SEMI-SERIOUS QUESTIONS. FULLY COMMITTED EGO BOOSTS.
A little context.
The questions are real.
I’m Eric Sun. I wrote ten English questions and gave seven AI models a closed-book exam. This game features two: each was solved by one model; none solved both. Results are from 6 September 2026.
I designed the benchmark and directed the project. ChatGPT built the website and generated its illustrations.
Scoring is strict: select every correct choice and no incorrect ones.
The full ten-question results
| Tested model | Correct out of 10 | Confidence (−40 to +40) |
|---|---|---|
| Claude Opus 5 (Max) | 2 | +29 |
| Claude Fable 5.1 (Max) | 1 | +27 |
| Gemini 3.1 Pro (Extended) | 1 | +40 |
| GPT-6 Astra (Max) | 1 | +16 |
| Gemini 3.8 Flash (Extended) | 1 | +36 |
| Claude Opus 4.8 (High) | 0 | +19 |
| GPT-5.6 Sol (High) | 0 | +28 |
How the confidence bars work
Models estimated their accuracy across 40 answer-choice decisions: +1 for each right classification, −1 for each wrong one. We convert that confidence score C into an estimated question score: 10 × ((C + 40) / 80)4.
This assumes four equally reliable, independent choices per question. It’s an estimate, not a direct prediction from the model. For example, +29 becomes about 5.5 out of 10.
The IQ chart takes liberties.
Model scores come from TrackingAI’s Offline Test, frozen on 6 September 2026. Some use related models or different thinking settings as proxies. These aren’t human-equivalent IQ measurements.
Your score? The highest reference you beat, plus 7 for one correct answer or 14 for two. Pure celebratory arithmetic.
Reference models and sources
| Game model | Reference | Score |
|---|---|---|
| Opus 5 | Claude-5-Opus-Max · Offline Test; 4 runs | 128* |
| Fable 5.1 | Claude-Fable · Offline Test; latest run; version/settings proxy | 130* |
| Gemini 3.1 Pro | Gemini Pro · Offline Test; 7 runs; settings proxy | 122* |
| GPT-6 Astra | GPT-Sol-Ultra · Offline Test; 7 runs; different-model proxy | 130* |
| Gemini 3.8 Flash | Gemini Thinking · Offline Test; 7 runs; older Flash proxy | 128* |
| Opus 4.8 | Claude-4-Opus-Extended · Offline Test; 7 runs; version/settings proxy | 118* |
| GPT-5.6 Sol | GPT-Sol-Ultra · Offline Test; 7 runs; Ultra rather than High | 130* |
We follow TrackingAI’s non-vision scoring and averaging rules. Astra uses Sol Ultra; Flash uses an older Gemini Thinking series. Exact runs and calculations are in the snapshot below.
Offline snapshot · Original data · Model details
Older shared cards retain their original scores. Previous Mensa Norway snapshot.
And the math?
The example comes from 2025 AIME I, Problem 5. The worked response is from Gemini 3.1 Pro (Extended); its result was checked by enumeration. Separate math benchmarks illustrate models’ strengths; we didn’t test all seven on this exact problem.
The small print, kept small.
Custom creators choose their answer keys. Challenges last seven days; exports keep the questions and answers. The leaderboard resets Mondays at 00:00 UTC; the secret guestbook covers the last seven days.
An anonymous cookie remembers your run. Revealing answers closes it. We count plays and sharing actions, limit spam, and publish names only when you add them to a board or shared card. No advertising trackers or live AI calls.
Visual inspiration: Any Human Ever.