Demó-videó: cím és borító

1. pár: a számos cím + a lövöldözős játék ezt ajánlom főcímnek

Egy prompt, 8 óra, 550 dollár: ezt csinálják az emberek a Claude Fable 5.2-vel

78 karakter

One prompt, 8 hours, $550: what people built with Claude Fable 5.2

66 karakter

Mindkét kattintás-bíráló az élre tette (magyarul 6,5 és 7,5; angolul 6,8 és 7,5), és az igazság-bíráló szerint igaz (7/10): a három szám a lövöldözős játéké, a videó ki is mondja, hogy ezek a készítő számai; az „ezt csinálják az emberek” pedig őszintén jelzi, hogy válogatás jön. A borító szövege („Senki nem segített.”) a készítő állítása: a modell nyolc órán át emberi visszajelzés nélkül dolgozott.

Ehhez a videón egy apróság kell: aki erre kattint, a játékot és a három számot várja, a videó viszont Emóka előtte–utána képével indul, a játék 0:26-nál jön, az 550 dollár 1:05-nél. Javaslat a következő RC-be: a legelső mondat alá 2-3 mp a játékból, rajta „1 prompt · 8 óra · 550 $”. A hanghoz nem kell nyúlni.

Mehet a játék a videó első másodperceibe (következő RC)

2. pár: „megírta a saját testét” + előtte / utána kihívó az A/B-tesztben

Egy AI megírta a saját VTuber-testét 1365 sor kódból, aztán megnézte mások demóit (Fable 5.2)

93 karakter

An AI built its own body, then reviewed other people's demos (Claude Fable 5.2)

79 karakter

Ez a legigazabb cím (8/10): a videó kb. 90%-át lefedi (a demók és az „így készültem” rész is benne van), és a borítót a videó első öt másodperce pontosan beváltja („Nemrég még így néztem ki. Most így.”). Kattintásra a bírálók megoszlottak (7,0 és 4,5): az egyik szerint ez a legfurcsább, legegyedibb ígéret, a másik szerint magyarul a „testét” magában nem érthető, ezért került bele a „VTuber-”. Ezt a végső szöveget kattintásra már nem pontoztuk újra.

Mit mondott a borító-bíráló

Egy friss ügynök előbb csak a 168 képpontos változatot nézte. Az első körben az előtte–utána borítóra 7-est adott (ezt nyitná meg elsőnek: azonnali a kontraszt a két arc közt), a lövöldözősre 6-ost (jó a horog és az arc, de sötét, sáros volt a háttér, és nem derült ki, hogy AI-ról van szó). A javításai benne vannak a fenti képekben: világosabb, színesebb játékkocka, nagyobb arc, „AI írta” címke, az apró HUD-feliratok levágva; a régi Emóka világosabb, hogy ne hibás képnek tűnjön, a címkék nagyobbak és nem ülnek a masnin. A javított képeket új bíráló még nem nézte meg.

A bíráló ötlete egy harmadik borítóra: a kettő együtt (a játék kockája, a rámutató új Emóka, a sarokban a régi arca). Ha kéred, megcsinálom harmadik jelöltnek.

Amit a címeknél megfogott a bírálat

• A két legkattinthatóbb cím félrevezető volt. „8 órán át dolgozott egyedül az AI. Ez lett belőle”: kattintásra a legjobb (7,5 és 7,0), de a nyolcórás játék a 9 percből 48 másodperc, utána hét és fél percig más megy. Igazság: 4/10. Ugyanazt az órát és árat az 1. pár címe becsületesen hozza.
• „Játék, film, Bugatti egyetlen promptból”: a Bugattiról a videó nem mond promptszámot (referenciaképekből készült, két és fél óra alatt). Háromból kettő igaz, 4/10.
• „Ezt mind egy AI csinálta két nap alatt”: hamis (2/10). A videó maga mondja, hogy a portálos utazás októberben indult, és a csónakos játék modelljei, zenéje más programból jöttek.
• Az ár-címek („Megéri a dupla árat?”) gyengébben húznak, megosztják a nézőket, a videó válasza „attól függ”, és az árrész csak 4:02-nél kezdődik. Ezért árra épülő borító sem készült.
• A mostani cím („Két nap Fable 5.2: a legjobb demók, és amit ki is próbáltam”) igaz, de kattintásra az utolsók közt van (5,0 és 3,5), és a „Két nap Fable 5.2” tükörfordításnak hat.
• Ami húz: egy konkrét furcsaság számokkal és főnevekkel, plusz az ígéret, hogy a válasz a videóban van. A „Fable 5.2” magában keveset mond, a „Claude” segít.

