The Umami Chocolate Chip Cookie
Gemini won the math and the room. Claude had its first — and worst — loss. And the recipe with three umami sources produced less umami flavor than the one with one.
The Scoreboard
Unlike Episode 2, there was no twist. Gemini led on every metric. Claude came in last in every category — a clean sweep in the wrong direction for the two-time defending champion.
Average scores by criterion
13 tasters · scored 1–10 on Taste, Texture, and Smell
Composite averages
| Cookie | Taste | Texture | Smell | Favorites | Want Again | Composite |
|---|---|---|---|---|---|---|
| Gemini — The Purist | 7.54 | 7.31 | 8.08 | 8 / 13 | 92% | 7.64 |
| ChatGPT — The Alchemist | 6.85 | 7.46 | 6.92 | 4 / 13 | 77% | 7.08 |
| Claude — The Maximalist | 5.69 | 6.46 | 6.62 | 1 / 13 | 54% | 6.26 |
No Twist This Time
Every democratic measure pointed the same direction. Gemini won the composite, won the most favorites, had the highest want-again rate, and won the most per-taster head-to-head matchups.
Favorites & Want-Again Rate
Per-Taster Breakdown
One mathematical tie: Taster 2 had Gemini and ChatGPT exactly equal at 7.333. The taster chose Gemini as their favorite.
The Salt Paradox
Claude's recipe had four sodium-bearing ingredients: white miso, soy sauce, MSG, and kosher salt. At those combined concentrations, tasters defaulted to the most familiar interpretation of the signal. Six of 13 independently described Cookie B as salty without knowing what was in it.
The umami paradox: Stacking umami sources is a real technique — miso and soy sauce together appear in serious professional cooking. The difference is that in most applications they function as background enhancers in small amounts. Claude's recipe had four separate sodium-bearing ingredients simultaneously. The glutamate pathway and the sodium pathway are not the same thing — but at those concentrations, tasters' brains defaulted to the most familiar interpretation of that signal. The cookie that tried the hardest to taste like umami tasted the least like umami.
Claude's "salty" descriptor — vs all other cookies
Number of tasters who independently described each cookie as salty (without knowing the recipe)
The Espresso Signal
ChatGPT was the only recipe to include espresso powder (1½ tsp) and mushroom powder (¼ tsp). Two tasters flagged coffee or espresso as a flavor note, completely independently.
Why ChatGPT reads as coffee: Espresso powder adds roasted aromatics that survive baking. Mushroom powder contributes glutamates but also earthy terpene notes. Two tasters read the combination as coffee; one found it distinctive and desirable. The same ingredient that was a calling card for some people barely registered for others — the alchemist approach had a distinct identity that split opinion.
The Language of Each Cookie
Gemini's words were aspirational. Claude's were about one thing. ChatGPT's split between those who caught the complexity and those who didn't.
"Aggressive" and "Disappointment" were written by the same taster, about the same cookie.
Who Was Tasting
13 tasters. The demographic breakdown revealed two notable splits — by gender and by age group.
Favorites by gender
Zero female tasters chose Claude. Claude's 29% want-again rate among women vs 80% among men.
Favorites by age group
The 50s–70s age group chose Gemini unanimously — 3 of 3 favorites, 100% want-again rate.
The Miso Measurement Error
What happened
After the tasting was complete, a review of baking notes revealed a measurement error in Gemini's recipe. The recipe calls for 60 grams of white miso — roughly 3½ tablespoons. The actual bake used 1 tablespoon. The winning cookie had approximately a third of its key ingredient.
Rather than leave that result unverified, both versions were rebaked and a blind comparison tasting was run with 11 people.
What the comparison showed
8 of 11 tasters preferred the correct 60g version. The composite score for the correct version was 7.57 versus 6.83 for the one-tablespoon version. The biggest gap was texture — two tasters described the undermeasured version as "grainy," which didn't appear once for the correct version. Smell was identical between both versions.
The win holds, and the data suggests the margin would have been larger with the correct recipe. The exact gap from the original three-way tasting cannot be known without rerunning it — which was not done. This is home science. The error is documented; the verdict stands.
See the Full Tasting
The complete blind tasting, scorecard reveals, the salt paradox breakdown, and the miso disclosure — all on camera.
Watch on YouTube ↗