Seven questions to ask about any accuracy claim, and what ReasonEra's published figures do and do not show.
Treat any single accuracy number with caution. A useful claim states the question types, the sample size, where the items came from, how it was tested, the date and whether anyone independent checked it. ReasonEra reports 98%+ on Matrigma-style matrices and published a benchmark with category ranges; those are its own figures and have not been independently verified.
| Question | Why it matters |
|---|---|
| Which question types does it cover? | An overall score can hide weak categories |
| How many items were tested? | Small samples give unreliable percentages |
| Where did the items come from? | Public analogous items are safer than proprietary ones; made-up easy items inflate scores |
| How was it tested? | Blind testing with identical inputs avoids bias |
| Which models were compared? | Old versions make a comparison look better than it is |
| When was it measured? | Models change quickly, so a date is essential |
| Did anyone independent check it? | Self-reported figures are a starting point, not proof |
| Category in the benchmark | General models tested | Specialised models |
|---|---|---|
| Dynamic and game-based reasoning (Aon, Arctic Shores style) | 43–64% | 92–97% |
| Advanced matrix reasoning (Raven's, Matrigma, Alva Labs style) | 51–76% | 96–99% |
| Deductive and verification logic (SHL, Talent Q style) | 54–77% | 94–97% |
These numbers come from ReasonEra's own benchmark of 30 November 2025. It used 500+ analogous public items from 15 providers, with blind testing, identical inputs and no proprietary items. The general models (GPT-4o, Claude Opus 4.5, Gemini 2.0 Pro, GPT-o1, Grok 2) were the versions available then and have since been updated. No independent party has verified the results. Read the benchmark.
ReasonEra also reports 98%+ accuracy on Matrigma-style matrices.
Accuracy below 100% is expected. An explanation that looks wrong is worth checking, and a tool that shows its reasoning lets you do that. Be wary of any claim of perfect accuracy.
Use free trials on questions you care about and compare tools on the same set. See our guide to testing an AI yourself.
ReasonEra is designed for practice and learning. Employers set their own conditions for assessments, and many forbid outside help during a real test. Check your invitation and the assessment rules before you use any tool, and use ReasonEra to build the skill, not to replace it.
ReasonEra reports 98%+ accuracy on Matrigma-style matrices, and its published benchmark reported 92–99% for specialised models across dynamic, matrix and deductive categories. These are ReasonEra's own figures from 2025 and have not been independently verified.
Treat it as a starting point. Look for question types, sample size, item sources, method and date, and test the tool on your own questions.
Blurry images, ambiguous puzzles, fast-moving game elements and misread details can all cause errors.
ReasonEra's benchmark cited roughly 85–95% for 99th-percentile humans on advanced matrix items. This is a figure from that write-up, not an independent source.
Models change quickly, so any benchmark is dated. Check the date on every accuracy claim, including ours.
Free trials come with every account. If it helps, a 14-Day Pass for 49 € removes the limit for two weeks.
You pay through Paddle, and the Refund & Cancellation Policy covers refunds.
Sources and notes. Accuracy numbers here are taken from ReasonEra's November 2025 benchmark and have not been independently checked. The examples are original to this page, not real test items. Send corrections to info@reasonera.com. No test publisher mentioned is affiliated with ReasonEra.