DeepSeek vs ChatGPT on an Impossible Math Question

I pasted the same impossible math question into DeepSeek R1 and ChatGPT. ChatGPT hit “cannot be determined” in 4 seconds. Wrong. DeepSeek chewed on it for 280 seconds, caught a copy-paste ambiguity, tried alternate readings of the formula, and landed on “A”. Correct. The 70x time gap is the whole point: on adversarial math, chain-of-thought reasoning beats fast pattern-matching, and that is the switching rule I now use.

The rig I used, and what it doesn’t prove

I pulled the item from a puzzle blog listing what it called the oddest math questions ever written. The answer key stated “A”, so I had a ground truth before either model saw the prompt. Same expression pasted into both chat windows. No system prompt tweaks. No temperature knob. Nothing else in the context window.

What this run doesn’t prove: nothing about average math accuracy across the field, nothing about consistency across retries, nothing about how either model handles the same trap when it’s phrased cleanly. One question, two models, one round. What it does show is a specific failure mode ChatGPT can hit and a specific mechanism DeepSeek uses to catch it.

ChatGPT answered in 4 seconds and got it wrong

I fed the expression to ChatGPT first. It read the question, walked through a short chain of algebra, and stopped at “D, the value cannot be determined”. Four seconds end to end. The trap in the question was a formatting choice that made one operator look ambiguous, and ChatGPT treated the ambiguity as a dead end rather than a lead to investigate. Confident answer. Wrong answer.

If you have ever pasted a slightly mangled formula from a PDF and got a clean “no solution” back, this is that failure mode. Fast pattern-matching hits the surface, calls the question ill-posed, moves on.

DeepSeek took 280 seconds and got it right

I tried DeepSeek second. The server was busy on the first attempt, which is worth flagging if you plan to lean on it inside a client demo. Expect a retry. On the second try the reasoning pane opened and stayed open for four and a half minutes.

Somewhere in the middle, DeepSeek landed on the same “D” ChatGPT had. Then it talked itself out of it and asked whether the copy-paste had garbled the formula. From the reasoning trace:

Maybe the person who typed this has typed it wrong.

That sentence is the whole story. R1 ran the expression under alternate readings of the ambiguous operator, tested each, discarded the ones that produced nonsense, and only committed to “A” once one interpretation held up. 280 seconds is not a bug. It’s the product. The mechanism R1 uses in that window is sustained reasoning that questions its own first guess, and that’s what catches the trap.

The switching rule I now use

The takeaway is not “DeepSeek is smarter than ChatGPT.” It’s that a 70x latency budget bought a correct answer on an adversarial input. If the question had been “what is 17 times 23”, ChatGPT’s four seconds would have been right and 280 seconds would have been dead time. Most of the math I hand an AI in a real working day is closer to the puzzle: something with a trap, a formatting quirk, or an assumption I’ve made without noticing.

The rule I now use:

• Quick single-step arithmetic, unit conversion, or a spreadsheet formula: ChatGPT is fine. Speed wins.
• Anything with an ambiguity, a suspected typo, or a “does this even make sense” gut check: hand it to R1 or another reasoning model and take the coffee break. The latency is the feature.

What would change my mind: seeing R1 hit the same false-confident “D” that ChatGPT did on a batch of ten adversarial questions, or seeing ChatGPT’s newer reasoning modes catch the copy-paste ambiguity in under 30 seconds. Either result would collapse the rule above. I’ll run the batch next.

For the wider field of reasoning-capable assistants I’ve tested, see DeepSeek alternatives and ChatGPT alternatives. For hands-on verdicts on individual models, browse the AI tool reviews.