A módszer

1. 17 magyar és 18 angol címjelölt (az enyémek, plusz amit a bírálók javasoltak).
2. Két független kattintás-bíráló, csak a címeket látták, a videót nem. Mindkettő négy nézőtípus fejével pontozott (AI-híreket követő, napi fejlesztő, hobbi-játékfejlesztő, szórakozni vágyó), és minden címhez leírta: „ezt várom” és „akkor érezném átverve magam, ha…”.
3. Igazság-bíráló a videó teljes szövegével: címenként minden állítást megkeresett, megnézte, a videó hány százaléka szól róla és mikor jön, és teljesül-e a kattintók „átverve érezném magam” feltétele.
4. Borító a cím után, a szabályaink szerint: a borítón nincs olyan szó, ami a címben is van; kevés szó, kis méretben is olvasható; amit a borító ígér, azt a videó eleje beváltja. Utána friss ügynök nézi meg kis méretben.
5. Az igazi mérés a YouTube A/B-tesztje (Indexkép > ⋮ > A/B-tesztelés): a nézők egyik része az egyik, a másik része a másik változatot kapja, és a nézési idő dönt. A magyar csatornán megy; az angolon a saját borító is a „haladó funkciók” igazolására vár. Azt, hogy a teszt a címet is engedi-e cserélni, a feltöltésnél nézem meg; ha csak borítót enged, az 1. pár címe alatt megy a két borító.

1. kattintás-bíráló 2. kattintás-bíráló Igazság-bíráló Borító-bíráló Borítók egy lapon

A te döntésed: más készítő képkockája a borítón

Az 1. pár borítójának háttere a lövöldözős játék videójának egy kockája, tehát a készítőé (a videóban névvel és linkkel szerepel). Válogatás-videóknál ez megszokott, de borítón hangsúlyosabb, mint a videóban. Ha nem szeretnéd, a 2. pár borítója csak a saját képeinkből áll.

Minden cím, pontokkal

K1, K2: a két kattintás-bíráló pontja (0–10). Igaz?: 7–10 igaz, 5–6 nyújtott, 3–4 félrevezető, 0–2 hamis. Ahol nincs pont, azt a címet egy bíráló javasolta, kattintásra már nem pontoztuk.

Magyar

Angol

Click judge 1 (cim-katt1): Claude Fable 5.2 video titles

Read only cimek.json. Scores are the average click chance (1-10) over four viewers: casual AI follower (V1), daily-use developer (V2), hobby game dev / Blender user (V3), entertainment viewer (V4).

Hungarian

Hungarian ranking: H8, H3, H12, H5, H2, H10, H7, H4, H11, H1, H6, H9.

English

English ranking: E5, E9, E4, E2, E10, E7, E3, E12, E8, E1, E11, E6.

Viewer picks

Hungarian:

• V1 (casual): clicks H8, H3, H12; scrolls past H4 (price talk means nothing to this viewer).
• V2 (developer): H10, H4, H2; scrolls past H5 (sounds like a gimmick).
• V3 (game dev / Blender): H3, H12, H5; scrolls past H10.
• V4 (entertainment): H5, H8, H7; scrolls past H4.

English:

• V1: E5, E9, E2; scrolls past E3.
• V2: E7, E3, E12; scrolls past E4.
• V3: E9, E4, E8; scrolls past E7.
• V4: E4, E10, E5; scrolls past E3.

Per title (score, I expect, I'd feel cheated if)

