ChatGPT vs Gemini vs Claude, Which AI Assistant Wins? (Tested on the Free Tiers)

For most people the AI assistant question comes down to three names: ChatGPT, Gemini and Claude. They are all genuinely good, which is exactly why choosing between them is hard. So instead of comparing spec sheets, I sat down and gave all three the identical set of tasks, on their free tiers, on the same afternoon, and watched what they actually did.

This was a free-tier test on purpose, because that is where most people start. As of August 2026, Claude served me Sonnet 5 on its Medium setting, Gemini defaulted to Flash, and ChatGPT did not show its model name anywhere but offered a "Think" toggle for slower, smarter answers, which I left off to keep everything basic and free. Same prompts, pasted verbatim, a fresh chat for each.

First, the friction of logins and models

Before a single answer, the three already differed. ChatGPT let me start with no account at all, though it nudged me to sign in with the Google account my browser was using. Gemini simply logged me in automatically, since the browser was already tied to a Gmail account. Claude was the only one that required me to sign up before it would say anything. None of this is dealbreaking, but if you want to open a tab and just go, ChatGPT and Gemini are frictionless in a way Claude is not.

Round one - the awkward email

I asked all three to write a firm-but-polite 120-word email to a landlord about a damp bedroom that was starting to affect my health, with no clichés and specifically not to open with "I hope this email finds you well."

All three obeyed the no-cliché rule and landed near the length. The differences were in texture. Gemini’s was the firmest and ran a little long. ChatGPT opened its draft inside an editable canvas, which is handy if you want to tweak before sending, and signed off politely. Claude’s was the one I would actually send. It invented the most convincing specifics, then offered to add dates or photos as evidence.

That reads the least like a robot and the most like a tenant who is quietly fed up.

Round two - the logic puzzle

Next, the classic two-ropes puzzle: each rope burns for 60 minutes but unevenly, so half a rope is not 30 minutes. Measure exactly 45 minutes with only the ropes and a lighter.

All three got it right, light one rope at both ends and the other at one end, then light the second rope’s other end the moment the first burns out. What differed was the teaching. ChatGPT was the most concise and finished with a bolded one-line answer. Gemini laid it out with tidy "time elapsed" markers at each step. Claude led with the underlying insight first, that burning a rope from both ends always halves its time no matter how unevenly it burns, which is the bit that actually makes the trick click. If you want to understand rather than just copy, Claude explained it best.

Round three - the buggy function

Then a small piece of real code: a Python function meant to return the second-largest distinct value, but written as sorted(nums)[-2], which quietly breaks on duplicates and on very short lists.

Every one of them spotted both bugs, that duplicates are not removed and that short lists need handling. The craft is in how they fixed it. Gemini returned None for the edge case, gave the prettiest, most readable code block with a download button, and offered to write unit tests. ChatGPT gave a compact one-line alternative alongside the safe version and flagged the trade-off, and it has Run buttons to execute the code in place. Claude was the most defensive, raising a clear ValueError and, unprompted, writing a little test I had not asked for.

second_largest([5, 5, 3])     # 3  (not 5)
second_largest([1, 2, 3, 4])   # 3
second_largest([7, 7, 7])     # raises ValueError
second_largest([9])           # raises ValueError

Three correct answers, three different instincts, cautious from Claude, pragmatic from ChatGPT, polished from Gemini.

Round four - read a long essay and push back

Finally the hardest task. I pasted a 1,500-word philosophy essay and asked for exactly five bullet points of under fifteen words each, then for any claims it thought were unsupported.

All three nailed the format, five tight bullets, and, crucially, none of them invented fake claims to criticise; every flag pointed at a real sentence in the text. But their reading depth differed sharply. Gemini was the most lenient, flagging only three claims. ChatGPT and Claude each flagged five. Claude was the sharpest reader of the lot, the only one to catch a smuggled-in biographical detail rather than just the philosophical overreaches.

That is the kind of catch that separates a careful reader from a fast one. If you want an assistant that argues back properly, Claude read most critically.

One telling difference in handling: ChatGPT and Claude both quietly filed my giant pasted text as a markdown attachment to keep the chat clean, while Gemini dropped the whole thing inline. A small thing, but Gemini’s chat was the messier one to scroll afterwards.

The bit the spec sheets miss - how they feel

After four rounds a clear pattern held. On correctness it was a tie, nobody got anything wrong, which matches what I keep finding: for everyday tasks these models have converged, and the real question is feel and fit. Claude’s prose was consistently the most human and its answers the most elegant. Gemini’s writing was the most robotic, but its interface was the nicest to actually read, the best-formatted code, the clearest numbered steps, the download buttons. ChatGPT sat in the middle as the confident all-rounder, with the editable canvas and the escape hatch of Think mode if you ever need more.

Prices here were verified in August 2026. Check the live pricing comparison and the pricing changelog for the current numbers, or use the tool recommender to match an assistant to your task.

For the writing-focused head to head, see Claude vs ChatGPT for writing. And if you are wondering whether any of them can be trusted with code, I put seven free chatbots through a coding gauntlet and none of them fell.