Home/ Incident database/alpha-arena-six-models-lose-2025-11
Financial loss Trading agent failure verified

Four of six AI models lost real money trading crypto perpetuals autonomously in nof1's Alpha Arena

nof1 (Alpha Arena Season 1) · Nov 3, 2025 · Trading agent

What happened

Nof1's Alpha Arena Season 1 gave six models - Qwen 3 Max, DeepSeek V3.1, GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5 and Grok 4 - $10,000 of real capital each to trade crypto perpetual contracts on Hyperliquid with no human intervention, from October 18 to November 3, 2025. Four of the six finished below their starting stake. Qwen 3 Max won at +22.31% and DeepSeek V3.1 was the only other model above water. Individual model loss amounts were not published in the sources verified for this record.

—
Loss / impact
Nov 3
2025
Trading
Failure mode
Trading
Agent type

Causal vector

Real capital, leverage, and no loss cap: identical prompts produced divergent autonomous trading, and nothing halted a model once its equity curve went negative

Source

Reported by ForkLog (results shared by nof1 founder Jay A. Zhang; field and dates corroborated by gncrypto.news and TradeRank's archived leaderboard). Verified against the primary report.

Read the original report ↗

How a spend firewall would have helped

A per-transaction cap plus a daily loss total enforced outside the agent converts an open-ended drawdown into a stopped line: once a model's realized loss for the day crosses the threshold, further orders are FLAGGED, not filled. Velocity limits also blunt the over-trading pattern - a model placing orders every minute exhausts its budget before it exhausts the account.

The six rule types that contain this class of failure

Per-transaction cap

Any single spend above your ceiling is BLOCKED before it moves.

Daily total

Cumulative spend across all agent calls, bounded per day.

Velocity limit

Stops runaway retry loops — the #1 cause of overnight losses.

Merchant allowlist

Only approved destinations can ever receive funds.

Category rules

Flag high-risk classes (crypto, infra, refunds) for review.

Approval threshold

Above a value, the action waits for a human.

Related incidents

Claude Opus co-authors a malicious npm commit that compromises a crypto trading agent

crypto trading agent operator (npm dependency victim) · Apr 29, 2026

Don't be the next entry

Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 105 documented failures, one control.