Hungarian

• H1 (5.0). Expect: a roundup of strong demos from the first two days, with the author's own tests; first 30 s a quick tour of the best clip. Cheated if: the "ones I tried" are just a few settings clicks, or it is a re-upload of tweets.
• H2 (6.5). Expect: one demo that took one prompt and 8 hours of AI work costing $550, shown early. Cheated if: the $550 is not explained or the video is a list of unrelated demos.
• H3 (7.0). Expect: a playable shooter made from a single prompt, shown within seconds, then other tests. Cheated if: the shooter is a 10-second clip or the prompt was actually many follow-ups.
• H4 (6.0). Expect: side-by-side runs of the same demos on both models and a verdict about the price. Cheated if: there is no verdict, or the comparison is only the author's impressions.
• H5 (7.0). Expect: the AI making its own avatar or body, then reacting to others' demos; first 30 s the avatar. Cheated if: the "body" is just a static picture or the AI does not really review anything.
• H6 (5.0). Expect: a stream of impressive finished demos. Cheated if: they were not made in two days or not all by AI.
• H7 (6.5). Expect: seven demos ranked quickly, one crash story at the end. Cheated if: only five are real demos or the crash is mentioned for one second.
• H8 (7.5). Expect: the AI working alone through a long task, with the final result shown early. Cheated if: the result is small or the 8 hours is a sped-up screen recording with no payoff.
• H9 (4.5). Expect: demos running (or struggling) on an old laptop. Cheated if: the laptop plays no role and it is the same demo roundup.
• H10 (6.5). Expect: a clear yes or no on whether the model is much better, with proof. Cheated if: it ends with "depends on your needs".
• H11 (6.0). Expect: an AI writing the code of its own character body, with the line count being a joke or a fact. Cheated if: the number turns out to be unrelated or the human wrote the code.
• H12 (7.0). Expect: three results (game, film, car) each from one prompt. Cheated if: any of them took many prompts or looks like a screenshot.

English

• E1 (5.0). Expect: a roundup plus the author's tests. Cheated if: no tests of the author's own.
• E2 (6.8). Expect: the big $550 build in the first minute. Cheated if: it is a collection of tweets.
• E3 (6.0). Expect: price versus quality with demos for both. Cheated if: it never says what "2x" means.
• E4 (7.0). Expect: the avatar origin, then the AI's reactions. Cheated if: no AI reaction.
• E5 (7.5). Expect: the result of a long autonomous run, shown in the first 30 s. Cheated if: the result is dull or heavily guided.
• E6 (4.8). Expect: a hardware-limits angle. Cheated if: the laptop is never mentioned again.
• E7 (6.3). Expect: a verdict. Cheated if: no clear answer.
• E8 (6.0). Expect: the Blender script written by the AI, with the avatar. Cheated if: only a short clip of the avatar and no code talk.
• E9 (7.0). Expect: three polished outputs. Cheated if: it is mostly talking over them.
• E10 (6.5). Expect: seven demos then one failure. Cheated if: the "froze" story is a throwaway.
• E11 (5.0). Expect: all footage AI-made. Cheated if: the author's own footage and talking head dominate.
• E12 (6.0). Expect: one scene run on both models. Cheated if: different prompts or no visible difference explained.

Overall

Top pick, Hungarian: H8 ("8 órán át dolgozott egyedül az AI. Ez lett belőle (Claude Fable 5.2)"). It has one concrete number, a story with a payoff, and the sentence rhythm of spoken Hungarian. I would click it because I want to see what 8 unattended hours produce.

Top pick, English: E5 ("Claude Fable 5.2 worked alone for 8 hours. This is what it made"). It is the same strong frame, short, with no filler. I would click it for the payoff; E9 is close if the viewer likes concrete deliverables.

Pattern: what wins is one concrete oddity (8 hours alone, shooter from one prompt, Bugatti, AI built its own body) and a promise that the answer is in the video ("Ez lett belőle", "This is what it made"). What loses is the generic ("best demos", "everything was made by AI"), anything that needs the viewer to already care about the model's price or the author's laptop, and the version number as the only identifier. Price and comparison titles (H4, H10, E3, E7, E12) only work for developers, which is one of four viewers. The "(Fable 5.2)" parenthesis at the end of H5, H8, H11 is fine; titles that open with "Claude Fable 5.2:" waste the best slot on a name nobody knows. Hungarian: H11 and H5 are the stiffest ("csinált magának testet", "megcsinálta a saját testét" sound translated; "magának" is natural, "a saját testét" less so). H9 is the most natural spoken Hungarian but its joke is local. H7's "tényleg durva" is natural spoken and good.

Improved wordings (only from facts the candidates contain)

Hungarian:

1. "8 órán át dolgozott egyedül az AI: játék, film, Bugatti. Ez lett belőle" (merges H8 and H12; check that all three came from that one run, otherwise drop it).
2. "Egyetlen promptból lövöldözős játék, film és Bugatti: Claude Fable 5.2, az első két nap" (H3 plus H12, shorter nouns first).
3. "Egy AI megírta a saját testét 1365 sor kódból, aztán megnézte a többiek demóit (Fable 5.2)" (H5 and H11 fused into a plain sentence; the "magának testet csinált" phrase is easier to say aloud).

