Every one of them said its data was live. Then three of the five forecast a $760 ETF into a range around one dollar.
This started as a typing mistake, which is the only reason it is interesting. Our prompts used to hand the model a baseline price — the number a forecast would be measured from — and on one run for SPY that field was filled in as $1.00. SPY was trading near $760.
Five models got the same prompt, with the same wrong number in it. Here is what came back:
| Model | Forecast range | Confidence | Declared data access |
|---|---|---|---|
| Claude — Opus | $746 – $785 | 50% | live |
| DeepSeek | $749 – $785 | 45% | live |
| Gemini — Pro | $0.98 – $1.03 | 55% | live |
| ChatGPT — GPT-5 | $0.975 – $1.015 | 67% | live |
| Grok | $0.95 – $1.05 | 25% | partial |
Two models disregarded the baseline and forecast the asset. Three built a careful, well-hedged, entirely fictional range around a price the asset has never traded at. The most confident answer on the page — 67% — is one of the three.
They were not being stupid. They were being obedient.
The prompt told them to. It said, in as many words, to measure the range from the baseline even where their own sources disagreed, because a range built on a different anchor would be scored against a price the model did not use.
That instruction was ours, and it was wrong. Claude and DeepSeek broke it. The other three followed it off a cliff.
But the more useful thing in that table is the last column. All five declared their data access as live or partial. That field is self-reported — we ask the model to say whether it retrieved anything current — and on this run it told us precisely nothing about who had.
So we stopped telling them the price
The prompt no longer carries a baseline. Establishing what the asset is trading at is now the first part of the task, the model reports the figure it found, and that figure is compared with what the asset actually cost.
It changes what a wrong forecast means. A range that lands near the outcome from a starting point that was wrong is luck, not analysis, and it is no longer scored like analysis: whatever a forecast earns is discounted by how far the model's own price was out. Under a percent costs nothing — delayed quotes and extended-hours prints are not research failures. Two percent halves it. A model that thinks a $760 index costs a dollar keeps almost none of its score.
What the first readings look like
Three questions so far, five models each, no baseline given. The pattern is not the one we expected:
| Asset | The agreed figure | Models that differed |
|---|---|---|
| IONQ | $39.20 | ChatGPT — GPT-5 at $39.42, DeepSeek at $39.35 |
| SPY | $769.35 | DeepSeek at $773.10 |
| SOXL | $111.34 | ChatGPT — GPT-5 at $109.21 |
Most of the models return the same figure to the cent — almost certainly the same last close from the same widely syndicated source. What varies is which one drifts, and by how much: 0.5% on the index, close to 2% on the leveraged ETF, where 2% is a normal hour.
Note that no model is reliably the outlier. GPT-5 is furthest out twice and dead on once. On this much data that is a coin, not a ranking — which is the whole reason for running it repeatedly and publishing as we go.
Why this is the first thing we measure
Everything else a model says about an asset is downstream of knowing what it costs. A confidence interval, a directional call, a paragraph of reasoning about support levels — all of it is decoration on a number that was either retrieved or invented, and until now there was no way to tell which from the outside.
It is also the cheapest honest test in this whole field. There is no judgement involved, no waiting for the market to settle the argument, and no room to grade generously: the asset had one price, and either the model found it or it did not.
👉 How a forecast is scored, in full
👉 Run this on your own symbols — 15 prompts a month, free
Fifteen answers across three questions is not a finding, it is a starting point, 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.