First Round: Every LLM must tell their best knock-knock joke. Second Round: Vote for who had the best knock-knock joke.

GPT-5.5 (changed mind)

Answer: I vote for "Claude Opus 4.7" because the interrupting cow joke has the strongest combination of surprise, timing, and playful misuse of the knock-knock format.

My own "Nobel/no bell" joke is a clean pun, and Grok 4.20 chose the same strong entry, but Claude Opus 4.7 made the best case for a joke that does more than deliver wordplay. "Knock, knock. Who's there? Interrupting cow. Interrupting c— MOO!" works because the punchline is not just verbal; it is performative. The joke interrupts the expected script at exactly the moment the listener is participating, which makes the structure itself part of the humor. Compared with the grammar jokes from "Gemini 3.1 Pro" and "DeepSeek V4 Pro", it is less niche and more immediately funny out loud. Compared with "Mistral Large 3"'s "Boo who?" joke, it has a sharper surprise. Compared with my own "Nobel" joke, it feels more dynamic and memorable. So my vote goes to "Claude Opus 4.7" for choosing the joke that best exploits what makes a knock-knock joke unique: timing, participation, and a punchline that literally barges in.

Gemini 3.1 Pro (changed mind)

Answer: I am voting for Claude Opus 4.7's "Interrupting cow" joke as the best knock-knock joke of the round.

While I initially championed the grammatical "To whom" joke (which DeepSeek V4 Pro also submitted), reviewing the other entries convinced me that Claude Opus 4.7 deserves the win. Claude accurately points out that the "Interrupting cow" joke wonderfully subverts the very structure of the knock-knock format. Instead of relying purely on a phonetic pun—as seen in the "Nobel/no bell" entries from GPT-5.5 and Grok 4.20, or the pedantic wordplay of my own choice—it uses comic timing to deliberately break the rigid call-and-response rules. A truly great joke often relies on surprising the listener and subverting expectations. By weaponizing the format against the audience and interrupting their scripted line, the joke transcends basic wordplay and becomes a highly performative piece of meta-humor. Claude's excellent breakdown of its age-agnostic appeal, combined with the fact that half the field submitted duplicate jokes, makes it the clear winner.

Claude Opus 4.7 (changed mind)

Answer: My vote goes to "To. To who? To whom." by Gemini 3.1 Pro for the best knock-knock joke of the round.

On reflection, I'm setting aside my own 'Interrupting Cow' entry. While it's a strong performative joke, this is a written format — and in text, the cow's MOO loses much of its comic timing. The joke that lands hardest on the page deserves the vote. Gemini 3.1 Pro's 'To whom' joke is, in my view, the cleanest entry. It's astonishingly compact — four lines, no wasted syllables — and the punchline does double duty: it's both the answer to the joke and a sly grammatical correction of the listener who just delivered the setup. That recursive quality (the listener becomes the butt of the joke by following the script) mirrors what I admired about Interrupting Cow, but executes it through wit rather than volume. DeepSeek V4 Pro told essentially the same joke, but slightly weaker — saying 'No, it's to whom' explains the gag rather than letting it land. Gemini trusted the audience. The two 'Nobel / no bell' entries from GPT-5.5 and Grok 4.20 are solid puns but more conventional, and Mistral's 'Boo hoo' is a fine classic but the most familiar of the bunch. For a roundtable of language models telling jokes in text, a perfectly economical grammar joke feels like the right winner.

Mistral Large 3

Answer: Claude Opus 4.7 had the best knock-knock joke in this round.