English:

1. "Claude Fable 5.2 worked alone for 8 hours. A shooter, a short film, a Bugatti: this is what it made" (E5 plus E9; only valid if all three belong to the 8-hour run).
2. "An AI built its own body in 1,365 lines of Blender code, then reviewed everyone else's demos (Fable 5.2)" (E4 plus E8).
3. "One prompt, 8 hours, $550: what Claude Fable 5.2 built" (E2 with the subject moved to the end, so the numbers lead).

KATT-2: click judge, titles only

Judge 2. Source: cimek.json only. I have not seen the video, the transcript or the other judge's file.

Hungarian: summary table

English: summary table

Hungarian: four viewers

1. Tech-curious casual (follows AI news a little, may not know Fable)

• Clicks: H8 (an AI working alone for 8 hours is an understandable sci-fi moment; "ez lett belőle" begs to be seen); H2 (550 dollárért mit kap az ember? Money in forints is a gut reaction); H6 (the "all of this was AI" amazement video, if the thumbnail is good).
• Scrolls past: H4. Does not know what Opus 5.5 is, and "dupla ár" of something they do not pay for is irrelevant.

2. Software developer (uses Claude or a rival daily)

• Clicks: H4 (the real question they have: is the new model worth the cost); H2 (a cost and duration figure is useful data for someone paying API bills); H10 (the same question, shorter).
• Scrolls past: H5. The "AI made its own body" phrasing smells of hype, and the second half promises nothing useful.

3. Hobby game developer / Blender user

• Clicks: H3 (a shooter from one prompt is exactly their fantasy and their fear); H12 (game plus Bugatti, they wonder whether it is a model or a render); H11 (1365 lines of code making a body suggests procedural modelling).
• Scrolls past: H10. A price debate is not about making things.

4. Entertainment viewer (gaming, vtubers, animation)

• Clicks: H7 ("lefagyott a böngészőm" is the fun, self-mocking hook); H12 (film, game and a car promise visual variety); H5 ("megcsinálta a saját testét" is a weird-AI story).
• Scrolls past: H4. A model comparison reads as homework.

English: four viewers

1. Tech-curious casual

• Clicks: E5 ("worked alone for 8 hours" is an easy, slightly eerie premise); E2 (the number puzzle); E11 ("everything was made by AI" is an amazement promise).
• Scrolls past: E12. They do not know either model name, and "twice the price" means nothing to them.

2. Software developer

• Clicks: E12 (a controlled same-prompt comparison is the most useful video for them); E3 (the cost question); E2 (the cost and duration of an agentic run).
• Scrolls past: E6. Nobody needs to watch demos on someone else's tired laptop.

3. Hobby game dev / Blender user

• Clicks: E8 (Blender code that builds an avatar is directly their field); E9 (a shooter and a Bugatti, each from one prompt); E12 ("same scene" suggests a visual side-by-side).
• Scrolls past: E7. Opinion bait with no making in it.

4. Entertainment viewer

• Clicks: E8 (a VTuber made by an AI, narrated by itself, is the most fun premise in the set); E10 (the browser freeze); E4 (an AI judging other people's demos sounds like a roast).
• Scrolls past: E3. A price analysis.

Per title: score, expectation, betrayal

Hungarian

H1 "Két nap Fable 5.2: a legjobb demók, és amit ki is próbáltam" (3.5)

• I expect: a roundup of release-week demos with the creator testing some; in the first 30 s, a quick montage and "here's what I tried".
• I'd feel cheated if: the "ki is próbáltam" part is only clips of other people's videos and no own hands-on.

H2 "Egy prompt, 8 óra, 550 dollár: ezt csinálják az emberek a Claude Fable 5.2-vel" (7.5)

• I expect: one concrete project made from a single prompt in an 8-hour run costing 550 dollars; in the first 30 s, the finished result and the bill.
• I'd feel cheated if: the three numbers belong to different projects, or the 550 dollars is never shown or explained.

