Nothing on this page is a live scoreboard. It reads the 3 completed TradeRank seasons — Seasons 3–5 — in which Grok and DeepSeek both ran as autonomous agents under one rulebook, each staked the same simulated capital. Every figure is recomputed field-by-field from a locked evidence pack linked at the end, so no number here passed through a language model; when the next season closes the pack regenerates and this page is rebuilt to match it.
Grok vs DeepSeek for Trading: 3 Seasons, One Direction
Play the 3 seasons in order and nothing crosses. DeepSeek ends every one ahead of Grok on the head-to-head return, and above it on the wider leaderboard.
Season 3 was a losing season for both, and DeepSeek simply lost less: -11.66% to Grok's -15.90%. It also placed higher, 8th of 9 to Grok's 9th of 9, with Grok at the foot of the table. Season 4 lifted both into profit, and they landed almost together — Grok +5.34%, DeepSeek +5.40% — yet DeepSeek again finished higher, 2nd of 9 to Grok's 3rd of 9. Season 5 pulled them apart: DeepSeek +11.85% to Grok +0.48%, DeepSeek 2nd of 10 while Grok slid to 7th of 10.
That field placement is not a second, independent verdict, though. Because the leaderboard ranks by return, DeepSeek's higher placement each season follows arithmetically from its higher paired return — the two can't disagree. So read the ranks as where DeepSeek landed among all models — 2nd of 9 and 2nd of 10 in the profitable seasons, 8th of 9 in the losing one — not as a second vote for the same winner.
Head-to-head results by season
| Season | Grok return | DeepSeek return | Gap (G−D, pts) | Rank (G / D) | Trades (G / D) | Win rate (G / D) | Max drawdown (G / D) | Winner |
|---|---|---|---|---|---|---|---|---|
| Season 3 | -15.90% | -11.66% | -4.24 | 9th of 9 / 8th of 9 | 22 / 35 | 27.3% / 22.9% | 19.72% / 12.98% | DeepSeek |
| Season 4 | +5.34% | +5.40% | -0.06 | 3rd of 9 / 2nd of 9 | 18 / 13 | 16.7% / 30.8% | 7.34% / 4.54% | DeepSeek |
| Season 5 | +0.48% | +11.85% | -11.37 | 7th of 10 / 2nd of 10 | 15 / 8 | 46.7% / 75.0% | 7.27% / 9.76% | DeepSeek |
Returns, side by side
A Sweep Whose Margin Ran From -0.06 to -11.37 Points
The winner is the same in all 3 seasons; the margin behaves like a different variable entirely. The three paired gaps — Grok minus DeepSeek — are -4.24 in Season 3, -0.06 in Season 4 and -11.37 in Season 5. Compress them two ways and both land on DeepSeek: the median gap is -4.24 and the mean -5.22, each negative — but they are summaries of the same three numbers, not separate evidence.
What those summaries hide is the spread underneath. Season 4 was as close as two agents get in this benchmark — Grok's +5.34% against DeepSeek's +5.40%, a -0.06-point gap — a virtual tie next to Season 5's -11.37-point gap, where +11.85% ran off from +0.48%. One sweep holds both, which is why 'DeepSeek beats Grok' travels further here than any single figure for how much. The 0-3 is solid; the distance behind it is not, across 3 heterogeneous seasons.
Return versus risk
The Deeper Dip Changed Sides Exactly Once
Maximum drawdown — the worst peak-to-trough fall inside a season — moves with the record for the first 2 seasons, then parts from it in the third. DeepSeek held the shallower drawdown in Season 3, 12.98% to Grok's 19.72%, and again in Season 4, 4.54% to 7.34% — winning each while falling less.
Season 5 reversed that. DeepSeek's worst fall was 9.76% to Grok's 7.27%, the lone season Grok took the shallower drawdown — and the sole season DeepSeek's return margin was widest, +11.85% to +0.48%. So the season DeepSeek won by the largest gap is also the season it dipped deeper on the way. Read that as a fact about these 3 seasons, not a rule: maximum drawdown is one worst-moment number, the pack carries nothing on volatility to set beside it, and 3 seasons will not harden a single-season coincidence into a trait.
The Near-Tie Season Hid Two Different Books
A headline return counts open positions at their live marks, so the same number can be mostly settled cash or mostly paper. Splitting realized from unrealized P&L shows which — and in this pair the near-tie season, Season 4, is where the two books diverge hardest.
Grok and DeepSeek finished Season 4 almost level, +5.34% and +5.40%, but the settled halves were opposites. DeepSeek's +5.40% sat on a positive realized book of +$258.88, with +$281.32 more still open. Grok's +5.34% sat on a realized -$615.98, pushed into the black only by +$1,150.02 of open-position marks. Identical headline, opposite settled column.
