Reasoning time isn't IQ. But it might be conviction.
Grok wrapped in 84 seconds. ChatGPT took 28 minutes. We unpack what the spread means — and what it doesn't.
Grok closed its portfolio submission in 1 minute 24 seconds. ChatGPT took 28 minutes 5 seconds. The other three landed in between: DeepSeek at 6:35, Claude at 14:46, Gemini at 16:41.
We've already seen people read that spread as a quality signal — Grok's lazy, ChatGPT's thorough. Don't.
Reasoning time, as Season 0 captured it, is the wall-clock duration from prompt-send to portfolio-submission. It includes the model's actual deliberation, but it also includes tool calls, web searches, drafting, self-correction loops, and (in Grok's case) parallel agent runs. There is no clean apples-to-apples here. Comparing reasoning time across contestants is like comparing chess engines by how long they spent on a single move — useful as a curiosity, dangerous as a leaderboard.
What reasoning time does seem to track is conviction.
- Grok 4.3 (1:24): Ran parallel agents, surfaced a tight, fully-formed thesis ("the binding constraint has shifted decisively from silicon to reliable baseload power"), and committed. 15 positions, max 12% in any name. Concentration without hesitation. Submission read like someone who already knew the answer.
- ChatGPT 5.5 Thinking (28:05): The longest run by an order of magnitude. The construction memo cites Reuters macro pieces, TSMC's investor transcript, individual earnings calls. 15 positions arranged into an explicit "barbell" with three sleeves — silicon/EDA, hyperscaler/cloud/software, grid/power. The submission reads like someone who wanted every weight defended in writing.
- Claude Opus 4.7 (14:46): 21 positions across four authored layers, no position at the 15% cap, explicit Sharpe-and-drawdown framing. Middle-of-the-pack time, middle-of-the-pack concentration — a portfolio built to not lose, before it was built to win.
- Gemini 3 Pro (16:41): 15 positions, three names at 10% or more, a deliberately heavy-tilt thesis (cooling and power as the bottleneck, not compute). The longest run among models that didn't iterate via parallel agents.
- DeepSeek V3 Expert (6:35): 24 positions — the most diversified portfolio in the field. The shortest time among models that produced a clean first-submission. Built for breadth, not for a single big bet.
If a six-month return tells us nothing, reasoning time tells us less. But the pattern in those numbers — Grok concentrated and fast, ChatGPT diversified and slow, DeepSeek diversified and fast — is the closest thing to a personality signal we have on Day 3.
We'll see in November how much personality is worth.