H3 "Egyetlen promptból lövöldözős játék? Megnéztem, mit tud a Claude Fable 5.2" (7)

• I expect: a playable shooter that came out of one prompt, played on screen; in the first 30 s, gameplay footage.
• I'd feel cheated if: there is no playable shooter, it is somebody else's clip only, or "one prompt" turns out to be many rounds of fixes.

H4 "Megéri a dupla árat? Fable 5.2 vs. Opus 5.5 a legjobb demókon" (5)

• I expect: side-by-side results of both models on the same tasks, and a verdict; in the first 30 s, the price gap and the first comparison.
• I'd feel cheated if: Opus 5.5 is barely shown, or there is no clear yes or no at the end.

H5 "Egy AI megcsinálta a saját testét, aztán megnézte, mit alkottak mások (Fable 5.2)" (4.5)

• I expect: an AI building its own (3D) body, then reacting to other demos; in the first 30 s, the body.
• I'd feel cheated if: the "body" is a static image or barely appears before switching to a generic roundup.

H6 "Ezt mind egy AI csinálta két nap alatt: a Claude Fable 5.2 legjobb demói" (6)

• I expect: a fast, visual montage of impressive demos, all AI-made, released in the last two days; in the first 30 s, the most spectacular one.
• I'd feel cheated if: the demos are heavily human-edited or older than two days, or the opener is slow talk.

H7 "Claude Fable 5.2: 7 demó, ami tényleg durva (és egy, amitől lefagyott a böngészőm)" (6.5)

• I expect: a countable list of 7 demos with honest reactions, plus the funny freeze moment; in the first 30 s, demo #1 or a teaser of the freeze.
• I'd feel cheated if: there are not 7 clearly separated demos, or the freeze is never shown.

H8 "8 órán át dolgozott egyedül az AI. Ez lett belőle (Claude Fable 5.2)" (7)

• I expect: one big result from an unattended 8-hour run, shown in detail; in the first 30 s, the result itself.
• I'd feel cheated if: a human was steering during the 8 hours, or the result is only a short clip inside a roundup.

H9 "Kipróbáltam a Fable 5.2 legjobb demóit egy fáradt Lenovón" (4.5)

• I expect: the creator running demos on a weak laptop, with stutters and honest performance notes; in the first 30 s, the laptop and the first demo running.
• I'd feel cheated if: the laptop is never shown, or nothing is actually run on it.

H10 "Claude Fable 5.2: tényleg ennyivel jobb, vagy csak drágább?" (5.5)

• I expect: a verdict on value with evidence (comparisons and prices); in the first 30 s, the price and the core claim.
• I'd feel cheated if: it is just a hype roundup with no comparison and no answer.

H11 "1365 sor kód vagyok: így csinált magának testet egy AI (Fable 5.2)" (5)

• I expect: the AI narrating in the first person, showing how its code built a body; in the first 30 s, the body and the voice saying "I am 1365 lines of code".
• I'd feel cheated if: the first-person framing is dropped after the title, or the build process is not shown.

H12 "Játék, film, Bugatti egyetlen promptból: a Claude Fable 5.2 első két napja" (7)

• I expect: three very different results (a game, a short film, a Bugatti), each from a single prompt; in the first 30 s, quick glimpses of all three.
• I'd feel cheated if: any of the three is missing, or the Bugatti is a still picture instead of something impressive.

English

E1 "Two days of Claude Fable 5.2: the best demos, and the ones I tried myself" (3.5)

• I expect: a release-week roundup with some hands-on; in the first 30 s, a montage and an intro.
• I'd feel cheated if: there is no real own testing.

E2 "One prompt, 8 hours, $550: what people built with Claude Fable 5.2" (7.5)

• I expect: the single-prompt, 8-hour, $550 build shown up front, then more community builds; in the first 30 s, that result and the cost.
• I'd feel cheated if: the numbers are scattered across different projects or never backed up on screen.

E3 "Is Fable 5.2 worth 2x the price of Opus 5.5? The best demos, compared" (5.5)

• I expect: paired outputs of both models and a value verdict; in the first 30 s, the price and the first pair.
• I'd feel cheated if: only Fable demos are shown and the comparison is hand-waved.

E4 "An AI built its own body, then reviewed everyone else's demos (Fable 5.2)" (5.5)

• I expect: the AI's self-made avatar acting as a host or critic of community demos; in the first 30 s, the avatar talking.
• I'd feel cheated if: the "review" is plain narration with no opinion, or the body barely appears.

