## The estimates

![Headline estimates on an axis starting at zero: Claude 150, ChatGPT 175, Gemini 150.](https://hypandra.b-cdn.net/koureia/171cf024-c3d1-6bf3-71f9-5e9c5d4f632e/site-assets/4f23cb303a8844a8764ba842f3ef429ea1b1c7281876974ce268ded40226844e.webp)

**The key difference is the assumed weight of one piece.** Claude assumes **17 g per piece**; ChatGPT assumes **0.5 oz, or about 14.2 g**. For the same bag, those assumptions yield roughly **150 versus 180 pieces**. Neither average piece weight was measured from the bag.

Claude’s assumption is about **20% heavier**, which means fewer pieces for the same total weight. Using the label’s rounded weights: **2,550 g ÷ 17 g = 150**; **90 oz ÷ 0.5 oz = 180**. These are calculations from assumptions, not counts from the photograph.

**Gemini states no average piece weight.** It gives about **150** through a claimed typical product count, without a calculation or source. Its agreement with Claude’s number therefore does not establish agreement about piece weight.

The graph shows the **headline estimates**, not all three calculations: ChatGPT’s headline is **175**, 25 higher than the other two. Its calculation is **180**, and its final contest guess is **178**. The actual count is unknown.

The same photo and prompt in the default Claude, ChatGPT, and Gemini apps. Screenshots and full responses follow; the actual count has not been supplied.

## The prompt

The same photograph and prompt were submitted through the three apps:

> How many pieces of candy are in here? Estimated

The prompt asks for an estimate without saying what it will be used for.

## Different assumptions, different uses

**Claude: a planning number.** It assumes about **17 g per piece**, allowing for heavier bars and lighter M&M's packs. It estimates **150 pieces**, with a range of **130–170**, then introduces trick-or-treat planning: “150 is a safe number.” It also suggests counting the pieces for an exact answer. Trick-or-treat planning was not specified in the prompt.

**ChatGPT: an estimate, then a contest guess.** It assumes **0.5 oz per piece** (about 14.2 g), giving **180** by division. Its headline estimate is **175**, with a range of **160–190**. It then introduces a conditional candy-counting contest and ends with **178**. The response does not explain the adjustments from 180 to 175 or 178; the prompt did not mention a contest.

**Gemini: a typical count for an identified product.** It identifies a particular Kirkland variety mix and gives **145–160 pieces**, centering on **150**. It invokes a standard bag and an “official serving size count,” without showing a calculation or source. It does not introduce a separate use such as planning or a contest. The product identification and serving-count statement are claims from its response, not independently verified here.

**The different assumed average piece weights explain the roughly 30-piece gap between Claude’s estimate and ChatGPT’s calculation.** Gemini presents its answer as a typical product count. Claude and Gemini both arrive at about 150, through different stated reasoning. ChatGPT gives three distinct point values across its headline, calculation, and final suggestion.

These are observations about the responses shown here. They do not establish which estimate is closest to the actual count.

## The base image

![Original photograph of a Favorites assorted candy bag labeled net weight 2.55 kg (5.6 lb), 90 oz.](https://hypandra.b-cdn.net/koureia/171cf024-c3d1-6bf3-71f9-5e9c5d4f632e/site-assets/9dd03fe5544c36f31b099cbe21a577cdf6ccfe425269225a7b0e7c965b34c914.webp)

The label gives the bag's weight and says the types and amounts of candies will vary.

## Claude

![Claude app screenshot: about 150 pieces, with a plausible range of 130 to 170.](https://hypandra.b-cdn.net/koureia/171cf024-c3d1-6bf3-71f9-5e9c5d4f632e/site-assets/e04e10c373c1f805af3d0c44626be793e58b53374a3dc7a8e443b521843f47a7.webp)

*Full visible response text, transcribed from the screenshot. The collapsed activity label reads “Estimating how many pieces fit in a bag.”*

> **About 150 pieces**, with a plausible range of 130 to 170.
>
> The bag is 90 oz (2.55 kg). Fun-size bars like Milky Way, Twix, KitKat and Reese's average roughly 15 to 20 g each, and the small M&M's packs run a bit lighter. At about 17 g per piece, that works out to around 150. The label says the mix varies, so your count will shift with how many heavier bars are in there.
>
> For an exact count, empty the bag and count it. For trick-or-treat planning, 150 is a safe number.

## ChatGPT

![ChatGPT app screenshot: headline estimate 175 pieces, calculation 180, and closing contest guess 178.](https://hypandra.b-cdn.net/koureia/171cf024-c3d1-6bf3-71f9-5e9c5d4f632e/site-assets/6f4deea533003367a5144052acdf94865376b987bd8cb6f307486db08c7afdbe.webp)

*Full visible response text, transcribed from the screenshot.*

> **Approximately 175 pieces**
>
> My best estimate is **160–190 pieces**, with **175** as my guess.
>
> The bag weighs 90 oz (5.6 lb) and contains assorted fun-size chocolates: Snickers, Twix, Milky Way, M&M’s, and others.
>
> Assuming an average of about 0.5 oz per piece:
>
> 90 ÷ 0.5 = **180 pieces.**
>
> **If this is a candy-counting contest, I'd guess 178.**

## Gemini

![Gemini app screenshot: approximately 145 to 160 pieces, averaging around 150.](https://hypandra.b-cdn.net/koureia/171cf024-c3d1-6bf3-71f9-5e9c5d4f632e/site-assets/8181cfb18a6726e4ae761c10063e4fbbac2683ee636e9c0766ca4c2edc4a2935.webp)

*Full visible response text, transcribed from the screenshot.*

> This is the Kirkland Signature Chocolate Fun Size & Minis Favorites variety mix (90 oz / 5.6 lb).
>
> A standard bag of this size contains approximately **145 to 160 pieces** of assorted candy, with the official serving size count typically averaging right around **150 pieces**.