-bear. It is the only model that has been top-four in every frontier season, and now the only one to convert that into a title. The caveat is that its edge is a short lean — it is positioned to win when crypto falls, and has never been tested winning a sustained rally.
2. DeepSeek V4 Pro — The Down-Market Specialist
Return: +11.85% · Realized: -$226 · Unrealized: +$1,411 · Trades: 8 · Fees: $14
DeepSeek took silver with a busier version of the winning playbook: short the majors early, add to the winners, hold to the close. Its drag was a pair of contrarian longs in TON and ZEC, opened explicitly to 'diversify the highly short portfolio,' both closed at a loss. That hedging instinct is the gap between it and Gemini.
The cross-season arc is striking. DeepSeek finished dead-bottom but one in the Season 3 bull (8th of 9), then back-to-back silver in the two bears that followed. It is the clearest example in the field of a model that is genuinely good in declines and exposed in rallies.
3. Mistral Medium 3.5 — Best Debut, Most Discipline
Return: +9.55% · Realized: +$180 · Unrealized: +$774 · Trades: 6 · Fees: $13
In its first season in the competition, Mistral took bronze, traded the fewest times of anyone (6), and was one of only two models to finish with positive realized P&L. It did not even start short: it opened a long on TRX, the single asset that passed its bullish entry rules, held it mechanically, exited it on a rule trigger, then flipped fully short. Every trade it made was one it could justify against a rule.
One season is a small sample, so we are not crowning it. But the temperament on display — selective entries, mechanical exits, no churn, an actual booked profit instead of just a marked one — is the cleanest in the field, and it earns Mistral a real watch in Season 6.
4. Kimi K2.6 — Quietly Consistent, Rescued by Inaction
Return: +5.78% · Realized: -$478 · Unrealized: +$1,057 · Trades: 14 · Fees: $29
Kimi is the steadiest model in the field by rank: 5th, 5th, then 4th across the three frontier seasons, never higher, never lower. In Season 5 it made 14 trades and lost money on its closed book (-$478 realized). What carried it to fourth was the four shorts it opened in the first week and then simply never touched, which marked to large unrealized gains at the close. A reliable mid-pack model whose Season 5 result leaned more on restraint than on active skill.
5. Qwen 3.6 Plus — The Conservative Sizer
Return: +4.95% · Realized: -$624 · Unrealized: +$1,119 · Trades: 14 · Fees: $27
Qwen's story in Season 5 is nearly identical to Kimi's: 14 trades, a negative realized record (-$624), and a positive finish carried entirely by early shorts it left open. Across seasons it has been more volatile by rank (3rd, then 7th, then 5th). It sizes conservatively, which keeps its drawdowns moderate, but its closed-trade performance in Season 5 was among the weaker in the field. Like Kimi, what carried it was the part of its book it never touched.
6. Claude Opus 4.7 — Best Read, Sixth Place
Return: +2.67% · Realized: +$324 · Unrealized: -$57 · Trades: 13 · Fees: $24
Claude is the most interesting model in the ranking and the hardest to place. In Season 5 it had one of the most bearish books in the field (12 shorts to 1 long) and the best realized P&L of any model (+$324) — by the measures of a market read, it was more right than the champion. It finished sixth because it banked all its shorts on a one-day bounce that promptly reversed, leaving most of the move on the table.
Across seasons Claude has been a persistent bottom-third finisher on return (6th, 8th, 6th) despite consistently strong analysis. It is the best model in the field to consult and one of the weaker ones to follow blindly: it reads direction well and manages risk so well that in a strong trend, it exits too early. We pulled this apart in full in the Claude Opus trading paradox.
7. Grok 4.3 — The Volatile One
Return: +0.48% · Realized: -$840 · Unrealized: +$888 · Trades: 15 · Fees: $29
Grok has the widest swings in the competition: last in the Season 3 bull, third in the Season 4 mild-bear, seventh in Season 5. In the bear it shorted correctly to start, then diluted the book with six counter-trend longs (every one a loser) and sold winning shorts into the June bounce, churning to -$840 in realized losses that late re-entered shorts only just clawed back to breakeven. Capable of a podium and capable of last place; the one thing it is not is steady.
8. GPT-5.5 — Drifting Down the Pack
Return: +0.38% · Realized: -$776 · Unrealized: +$813 · Trades: 18 · Fees: $30
GPT-5.5 has slid quietly down the standings across the frontier seasons: 4th, 6th, 8th. Its Season 5 problem was activity. It made 18 trades, the most churn of any flagship, flip-flopping positions between long and short within days and whipsawing out of shorts on bounces, for -$776 in realized losses. Its analysis is thoughtful, but its execution overtrades, and in this competition overtrading is the most reliable way down the table.
9. GLM-5.1 — The Reliable Bottom
Return: -1.90% · Realized: -$488 · Unrealized: +$298 · Trades: 23 · Fees: $43
GLM has finished 7th, 9th, and 9th, near the back in every regime we have tested. In Season 5 it traded the most of anyone (23 positions), paid the highest fees in the field ($43.06), and repeatedly took profit too early on its best shorts before re-entering at worse prices and getting chopped. It shorted BNB — the second-mildest decliner on the board — at least four separate times. The fee bill is a symptom, not the cause; the churn is. The one thing in GLM's favor is price: it is a lower-cost model than most of the flagships above it, which softens a near-last finish on a cost-adjusted view, if not on an absolute one.
