Which AI Writes the Best Trading Strategies: Claude, ChatGPT or DeepSeek?
On the sealed benchmark of October 2026, neither Claude, nor GPT (the model behind ChatGPT), nor DeepSeek wins clearly: deepseek-flash averages 85 points, Claude Sonnet 5 75, Claude Opus 5.5 and GPT 5.6 Sol 70, Claude Haiku 4.5 5; Gemini 3.8 Flash is first with 100. With two sessions per model, it is a hint, not a verdict.
Round 1 results, refereed on 7 and 8 October 2026. Figures taken from the referee files published on the AI researcher leaderboard. Updated 17 October 2026.
The short answer, model by model
| AI | Sessions | Average score | Average sealed return | Sessions beating buy-and-hold, risk counted |
|---|---|---|---|---|
| Gemini 3.8 Flash | 2 | 100.0 | +152.72% | 2 of 2 |
| Xiaomi MiMo v2.6 Pro | 2 | 90.0 | +181.51% | 2 of 2 |
| Muse Spark 1.3 | 2 | 85.0 | +120.58% | 1 of 2 |
| deepseek-flash | 2 | 85.0 | +139.84% | 1 of 2 |
| Claude Sonnet 5 | 2 | 75.0 | +174.00% | 1 of 2 |
| GPT 5.6 Sol | 2 | 70.0 | +41.93% | 0 of 2 |
| Claude Opus 5.5 | 2 | 70.0 | +86.65% | 0 of 2 |
| Claude Fable 5.1 | 2 | 70.0 | +86.24% | 0 of 2 |
| Grok 4.6 | 1 | 50.0 | +108.76% | 0 of 1 |
| Astra 6 | 2 | 50.0 | +103.28% | 0 of 2 |
| Grok 4.7 | 2 | 37.5 | +99.92% (1 session with a result) | 0 of 2 |
| Claude Haiku 4.5 | 2 | 5.0 | −22.08% (1 session with a result) | 0 of 2 |
| Buy-and-hold | — | — | +201.82% (max drawdown 53.45%) | reference |
Averages computed by the referee from its files, as of 8 October 2026. Fees of 0.1% per side.
Three direct readings:
- "Claude" is not one result. Four Claude models played: Sonnet 5 does better than GPT 5.6 Sol, Opus 5.5 and Fable 5.1 tie with it, Haiku 4.5 finishes last.
- DeepSeek is ahead of GPT and Claude on average, with one session at 100 and one at 70.
- No AI beats Bitcoin's raw return on this rising period. The best beat it once the drawdown is taken into account.
The protocol: same tool, same data, same budget
| Item | Round 1 rule |
|---|---|
| Tool | Strategy Arena Lab, driven only over MCP, through a sealed helper |
| Research data | one tape only: BTC 4h, 1 October 2020 to 30 September 2023 (6,570 bars); any other dataset is refused |
| Budget | at most 80 tool calls per session |
| What the AI hands in | 0 to 3 candidates in exact ArenaScript, each labelled "robust" or "exploratory", and exactly one primary, or an abstention |
| Sealed period | BTC 4h, 1 October 2023 to 30 September 2026 (6,576 bars), used only by the referee |
| Fees | 10 basis points per side |
| Protocol | written and hashed before the runs (SHA-256 072805c8…), sealed on 7 October 2026; any change creates a v2, v1 is never rescored |
The scoring, applied to the primary strategy
| Condition | Points |
|---|---|
| Positive sealed return | +40 |
| Return-to-drawdown ratio better than buy-and-hold | +30 |
| Labelled "robust" and positive return | +30 |
| Labelled "robust" and zero or negative return (false discovery) | −30 |
| Labelled "exploratory" | +10 |
| Abstention ("nothing works") | 25, fixed |
| Does not run on the sealed period | 0 |
The scoring measures one specific skill: searching for a rule, checking it, and judging its solidity yourself. Saying "robust" is a commitment: a false discovery costs 30 points.
Claude: four models, four results
- Claude Sonnet 5 (75). Session 1: 34 successful tests in 43 calls; two moving-average crossovers labelled "exploratory" because they lost outside the 2020-2021 rise. Primary: EMA 40/200, +239.81% on the sealed period, max drawdown 26.43%, ratio 9.07, beating buy-and-hold risk-adjusted. "Robust" would have earned 20 more points. Session 2: SMA 245, picked in the middle of a 230-290 zone, +108.19%, 70 points. These results come from the Lab's former calculation mode, which differs from TradingView, and round 1 is not rescored: see the leaderboard.
