GPT vs Claude vs Gemini: the same 3 prompts, real answers

I sent three short prompts to claude-sonnet-5, gpt-5.4-mini, and gemini-3.5-flash through Concentrate on 2026-09-11 and pasted the answers below exactly as they came back. All nine calls succeeded. The headline result: all three agreed on the facts every time, GPT was the shortest and cheapest, Claude was the most detailed, and Gemini spent roughly 500 to 900 hidden reasoning tokens per answer, which made it the slowest and most expensive by a wide margin.

Ground rules. This is n=1. Every number is one run. Run it again and the words change. The numbers come straight from the API response: latency is wall clock from my laptop, tokens and cost are what Concentrate reported. Output cap was 4,096 tokens and nothing got cut off. No system prompt.

Prompt 1: "What is the currency in Mesopotamia? Answer in under 60 words."

claude-sonnet-5

Ancient Mesopotamia did not have coined currency. Instead, they used a barter system supplemented by commodity money—primarily silver (measured by weight, often in units called shekels) and barley. These served as standards of value for trade and payments. Actual minted coins did not appear until much later, introduced by the Lydians around the 7th century BCE.

gpt-5.4-mini

Mesopotamia did not have a single official currency like modern money. People mainly used silver by weight as a medium of exchange, along with barter. In some periods, goods like barley also functioned as payment.

gemini-3.5-flash

Ancient Mesopotamia did not use minted coins for most of its history. Instead, they relied on a commodity system where barley (for everyday transactions) and silver (for larger trade) served as currency. The standard unit of value was the shekel, which represented a specific weight of barley or silver.

Same facts three times: no coins, silver and barley by weight. The differences are all in what each model chose to add. Claude added the Lydians and a date, which is the kind of detail that makes the answer feel authoritative and is also the kind of detail I would want to check. Gemini added that the shekel was a weight unit for both commodities. GPT added nothing and came in at 52 tokens. All three bolded the key words without being asked.

The thing that surprised me was the Gemini bill. 505 reasoning tokens to write 65 visible words. The answer is fine. It is not 20 times better than GPT's, which is what it cost.

Prompt 2: "Is a hot dog a sandwich? One sentence."

claude-sonnet-5

A hot dog fits the basic definition of a sandwich (filling enclosed in bread), but many people consider it its own distinct category due to cultural convention and its unique form.

gpt-5.4-mini

Yes—if you define a sandwich broadly as filling between bread, a hot dog counts, though some people dispute it.

gemini-3.5-flash

While a hot dog is technically a sandwich because it consists of filling nestled inside a split roll, culturally and culinarily it is widely regarded as its own distinct category of food.

Nobody committed. All three said "technically yes, culturally no" in one sentence. GPT is the only one that led with the word "Yes." Claude and Gemini both put the definition first and the hedge second, and they are close enough that the word diff on the site lights up mostly on synonyms: "enclosed in bread" versus "nestled inside a split roll."

This is post-training talking. Three labs, three sets of raters, and all three landed on the same diplomatic shape. I asked for one sentence and got one sentence from everyone, which was not a given.

Prompt 3: "Write a haiku about a database migration."

claude-sonnet-5

Old schema falls silent—
new tables rise from the dust,
data flows again.

gpt-5.4-mini

Tables shift in silence
Old keys bloom into new forms
Night commits, dawn sings

gemini-3.5-flash

Tables mapped anew,
Data flows through silent pipes,
No rows left behind.

Two of the three used the word "silent" or "silence." Two used "flows." Everyone has apparently decided database migrations happen quietly at night, which matches my experience only when they go well. GPT's is the most haiku-like to my ear. Gemini's "No rows left behind" is the only line that reads like someone who has actually run one.

Gemini spent 905 reasoning tokens on 17 syllables. Eight seconds, 75 times the cost of GPT's. The haiku is fine. I do not know what it was thinking about.

What I took from this

• On easy factual prompts the three agree on facts and differ on garnish. Length is the biggest visible difference, and it tracks cost almost exactly.
• Reasoning mode is on by default for the Gemini flash model through this router, and it costs 500 to 900 hidden tokens per call even when the prompt does not need it. Check the reasoning badge before you compare prices.
• GPT mini was the fastest and cheapest on all three prompts by a wide margin, and it never lost on facts. For a lot of workloads that is the whole story.
• The "via" field showed anthropic/, azure/, and vertex/. Three model IDs, three different clouds. Worth knowing before you conclude anything about "the model."

Run them yourself at fiddlellm.com. The Mesopotamia prompt is already in the box. Turn on the diff, pick GPT as the base, and watch the green light up.