At 4:22 PM on 4 November, with Micron (MU) at $855.10, five AI apps were asked the same question from the same evidence: where will this stock be in a week? Claude (Opus 5.5), ChatGPT (GPT-6 Astra), Gemini, Grok and Mistral each looked at it four ways: price patterns, company fundamentals, the news, and a combined view.
A week later MU closed at $811.30, down 5.1%.
What each one said
One-week forecasts, as the expected move from $855.10:
| View | Claude | ChatGPT | Gemini | Mistral | Grok |
|---|---|---|---|---|---|
| Price patterns | down −2.4% ✓ | down −2.4% ✓ | flat +0.6% ✗ | flat +0.6% ✗ | — |
| Fundamentals | — | — | flat +0.6% ✗ | — | — |
| News | down −4.7% ✓ | down −4.1% ✓ | flat −0.0% ✗ | down −2.9% ✓ | — |
| Combined | down −3.4% ✓ | down −3.1% ✓ | flat +0.6% ✗ | flat −0.9% ✗ | flat +0.6% ✗ |
✓ means the direction was right and the price landed within tolerance; ✗ means it wasn't. A dash means no forecast for that view: only Gemini gave a one-week fundamentals forecast, and Grok's per-view forecasts weren't captured.
Four things worth noticing
1. The news view was the most right. Three of the four news forecasts said down, and all three were right on direction and close on price. Whatever moved Micron that week, it was in the headlines before it was in the chart.
2. Mistral talked itself out of the right answer. Its own news view said down 2.9%. Its combined forecast, which is supposed to weigh all the views, came out flat. It had the right call and averaged it away.
3. Gemini didn't really take a position. Flat, within a fraction of a percent, on every view it gave. That's a defensible stance when you don't know, but in a week where the stock fell 5%, “flat” is just wrong.
4. Confidence didn't help. Each forecast comes with the model's own conviction score. The right calls carried 32–46%. The wrong ones carried 40–50%. If anything, the more sure-sounding forecasts were the ones that missed.
What this proves
Nothing, on its own. It's one stock over one week, and next week the flat-callers might be right. Across all 171 graded forecasts so far, these AIs get the direction right 39% of the time, barely above the 33% you'd get by guessing among up, flat and down.
What one week can do is show how differently models reason from identical evidence, and that's the point of putting them side by side and grading them in public. The next MU forecast will be graded the same way, whatever it says.