What general AI assistants can and cannot do on aptitude and reasoning questions, what a published benchmark found, and how to test it yourself.
General AI assistants such as ChatGPT can attempt every aptitude question type from an image, but reliability varies. On visual reasoning, ReasonEra's benchmark found general models scoring 43–77% where specialised models scored 92–99%. Numerical and verbal questions were not part of that benchmark, so test any assistant on your own questions.
Sometimes. A general assistant can read an image of a puzzle, describe it and propose an answer. How often that answer is right depends on the model version, the image quality and the question type.
Matrix, abstract and diagrammatic questions depend on spotting rules about position, orientation, count and shading. According to ReasonEra's benchmark write-up, general models tend to describe what they see, while a specialised model is trained to model the rules. A description can contain a wrong interpretation, and a wrong rule leads to a confident wrong answer.
| 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.
The benchmark did not cover numerical or verbal questions, so we do not quote a figure for them. A general assistant can read a table or a passage and answer, but numbers and units can be misread, so check the working. See the numerical and verbal solver pages.
ReasonEra is built only for pre-employment tests: you upload the image, choose Quick Verify or Tutor Explanation and get a structured answer with no prompt to write. ChatGPT is more flexible. See the full ReasonEra vs ChatGPT comparison.
Treat ReasonEra like a tutor you consult while studying. In a proctored or unaided assessment, the tutor stays at home.
It can attempt them from an image, but accuracy varies by model version and question. ReasonEra's published benchmark reported general models at 51–76% on advanced matrix reasoning, compared with 96–99% for specialised models. These are ReasonEra's own figures from 2025.
It can attempt them, but check the numbers and the passage carefully. ReasonEra's benchmark did not cover these question types.
If a model misreads a visual detail, it can build a consistent but wrong rule and present it confidently. Check the reasoning, not just the answer.
Possibly. The benchmark tested the models available in November 2025. General models have been updated since, so test current versions on your own questions.
Employers set their own rules and many forbid outside help during a real test. Check your invitation, and use AI to practise.
Pick a question that has been stumping you and upload it. The first three tries are free with an account.
Payment is taken by Paddle; see the Refund & Cancellation Policy for how refunds work.
Sources and notes. Where this comes from: ReasonEra's published benchmark supplies the accuracy figures, and it has not been independently verified. Examples are original and not from real tests. Please send corrections to info@reasonera.com. ReasonEra works independently of the test publishers named.