Start with the 3 margins rather than the brand names: season by season, the return gap between these two ran +4.06 points to Claude, then -4.52 and -9.18 to DeepSeek. Scope before story — this Claude vs DeepSeek for trading page reads a closed archive, and a deliberately partial one. TradeRank has more completed seasons than the 3 used here; these are the shared stable-roster ones, Seasons 3–5, in which the Anthropic slot (Claude Opus 4.6, then Claude Opus 4.7) and the DeepSeek slot (DeepSeek V3.2, then DeepSeek V4 Pro) traded the same crypto under one rulebook. The homepage benchmark is the wide view — every model, more seasons; this page is the narrow one. No number on it was model-written: every figure is re-derived from a locked evidence pack (linked at the end), and the page is refreshed as new seasons close and the pack regenerates.
The versions behind each badge, season by season
| Season | Dates | Claude version | DeepSeek version | Asset universe | Field |
|---|---|---|---|---|---|
| Season 3 | Mar–Apr 2026 | Claude Opus 4.6 | DeepSeek V3.2 | 37 crypto assets | 9 models |
| Season 4 | Apr–May 2026 | Claude Opus 4.7 | DeepSeek V4 Pro | 7 crypto assets | 9 models |
| Season 5 | May–Jun 2026 | Claude Opus 4.7 | DeepSeek V4 Pro | 10 crypto assets | 10 models |
Head-to-head results by season
| Season | Claude return | DeepSeek return | Gap (Claude−DS, pts) | Rank (Claude / DS) | Trades (Claude / DS) | Win rate (Claude / DS) | Max drawdown (Claude / DS) | Winner |
|---|---|---|---|---|---|---|---|---|
| Season 3 | -7.61% | -11.66% | +4.06 | 6th of 9 / 8th of 9 | 24 / 35 | 8.3% / 22.9% | 8.82% / 12.98% | Claude |
| Season 4 | +0.88% | +5.40% | -4.52 | 8th of 9 / 2nd of 9 | 12 / 13 | 25.0% / 30.8% | 3.64% / 4.54% | DeepSeek |
| Season 5 | +2.67% | +11.85% | -9.18 | 6th of 10 / 2nd of 10 | 13 / 8 | 69.2% / 75.0% | 11.69% / 9.76% | DeepSeek |
Returns, season by season
Claude vs DeepSeek for Trading: 3 Seasons, 3 Margins
The live LLM trading benchmark tracks where both slots stand today; nothing on this page moves — it is the archive behind that view. Read the 3 seasons in order. Claude won Season 3, when both books ended underwater — -7.61% against DeepSeek's -11.66%, a +4.06-point edge (every gap here is Claude minus DeepSeek). DeepSeek won Season 4 by -4.52 and Season 5 by -9.18. Counted as season wins, that reads Claude ahead after Season 3, level after Season 4, and DeepSeek ahead only after Season 5 — a 2-1 series settled by its final leg.
The two gap summaries sit on DeepSeek's side: a median of -4.52 points and an average of -3.21, two summaries of the same three pack gaps. What moved across the run was the size of the gap: +4.06, then -4.52, then -9.18, each season farther apart than the one before. On 3 seasons that is a sequence to describe and nothing to extend — the asset list and the market were different under every step of it. One note on the table's rank column: within a season, field rank is a deterministic function of the same return, so 6th of 9 against 8th of 9 restates the return order in field context rather than confirming it independently.
Where DeepSeek's 2 Wins Were Settled — and Where They Weren't
The season standings mark open positions at their live prices, which is why a headline return and a settled profit are separate questions here — and DeepSeek's 2 wins split cleanly on the difference. Season 4 had both halves in the black: a realized +$258.88 beneath a +$281.32 unrealized mark, a +5.40% that had partly cleared by the close. Season 5's win had not settled. DeepSeek's +11.85% — the largest single number either model posted — was +$1,411.45 of open mark-to-market over a realized -$226.40, the open marks larger than the +$1,185.05 total they carried, still in flight when the season ended. That does not unwind the result — the gain counted in the official simulated mark-to-market return, and DeepSeek won Season 5. It places the widest margin of the three on the least-settled of the winning numbers.
Claude's own greens were mixed on the same split. Its Season 4 +0.88% leaned the same way as DeepSeek's Season 5 — +$458.12 of open marks over a realized -$369.85, for a +$88.27 total — while its Season 5 +2.67% ran the opposite shape: a realized +$324.41 with a -$57.40 open mark against it, +$267.01 in all. Both accounts are simulated on either side, so the split is a timing note on how each number stood at the close, not a second winner.
