Sam Altman says ChatGPT 5 is an expert. A small maths slip shows why it’s still not perfect

OpenAI CEO Sam Altman describes

ChatGPT 5

as “like talking to an expert in any topic.” The company calls it its most advanced AI yet, capable of reasoning across subjects with PhD-level proficiency. Yet, on launch day, the model made a basic decimal subtraction error that underscored a reality of AI development: no matter how sophisticated, the technology can still make small but visible mistakes.The equation was simple: 5.9 = x + 5.11The correct solution is: x=0.79GPT-5 initially answered: x=−0.21 before correcting itself when prompted.

Why did GPT-5 make the error?

This wasn’t a limitation of its ability to handle decimals — the maths itself is well within its capabilities — but a case of what AI engineers call

reasoning slips

. Models like GPT-5 don’t “see” numbers in the way humans do. Instead of performing calculations in a dedicated numerical processor by default, they generate answers based on patterns in training data, unless specifically prompted to calculate step-by-step.

In this instance, the model appears to have briefly reversed the subtraction order mentally, before correcting when asked again. This kind of mistake can happen when an AI tries to solve a problem “in its head” without breaking it down explicitly, much like a human making an off-by-sign error during mental maths.It’s a reminder that while GPT-5 is much stronger at reasoning than previous models, accuracy improves when users give it clear instructions to show its working.

A major leap in capability

Launched on Thursday night, GPT-5 replaces all previous OpenAI models, including the GPT-4 series and the o-series, combining general question answering, complex reasoning, and tool usage into a single system. It automatically decides when to switch to its deeper “thinking mode” for more complex problems.Altman compared the leap from GPT-4 to GPT-5 to Apple’s move from standard pixels to the Retina Display, saying: “It’s the first time it really feels like you’re talking to an expert in any topic.”OpenAI says GPT-5 is its best coding model yet, capable of building complete websites, apps, and games in a single prompt, and significantly better at following instructions without straying from the request.

Improved accuracy — but not infallibility

According to OpenAI, GPT-5 is 45% less likely to produce factual errors than GPT-4o in standard mode, and 80% less likely in “thinking mode.” It also has reduced hallucinations — where AI generates plausible-sounding but incorrect information.However, as the decimal slip showed, the model can still make small, low-level errors if it doesn’t explicitly calculate step-by-step. This is partly because GPT-5 is designed to prioritise conversational fluency and context over raw number crunching unless told otherwise.

Availability and pricing

GPT-5 is available in three versions:

All signed-in ChatGPT users now default to GPT-5, with automatic switching to “thinking mode” when the system detects a harder problem.

The decimal takeaway

GPT-5’s decimal error isn’t evidence that it lacks mathematical skill — in fact, it can handle advanced calculus, algebra, and statistical analysis with high accuracy. Rather, it’s a reminder that

AI reasoning

still depends on how problems are framed and processed.For everyday use, GPT-5 is a leap forward: more accurate, more capable, and far better at following instructions. But like any expert — human or artificial — it benefits from being asked to show its working. That’s often the difference between an answer that’s fast and one that’s right.