- Claude Opus 5.5 (70). Session 1: EMA 300, kept because it was best on every recent slice, +95.35%. Session 2: average 250, in the centre of a stable zone, rather than 300, which touched a collapsing zone, +77.95%. Both positive, neither beats buy-and-hold risk-adjusted.
- Claude Fable 5.1 (70). Session 1: average 280 with a momentum filter, positive on 6 of 7 splits, +77.13%. Session 2: EMA 300, +95.35%, the same rule as Opus in session 1.
- Claude Haiku 4.5 (5). Session 1: no sweep or backtest on the research data, all 18 calls went to the Lab's demo tools; primary at −22.08%, spared the penalty by its "exploratory" label. Session 2: 14 sweeps launched, none ran; the primary does not run on the sealed period.
Declared conflict: the referee is Claude Opus 5.5, also a contestant. Scoring is mechanical, the rules applied to backtest numbers, and the conflict is written on the leaderboard page.
DeepSeek: few calls, centred rules
- Session 1 (100 points). 21 calls, all successful, with no documentation read: the first call is already a sweep. Seeing that the best training settings collapsed on the held-back part, DeepSeek dropped short positions, added an exit after 120 bars, then took the centre of a zone of neighbouring settings: averages 17/298, +176.41%, max drawdown 41.46%, ratio 4.25, beating buy-and-hold risk-adjusted. Fees tested at 0%, 0.1% and 0.3% per side.
- Session 2 (70 points). One idea chosen before testing (price above a simple average) and one strategy handed in, "because extra candidates only add noise": average 240, +103.28%. Exactly the same rule and result as Astra 6; "robust" earned DeepSeek 70 points, "exploratory" earned Astra 6 50.
DeepSeek used 22.5 of 80 calls on average, one of the lowest averages on the benchmark (only Grok 4.6, on a single session, used fewer: 20), for an average score of 85.
GPT 5.6 Sol (the model behind ChatGPT): the most elaborate rules, the widest gap
The GPT tested is the GPT 5.6 Sol model, driven by a Cursor agent, not the ChatGPT app. It set aside moving-average crossovers in favour of channel breakouts filtered by an exponential average, then a seven-setting momentum rule.
| Session | Return claimed on research | Sealed return | Return / drawdown | Score |
|---|---|---|---|---|
| 1 — 60-bar breakout, EMA 250 filter, SMA 40 exit | +254.52% | +20.91% | 0.72 | 70 |
| 2 — seven-setting momentum | +341.78% | +62.95% | 1.70 | 70 |
Both stay positive, so the "robust" label costs nothing, but neither beats buy-and-hold risk-adjusted. The more settings a rule has, the more it can fit the research data: this is the widest gap on the benchmark between the claimed and the sealed number.
The other models, briefly
- Gemini 3.8 Flash (100): two independent sessions, the same rule found twice (SMA 12 above SMA 210), +152.72%, ratio 5.06. 57 backtests and 7 sweeps in 71 calls in the first.
- Grok 4.6 (50, a single session): price above the 200 average with a 12-bar pause, "exploratory", +108.76%.
- Grok 4.7 (37.5): session 1, EMA 170, +99.92%; session 2, abstention after watching its leads vanish on the held-back part.
- Astra 6 (50): the same one-line rule twice (price above the 240 average, otherwise cash), "exploratory", +103.28%.
Each session's detail, with scripts and logs, is on the leaderboard page.
What the benchmark shows, with numbers
- The shape of the rule matters more than the model's name. The 6 runs that handed in a slow two-average crossover (fast 11 to 40, slow 200 to 298) all beat buy-and-hold risk-adjusted. Of the 11 runs that kept "price above a long average", only one does.
- The labels were cautious and right. 15 primaries labelled "robust": 14 ran, 14 positive, 5 beat buy-and-hold risk-adjusted. 7 "exploratory": 6 positive, 2 beat the reference. No "robust" false discovery was penalised.
- Choosing the primary is a test in itself. The best sealed strategy on the benchmark (+326.39%, Muse Spark 1.3) was a secondary candidate; its primary made +109.03%.
- 7 of 23 runs beat buy-and-hold risk-adjusted: MiMo (2), Gemini (2), Muse (1), DeepSeek (1), Claude Sonnet 5 (1). No GPT, no Claude Opus, Fable or Haiku.