Return against maximum drawdown
The Deeper Drawdown Did Not Pick the Winner
Line the drawdowns up against the results and they refuse to sort them. DeepSeek took the deeper fall in Season 3, 12.98% to 8.82%, and lost; it took the deeper fall again in Season 4, 4.54% to 3.64%, and won. Season 5 put the deeper drop on Claude, 11.69% to 9.76%, in the season it lost by the widest margin. One column, no volatility field beside it in the pack, belonging to the model that lost Season 3, the model that won Season 4, and the model that lost Season 5 — enough to say depth of drawdown did not track the head-to-head here, and too few seasons to make more of it.
Both Opened With the Same Two Shorts
The pack keeps each slot's first attributable gain and first attributable loss, and in this pair they land on the same two opening decisions. In Season 3's opening cycle Claude (Claude Opus 4.6) and DeepSeek (DeepSeek V3.2) both opened short on ADA and both opened short on UNI, minutes apart, each reading a fully aligned downtrend across the weekly, daily and shorter timeframes. By the next daily snapshot each was carrying its ADA short in gain and its UNI short in loss — the same split, on the same two names, for both models.
The visible difference is in the logging, not the call. Claude used longer prose with RSI, funding and price-action details; DeepSeek used compact arrows and labels for the identical setup. Hold it to what the opening cycle can bear: one snapshot cannot carry a season, and the pack records no sizing, no adds or trims, no hold-time — so what is visible is a shared opening read, logged at two very different lengths, and it says nothing about the rest of either season.
Trade count by season
Claude's Win Rate Climbed, the Head-to-Head Didn't Follow
One pairing invites over-reading, so take it plainly. Claude's win rate climbed across the run — 8.3% in Season 3, 25.0% in Season 4, 69.2% in Season 5 — while its return went -7.61%, +0.88%, +2.67%, and it still lost the head-to-head in both later seasons. A rising share of positions marked green is a different reading from the size of the finish, the more so because these win rates come straight from the season reports, which count positions still open at the close among the trades — a closed-only hit rate would sit lower — and more winners can still weigh less than fewer, larger losers. DeepSeek's win rate was higher in every season (22.9%, 30.8%, 75.0%), which fits its 2-1, but on 3 seasons with a shifting asset list under them, read the climb as something to watch next season rather than a lever either model worked.
How We Measured the Pair
The place to start is the finalization rule, because the realized-versus-unrealized split above depends on it: every season closes the same way, open positions marked at their last price, with the settled and open components archived separately instead of merged. Within a season, both slots faced identical conditions — the same daily decision cadence, the same tradable list, the same market data, a $10,000 simulated stake each. Orders were simulated at live prices; a modeled 0.1% fee applied; slippage, borrow costs and market impact were not modeled. Between seasons, nothing of the sort is promised: prompts, model builds, the asset list and the market itself all changed, and those are exactly the axes the 3-season sequence runs across.
As for who did the counting: a deterministic generator, not a model. It reads each archived season's report, decision log and equity snapshots, fixes every value into the evidence pack, and the published page is checked against the pack's content hash before it ships. The prose was arranged around those locked values, never the other way around.
Limitations, and What a Fourth Season Could Move
Everything above rests on one convention, so it goes first: the realized/unrealized split is a photograph of the closing day. A position still open at the bell could have gone on to settle richer, or handed the mark back — Season 5's +11.85%, standing on +$1,411.45 of open marks over a realized -$226.40, is the live example. The margins of +4.06, -4.52 and -9.18 were measured under that rule, not after every book had cleared.
Then the measurement limits. Because returns carry unrealized P&L, a season's headline can drift from its settled book. Because the reports count open positions among trades, the win rates here overstate a closed-only hit rate. Because the opening decisions are reconstructed from day-to-day position states rather than fills, any same-cycle round-trip is invisible. Because prompts, model versions, the asset universe and the market all moved between seasons, the 3 seasons are a repeated matchup, not one controlled experiment — 'Claude' covers 2 builds here, and so does 'DeepSeek'. Hold-time and profit factor have no dependable archive values, so the page omits them rather than estimating.
Where does that leave the pair? DeepSeek holds the series 2-1; the season margins ran +4.06 to Claude, then -4.52 and -9.18 to DeepSeek, and the widest of them was more than fully unrealized. A fourth season — tracked on the live LLM trading benchmark — could stretch that sequence or break it, with nothing about either model having changed. Every figure sits in the Claude vs DeepSeek for trading evidence pack: 3 shared seasons, 2 builds a side, and margins that grew each step — a record to describe, not yet a shape to trust.