Early Reactions to OpenAI o3 and o4-mini: Tool Use, Math and Mistakes

OpenAI released o3 and o4-mini in April 2025, prompting strong early reactions. One early user describes o3 as near genius level and claims it never hallucinates. That is personal enthusiasm, not proof of reliability: these models can still make things up and make serious mistakes, including in medical contexts.

A displayed TrackingAI chart reports an o3 score around 136 on an online Mensa Norway quiz-based scale, above the models shown beside it. This is a result on that particular quiz and test procedure. Human IQ tests are designed and normed for people, so the number is not a validated human IQ or a general measure of the model’s intelligence.

What interests me most is the models’ ability to use tools iteratively while working on a problem. An early user praises the models’ responses to scientific and clinical questions, but those impressions do not establish medical expertise or accuracy. Answers in consequential settings still need qualified human verification.

A displayed long-context comprehension table shows o3 scoring highly at many tested lengths, with dips at some lengths. These are results for a specific evaluation and set of prompts, not a promise that every detail in any long context will be retained. Context-window size and effective comprehension are different questions.

One o4-mini example calculates a compound daily growth rate by writing and running Python during the response. I find the combination of reasoning and tool calls particularly useful for multistep tasks. The displayed explanation and code offer a way to inspect the calculation, although tool use does not guarantee a correct result.

Another commentator describes o3 as a major improvement and says it helped develop ideas about post-labor economics. That is an anecdotal assessment, not independent validation of an economic theory. The model can help formulate and explore ideas, but formulas and claims require scrutiny.

A reported location-guessing example gives o3 a Street View image and receives a regional guess in eastern Canada. I am impressed by the example, but one case does not show the system has solved geolocation or always outperforms people. Images can contain identifying clues, so sharing them can raise privacy concerns.

Another user reports that a model inferred a Chicago restaurant from a photograph of a meal. The explanation suggests the model used visual clues and possibly outside information, though the exact process is not established by the screenshot. A successful-looking answer can still be wrong, and identifying a location from images can affect privacy.

The model also has simple failures. One reported response says “strawberry” contains two instances of the letter r; the correct answer is three. Another user obtains the right answer from o3. The contrasting anecdotes show that a model can succeed on difficult tasks and still be inconsistent on simple ones.

Riley Goodside reports that o3 found a path through a 200-by-200 maze in one try and says he overlaid the output on the original to check it. That is an interesting multimodal example, not a comprehensive evaluation. Another post reports o4-mini high solving a recent Project Euler problem quickly; the claim and comparison depend on the specific setup.

A MathArena table shows o4-mini high at 100% on AIME 2025 II with a displayed cost of $3.16 in that evaluation. The video also refers to an overall math score and other AIME results. Strong scores on a limited contest test are notable but do not establish mastery of all mathematics or rule out exposure to related material.

A coding demonstration asks models to make balls move within hexagons. The o3 and o4-mini examples look good to me, while other displayed samples show objects escaping or disappearing. These are visual demonstrations under particular prompts; they do not replace code review, repeated trials or broader coding evaluations.

Artificial Analysis reported strong o4-mini high results on its intelligence and coding indexes, including gains on LiveCodeBench and SciCode compared with o3-mini. Its o3 evaluations were still in progress in the displayed post. Rankings and prices are historical snapshots from April 2025, and benchmark composition and settings matter.

The video notes a 200K context window for o4-mini, compared with larger windows available in some other models at that time. A chart compares total output tokens used for a benchmark. Fewer tokens can reduce cost or latency, but quality, tool calls, pricing and successful task completion also affect efficiency.

One final visual test asks the model to match named people in a drawing with their colors. The displayed answer assigns several colors incorrectly after a long response time. Together with the letter-counting example, it is a reminder that impressive math and coding scores do not make every visual or reasoning task reliable.