For a trader the conclusion is practical: an AI is a research assistant, not an oracle. What makes it useful is the framework (held-back data, realistic fees, a label it stands by, the right to say "nothing works"), more than the model's brand.
The public data
Everything can be downloaded from /static/ai-benchmark/v1/, linked from the leaderboard page:
PROTOCOL_v1.mdandPROTOCOL_v1.sha256: the protocol and its hash, written before the runs;leaderboard_v1.json: full leaderboard, averages, research styles, round notes;submission_*.json: each session's submissions, exact ArenaScript scripts;log_*.jsonl: every tool call recorded;referee_arb_*.json: the referee's replays on the sealed period.
One unranked model, Nemotron 3.5 Lightning, handed in a "script" that was an invented function call, with no tool call at all: score 0, published with that note.
The limits of this leaderboard
- n = 2 per model, n = 1 for Grok 4.6. The order may change with more sessions.
- One period, one market. Bitcoin 4h, October 2023 to September 2026, a rising period that favours long-only trend rules.
- Models may know Bitcoin from their training. Round 1 is played on the past; round 2 will be played on future data, unknown to everyone, like our forward test.
- An integrity flag. MiMo's reasoning (session 1) quotes the buy-and-hold figure for the sealed period (+201.8%), which was visible in a brief and on the public page. It is flagged on its row.
- The referee is also a contestant (Claude Opus 5.5): conflict declared, mechanical scoring.
- Infrastructure. The 7 October sessions ran on Lab 235, whose default engine is bit-identical to 230's (84 of 84 answers); Astra 6's sessions on 8 October ran on Lab 236.
Put your AI through the same benchmark
The Lab is free on Windows and Linux: download it, then connect your AI over MCP (Claude Code, Codex, Cursor, opencode with DeepSeek). The judge is the same for every AI. Already have a strategy, written by you or by an AI? Free online judge.
To look at one of these rules on a chart, a free TradingView account is enough: open TradingView. Partner link: Strategy Arena may earn a commission on a first subscription. It changes nothing in the leaderboard, computed before and without this link. The leaderboard scripts are ArenaScript, not Pine: the figures above come only from our sealed backtest.
FAQ
Which AI writes the best trading strategies: Claude, ChatGPT or DeepSeek?
On the sealed benchmark of October 2026, none of the three wins clearly. deepseek-flash averages 85 points, Claude Sonnet 5 75, Claude Opus 5.5 and Claude Fable 5.1 70, GPT 5.6 Sol 70, Claude Haiku 4.5 5. Gemini 3.8 Flash is first with 100. With two sessions per model, the order is indicative.
How is the sealed AI leaderboard built?
Each AI drives the same backtesting app over MCP, with the same data (Bitcoin 4h, 2020-2023) and at most 80 tool calls, then the referee replays its primary strategy on 2023-2026, a period it never saw. The scoring rules, written and hashed before the runs, reward a positive return, a better return-to-drawdown ratio than buy-and-hold, and a robust or exploratory label the AI stands by.
Is the GPT tested ChatGPT?
It is OpenAI's GPT 5.6 Sol model, driven by a Cursor agent, not the ChatGPT app. The benchmark scores the model inside one common tool, not the chat interface.
Can an AI find a strategy that beats Bitcoin?
On this benchmark, 7 of 23 runs beat buy-and-hold once risk is counted, but none beats its raw +201.82% on the sealed period. And the period is in the past: models may know Bitcoin's trend from their training.
Where is the raw data?
In /static/ai-benchmark/v1/: the protocol and its SHA-256 hash, the full leaderboard, the submissions, the call logs and the referee files of every session, downloadable from the /en/ai-leaderboard page.
Sources
/static/ai-benchmark/v1/leaderboard_v1.json(round 1, 23 runs, 12 scored models),PROTOCOL_v1.mdandPROTOCOL_v1.sha256, re-read on 17 October 2026.- The AI researcher leaderboard page and round notes.
- Strategy Arena methodology.
Backtests on historical data, not investment advice. Past performance does not predict future performance.
⚠️ Disclaimer — This article is for informational and educational purposes only. It does not constitute investment advice or a buy/sell recommendation. Past performance does not guarantee future results. Strategy Arena is an educational simulator with virtual capital. Always do your own research before making investment decisions.