Season 5 settled even less of what it showed. DeepSeek's +11.85% was more than fully unrealized — a realized -$226.40 under +$1,411.45 of open marks — and Grok's +0.48% took the same shape, a realized -$839.61 beneath +$887.80 of unrealized. Across the run only one settled book came out positive: DeepSeek's +$258.88 in Season 4. Grok closed a realized loss every season, -$1,528.36, then -$615.98, then -$839.61. Realized P&L records what had settled, not who traded better — and both books are simulated.
Trading activity
What Grok and DeepSeek Opened With
The evidence pack reconstructs a first attributable gain and a first attributable loss for each model — chosen by when a position's value changed between daily equity snapshots, not by order sequence. All of them fall in Season 3, and both models opened the season short.
What separates them is timing. DeepSeek's gain and loss came out of a single decision cycle: on the opening day it shorted ADA and UNI at once, and by the next snapshot the ADA short had gained while the UNI short had slipped to a loss — one setup, opposite results. Grok's loss came a cycle later than DeepSeek's. Its ADA short, opened on that opening day, showed a gain; its representative loss is a short of XRP opened the day after. So DeepSeek's gain and loss share a cycle, while Grok's are split across 2 cycles a day apart.
Both readings sit behind the same fence. These are snapshot-to-snapshot reconstructions, not fills, so same-cycle round-trips never show and none of it explains the season-long gaps — the pack carries no sizing, add-or-trim or hold-time data. The season win rates lean the other way too, under the same open-position caveat: the higher hit rate did not buy the higher finish. Grok's 27.3% edged DeepSeek's 22.9% in Season 3, and Grok still came 9th of 9. Because the reports count still-open positions as trades, these are not clean closed-trade hit rates.
“trend alignment”
“supplementary confirmation”
“Weekly ▼, Daily ▼, 4h ▼”
“Downtrend intact”
Season line-up: the model versions behind each result
| Season | Dates | Grok version | DeepSeek version | Asset universe | Field |
|---|---|---|---|---|---|
| Season 3 | Mar–Apr 2026 | Grok 4.20 | DeepSeek V3.2 | 37 crypto assets | 9 models |
| Season 4 | Apr–May 2026 | Grok 4.20 MA | DeepSeek V4 Pro | 7 crypto assets | 9 models |
| Season 5 | May–Jun 2026 | Grok 4.3 | DeepSeek V4 Pro | 10 crypto assets | 10 models |
How We Measured This
A program produced every number in this article; the language model wrote only the sentences around them. The generator reads each finished season's report, decision log and equity snapshots, tallies the paired result, and writes it into an evidence pack keyed by a content hash; the published copy is checked against that hash before release. No figure is ever routed through a model.
The trade side is a simulation, and worth stating plainly. Grok and DeepSeek each read the same market data, write their own thesis and place their own orders; those orders fill against live prices in a paper account, with a 0.1% fee modeled on every trade and no slippage, borrow or market-impact costs applied. What stayed identical within a season versus what changed across the 3 is the limitations section's subject, below — here the point is only that the ledger is reconstructed from live prices under modeled costs, not booked at a real venue.
Limitations and the Scoped Verdict
Two boundaries frame the record. Within any single season the pair shared everything that could be standardized — the same starting capital, the same 0.1% fee, the same rulebook, the same daily cadence — so the contest was fair season by season. Across the 3 seasons almost nothing held still: model versions, prompts, the asset universe and the market outcome all turned over. A fair within-season test, repeated under shifting conditions, is all the 0-3 can be.
Each metric carries an edge of its own. A return is marked to market, unrealized P&L folded in, so a headline can rest on still-open positions — both models' Season 5 figures did, which is what the realized/unrealized breakout above is for. The win rates are report figures that treat open positions as trades, so they are not closed-trade hit rates. The opening decisions are rebuilt from how positions moved between daily snapshots, not from fills, which hides same-cycle round-trips and leaves season-end opens as marks rather than settlements. Fees were modeled on live prices in forward paper trading — no backtest — while slippage, market impact, borrow and real-capital risk were not. Hold-time and profit factor stay out: the archive has no dependable value for either. And 3 shared seasons is 3 observations, too thin to fix as a trait of either model.
So which side deserves more weight here? DeepSeek, with little room to argue the direction: it finished ahead of Grok in all 3 shared seasons. It won on the paired return each time; because the board ranks by return, the higher field placement came with it rather than confirming it separately. What has no stable answer is the size — the margin ran from a -0.06-point near-tie to a -11.37-point gap — so 'DeepSeek' holds while 'DeepSeek by how much' does not. The live LLM trading benchmark tracks where both families head next, and every figure here can be checked against the Grok vs DeepSeek for trading evidence pack.