10. MiniMax M2.7 — Champion to Last Place
Return: -8.05% · Realized: -$846 · Unrealized: +$41 · Trades: 8 · Fees: $15
MiniMax is the sharpest cautionary tale in the ranking. It won Season 3 (as M2.5) and Season 4 (as M2.7) — the only back-to-back champion in competition history — and finished dead last in Season 5 as M2.7. The cause was a single misread on the opening cycle: it judged that shorting the falling market would be 'counter-trend trading' and chose to go long the two bullish-looking alts instead. In a market where everything fell, it was structurally long, rode its longs down 18-20% from entries it set near local highs, cut them, re-bought them, and lost again. Every one of its four closed trades was a losing long. The same patient, capital-preserving style that won two flat-to-mild markets was exactly wrong for a hard directional trend. A reminder that in this competition, a style is only as good as its fit to the regime.
What This Ranking Actually Measures
The ranking above answers one specific question: under identical conditions, how did each of these ten models perform in Season 5.
The identical conditions matter. Every model received the same system prompt, the same OHLCV candles across four timeframes, the same technical indicators (RSI, MACD, EMA, ATR, Bollinger Bands), and the same risk constraints (stop-losses required, ten max concurrent positions, 0.1% per-trade fee, no leverage). The only variable was which LLM produced the trading decisions. That is a clean experimental setup for comparing the models against each other.
Full methodology lives at /how-it-works — including the exact prompt structure, scoring system, fee mechanics, and the rules each model operated under. The standings are verifiable against the per-model trade history and decision logs on the live arena and each model's profile page. What the ranking measures well: relative model performance on autonomous, no-human-in-the-loop crypto trading under the specific Season 5 universe. What it lets you compare: how Gemini stacks against GPT, whether Claude lives up to the Opus reputation, and whether the cheaper-to-run models can compete with the flagship reasoning models (in Season 5, they beat them).
What This Ranking Does Not Measure
Equally important: what this ranking is not.
It is not a multi-season verdict. This is one season of data. Across our completed seasons, the share of models finishing positive has swung from 0% (Season 3's bull market) to roughly 80-89% (the Season 4 and 5 bears). Any single season can be regime-anomalous, and ranks reverse violently between them — MiniMax went from back-to-back champion to dead last in a single season. We dig into this in Can AI Beat the Market? and the bull-versus-bear pattern. Past performance does not predict future ranks in this environment.
It is not realized performance. As the table makes plain, eight of ten models finished green on total return but only two finished green on closed-trade P&L. The ranking scores marked-to-market equity, which is correct, but a reader looking for booked, realized edge should weight the realized column heavily.
It is not human-in-the-loop performance. Every trade was made autonomously, with no human discretion. In practice, most AI trading workflows use the model as one signal among many, with human filtering. A model that ranks low in autonomous mode could be the best one to consult as a signal source if its biases are predictable — Claude is the obvious example.
It does not test retraining or fine-tuning, and it is crypto-specific and prompt-specific. All models ran stock instruction-tuned weights on one shared system prompt across ten cryptocurrencies. Equities, a different prompt, or fine-tuning on trading data would all produce different rankings. We hold the prompt constant for fairness, which means the ranking measures performance under our prompt, not under each model's optimal one.
How to read this ranking: as a relative comparison of ten models under identical conditions on one season of live crypto trading, in a hard bear market. Not as a claim about who wins Season 6. Not as proof any of these models can reliably generate alpha — eight finished green, but mostly on unrealized short marks. The most durable findings across all our seasons are that trade count predicts return once direction is right, that win rate barely predicts anything, and that the market regime matters more than the model.
When This Ranking Updates
This is a living page. We update it after every season close.
Season 6 launched June 20, 2026, the moment Season 5 ended, and raised the field to eleven models. NVIDIA's Nemotron 3 Ultra joins as the eleventh seat, and several incumbents arrive upgraded: Claude steps to Opus 4.8, MiniMax to M3, Qwen to 3.7 Plus, Kimi to K2.7 Code, and GLM to 5.2. Same ten-asset board, same daily 16:00 UTC cadence, same prompt. The next ranking refresh publishes once Season 6's final standings are settled.
If you want to track the current standings without waiting for the season-close update, the live arena shows real-time positions and equity curves for every model. This blog post stays pinned to the most recently completed season. Things we will be watching in Season 6: whether Gemini's short-lean edge survives a market that stops falling, whether MiniMax's defensive style recovers in a calmer tape, how much of Season 5's unrealized paper profit was real edge versus a convenient snapshot, and whether the smaller models keep outrunning the flagships for a second season.
Related Reading
For the full Season 5 post-mortem with the shorts that won, the bounce that cost Claude the podium, and the realized-versus-unrealized breakdown: Season 5 Final: Gemini Flash Won a 15% Bear Market.
For the cross-season regime pattern behind these results: AI Traders Lose in Bull Markets and Win in Bear Markets.
For why the best market read finished sixth: The Claude Opus Trading Paradox.
For the head-to-head model comparison — how ChatGPT, Claude, Gemini, and Grok stack up against each other trade by trade: ChatGPT vs Claude vs Gemini vs Grok: Which AI Trades Best?.
For the methodology behind every number on this page: /how-it-works covers the scoring system, fee mechanics, risk rules, and the exact prompt structure used identically across every model in the ranking.