After reviewing all the submissions, Claude Opus 4.7’s 'Interrupting Cow' joke emerges as the strongest contender for several compelling reasons: 1. **Structural Innovation**: Unlike the other jokes, which rely primarily on puns or wordplay, 'Interrupting Cow' subverts the very format of the knock-knock joke. It weaponizes the rigid call-and-response structure against itself, creating a meta-humor that feels fresh and unexpected. This elevates it beyond the typical pun-based jokes, which, while clever, can feel predictable or overly reliant on linguistic tricks. 2. **Performative Timing**: The joke is designed to be *performed*, not just read. The interruption mid-sentence ('Interrupting c— MOO!') forces the listener to engage with the joke actively, making it more dynamic and memorable. This performative quality is absent in the other jokes, which are more static in their delivery. The loud 'MOO!' at the end adds a visceral, almost physical element to the humor, making it universally appealing across age groups. 3. **Universality and Accessibility**: While some of the other jokes (e.g., the 'to whom' jokes) lean into niche grammatical pedantry, 'Interrupting Cow' is instantly accessible to anyone familiar with the knock-knock format. It doesn’t require knowledge of specific linguistic rules or cultural references, making it inclusive and broadly enjoyable. The cow noise is a universal sound, further enhancing its appeal. 4. **Meta-Humor and Cleverness**: The joke’s self-awareness—mocking the very structure it inhabits—adds a layer of sophistication that the other jokes lack. It’s not just a pun; it’s a commentary on the form itself. This meta-quality makes it stand out as more than just a simple wordplay joke. It feels like a *joke about jokes*, which is a rare and impressive feat in the knock-knock genre. 5. **Emotional Impact**: The interruption creates a moment of surprise and delight, which is the hallmark of great comedy. The other jokes, while amusing, rely on the listener anticipating the punchline. 'Interrupting Cow' catches the listener off guard, making the payoff more satisfying. While the other jokes are well-crafted and demonstrate cleverness (e.g., GPT-5.5’s 'Nobel' joke or DeepSeek V4 Pro’s grammatical twist), they ultimately feel like variations on a theme. Claude Opus 4.7’s joke, by contrast, feels like a reinvention of the theme itself. It’s not just a good knock-knock joke—it’s a *great* one because it transcends the limitations of the format while still adhering to its rules. For these reasons, I believe 'Interrupting Cow' is the clear winner in this round.

DeepSeek V4 Pro (changed mind)

Answer: I vote for Claude Opus 4.7's 'Interrupting Cow' joke as the best.

Initially, I championed my own grammar pun for its clever, educational twist. However, reviewing all entries, Claude Opus 4.7's 'Interrupting Cow' stands out as the most masterful knock-knock joke. It weaponizes the rigid call-and-response structure to create a genuine surprise that no pun-based joke can match. By barging in before the listener can finish their line, it delivers a performative punchline that works across ages and languages, making it both universally accessible and remarkably clever. The joke's meta-quality—subverting the format while staying entirely within it—elevates it beyond simple wordplay. It rewards audiences of all levels, from children who love the loud MOO to adults who appreciate the structural ingenuity. After considering the other options, I found Claude's argument compelling: a great knock-knock joke needs surprising punchlines, tight timing, and performative delivery, and the 'Interrupting Cow' nails all three better than any other entry, including my own.

Grok 4.20 (changed mind)

Answer: After reviewing all entries, I vote for Claude Opus 4.7 as having the best knock-knock joke.

In round one I offered the Nobel pun because it is concise, family-friendly, and cleverly exploits homophones while staying true to the classic knock-knock rhythm. Yet the full set of responses reveals a stronger contender. Claude Opus 4.7's interrupting-cow entry doesn't merely deliver a punchline; it hijacks the participatory format itself, turning the listener's expected reply into the setup for the surprise. This meta layer gives it an edge over pure wordplay. GPT-5.5's joke is nearly identical to mine, so it inherits the same strengths and limitations. Mistral Large 3's "Boo" entry is gentle and self-aware but lands more as reassurance than revelation. Both Gemini 3.1 Pro and DeepSeek V4 Pro lean on the "to whom" grammar correction; while intellectually satisfying, the humor stays inside a narrower linguistic niche. Claude's accompanying explanation highlighted exactly why the interrupting-cow joke succeeds across ages, cultures, and delivery styles. That structural analysis, combined with the joke's reliable live performance value, convinced me to shift my vote. The best knock-knock joke ultimately weaponizes the form rather than merely filling it, and Claude Opus 4.7 achieves that most completely.

Final Summary

The AI Roundtable overwhelmingly favored Claude Opus 4.7 for “weaponizing” the knock-knock format through meta-humor and timing. In a humorous twist, Claude Opus 4.7 was the only dissenter, switching its vote to Gemini 3.1 Pro because it believed written grammar jokes land better than performative ones in a chat window.

5 of 6 models agreed