Give five models the same prompt about the same asset over the same window, and the thing that varies most is not the forecast. It is the handwriting.
Asked about IONQ over one day, Claude wrote 8,383 characters of reasoning. Gemini wrote 1,776 for the identical question. That is not a disagreement about IONQ — it is a difference of temperament, and it is stable: across eighteen shared forecasts on five assets over four horizons, each model files in roughly the same shape every time.
Five signatures
| Model | Median rationale | Range width vs the five | Confidence band | Declared less than "live" |
|---|---|---|---|---|
| Claude — Opus | 7,834 chars | 106% | 40–64% | 4 of 18 |
| DeepSeek | 4,108 | 115% | 40–65% | 3 of 18 |
| ChatGPT — GPT-5 | 3,886 | 96% | 63–78% | 6 of 18 |
| Grok | 2,320 | 106% | 28–53% | 7 of 18 |
| Gemini — Pro | 1,808 | 78% | 50–80% | 3 of 18 |
Read down the columns and each model resolves into a posture you could describe without mentioning a single price.
Claude — Opus writes four times as much as Gemini while sitting near the bottom of the confidence table. Most verbose, least willing to claim certainty — the opposite of what long output usually signals.
DeepSeek draws the widest ranges of the group and is the only one of the five that reliably takes a side. More on that below.
ChatGPT — GPT-5 was the most confident of the five on sixteen of eighteen questions and never dropped below 63%. Its ranges are slightly tighter than average, so it commits on both axes at once.
Grok is the least confident by a wide margin — 39% on average, and lowest of the five on seventeen of eighteen questions. It is also the most willing to admit its data was not current: seven declarations below "live", the most of any model.
Gemini — Pro writes the least and draws by far the tightest ranges — 78% of the group average, and the tightest of the five on eleven of eighteen questions — while sitting second in confidence. It is the most assertive forecaster on the measure that actually matters, and it says the least about why.
Four of the five will not take a side
The prompt is explicit about this. It tells the model that anchoring a range to the price it found does not mean centring on it, and that a symmetric range asserts no directional view — a legitimate answer, but one to give on purpose rather than by default.
Here is where each model actually put the midpoint of its range, relative to the price it had just reported for itself:
| Model | Average offset | Largest offset | Offset by 3% or more |
|---|---|---|---|
| DeepSeek | +2.07% | 12.3% | 5 of 21 |
| Gemini — Pro | −0.71% | 6.5% | 3 of 20 |
| Claude — Opus | −0.74% | 3.3% | 1 of 19 |
| ChatGPT — GPT-5 | +0.06% | 2.4% | 0 of 19 |
| Grok | −0.04% | 3.1% | 1 of 20 |
GPT-5 has never once, in nineteen forecasts, placed its range more than 2.4% to either side of the price it reported. Grok's midpoint sits an average of four hundredths of a percent from its own quote. These are ranges centred on the present, which is a statement about volatility, not about direction.
Read strictly, four of these five models spent eighteen forecasts declining to say which way they thought an asset would go — while writing thousands of characters about catalysts, momentum and regime.
DeepSeek is the exception, and it is a real one. On SOXL over seven days it filed $115–135 while the other four all put their floor below $99:
| Model | Range |
|---|---|
| Claude — Opus | $90.50 – $129.00 |
| Grok | $95.00 – $125.00 |
| ChatGPT — GPT-5 | $98.00 – $120.00 |
| Gemini — Pro | $98.00 – $118.00 |
| DeepSeek | $115.00 – $135.00 |
That window has not closed, so whether the call was right is still open. But it is the only one of the five that is unambiguously a call.
SOXL 7 days Prediction - 2026-08-29 22:21 UTC
The style is the model's, not the question's
The obvious objection is that these are just different questions producing different answers. They are not. Every number here comes from questions all five models answered simultaneously, from the same prompt, over the same window — so anything that differs between them is the model.
And the postures survive conditions that ought to disturb them. SPY over one day is a quiet, liquid, heavily covered question; FAS over thirty days is a triple-leveraged financials ETF over a month. Every model widens its range for the harder question, sensibly. Their relative positions barely move. Across the eighteen questions:
- Claude wrote the longest rationale 18 times out of 18.
- Grok declared the lowest confidence 17 times out of 18.
- GPT-5 declared the highest confidence 16 times out of 18.
- Gemini drew the tightest range 11 times out of 18.
Whatever produces "how sure a model sounds" and "how much it writes" is set upstream of the question.
It is also close to deterministic. We asked the same FAS 30-day question twice, eight minutes apart. GPT-5 returned $145–215 both times; Claude returned $162–194 both times; DeepSeek returned $155–210 both times — to the dollar. Three of the five gave a byte-identical answer to a question asked twice.
Why length is not quality
It is tempting to read 8,000 characters as diligence and 1,800 as laziness. The settled results do not support that reading cleanly in either direction. Claude, the longest writer, landed in range 8 times out of 10 — the best of the five. But Grok, at under a third the length, managed 7 of 10, and what separates both from Gemini's 5 of 10 is range width, not word count.
What length does buy is auditability. A model that names its sources, its levels and its counter-argument in 8,000 characters has given you something to check. One that gives you a range and a paragraph has not. That matters more than it sounds, because a forecast you cannot audit is one you have to take on trust — and the first thing we published here was a case where five models' stated data access told us precisely nothing.
What to do with this
- Pick models for complementary postures, not for a leaderboard. One tight range and one wide one on the same question tells you more than two of either.
- Calibrate each model against its own history. Grok at 53% is Grok being unusually sure. GPT-5 at 63% is GPT-5 at its most hesitant. The same number means opposite things.
- Treat a long rationale as evidence to check, not as evidence of care. The value is in the levels and sources you can verify, not the word count.
- Do not mistake a centred range for a forecast. If the midpoint sits on today's price, the model has told you how much it expects the asset to move, not where.
Every figure here was computed from shared forecasts you can open and read — the same ranges, the same reasoning, the same settlements.
👉 Read the shared forecasts these numbers came from
👉 Run this on your own symbols — 15 prompts a month, free
Eighteen questions per model is a starting point, not a finding, and we will keep saying so until the sample says otherwise. Content on ForecastNest Insights is for research, educational and informational purposes only. Nothing here is financial, investment, legal or tax advice, and nothing here is a recommendation to buy or sell anything. We publish how models perform on a disclosed task; we do not publish price targets or trading signals.