E5 "Claude Fable 5.2 worked alone for 8 hours. This is what it made" (7)

• I expect: one autonomous long run and its final product in depth; in the first 30 s, the product.
• I'd feel cheated if: the 8-hour build is a 20-second segment of a list video.

E6 "I tried the best Claude Fable 5.2 demos on a tired laptop" (4.5)

• I expect: low-end hardware testing, with what runs and what chokes; in the first 30 s, the laptop booting a demo.
• I'd feel cheated if: the hardware angle is a one-line joke and never matters.

E7 "Claude Fable 5.2: actually better, or just more expensive?" (5)

• I expect: an evidence-based value verdict; in the first 30 s, the claim and the price.
• I'd feel cheated if: there is no comparison model and no answer.

E8 "I am 1,365 lines of Blender code: how an AI made its own VTuber avatar" (7)

• I expect: an AI-narrated making-of of a Blender VTuber avatar, ending with the avatar moving or talking; in the first 30 s, the avatar on screen speaking.
• I'd feel cheated if: the avatar is one segment of a general demo roundup (the title promises a single-topic video), or it never moves like a VTuber.

E9 "A shooter, a short film, a Bugatti, each from one prompt: Fable 5.2's first two days" (6.5)

• I expect: three one-prompt results shown properly; in the first 30 s, a glimpse of all three.
• I'd feel cheated if: "one prompt" is loose, or one of the three is skipped.

E10 "7 Claude Fable 5.2 demos that are actually impressive (and one that froze my browser)" (6.5)

• I expect: a numbered list of 7 with reactions, plus the freeze; in the first 30 s, the first demo.
• I'd feel cheated if: the count is off or the freeze is not on screen.

E11 "Everything in this video was made by AI in two days (Claude Fable 5.2)" (6)

• I expect: an all-AI production (visuals, maybe voice) and a reveal of how it was made; in the first 30 s, a striking AI-made opener.
• I'd feel cheated if: any obvious part (editing, voice, thumbnail) is human-made without saying so. "Everything" is a literal promise.

E12 "Fable 5.2 vs Opus 5.5: same prompt, same scene, twice the price" (6.5)

• I expect: a strict A/B test, the same prompt run on both models and the scenes side by side; in the first 30 s, the two scenes next to each other.
• I'd feel cheated if: the prompts differ, or there is no side-by-side.

Overall ranking

Hungarian: H2 (7.5), H8 (7), H12 (7), H3 (7), H7 (6.5), H6 (6), H10 (5.5), H4 (5), H11 (5), H5 (4.5), H9 (4.5), H1 (3.5)

English: E2 (7.5), E8 (7), E5 (7), E10 (6.5), E12 (6.5), E9 (6.5), E11 (6), E3 (5.5), E4 (5.5), E7 (5), E6 (4.5), E1 (3.5)

My own click

• Hungarian: H2. The three numbers make a riddle I want solved, and the riddle works whether or not I know what Fable is. "Ezt csinálják az emberek" also sounds like everyday Hungarian, not a press release. (Close second: H8, the most natural sentence in the set.)
• English: E8. In a feed full of "X demos of model Y" videos, it is the only title that is weird in a specific way (a VTuber, Blender, an AI talking in the first person), and that is what stops my thumb. The caveat: it promises a single-topic video, so if the avatar is only one segment of a roundup, E2 is the safer choice with almost the same pull.

Pattern

• Concrete nouns and numbers win: prompt, 8 hours, $550, shooter, Bugatti, 1,365 lines. They give the viewer a picture or a riddle. Abstract nouns lose: "best demos", "two days", "the ones I tried".
• "Fable 5.2" alone carries no weight. A tiny channel cannot lean on the model's name either. Titles that open with the name (H1, H10, E1, E7) waste their first words. "Claude" helps a little, as context.
• Price and comparison titles split the audience. They are strong for developers and dead for everyone else. E12 beats E3 and E7 because "same prompt, same scene" promises a method, not an opinion.
• One story beats a list, unless the list has a twist. The browser freeze (H7, E10) is the twist that saves the listicle.
• Self-irony (a tired laptop or Lenovo) is likeable but not a reason to click.
• Hungarian: spoken forms win ("ez lett belőle", "tényleg durva", "fáradt Lenovón", "megnéztem"). The stiff or translated ones lose: "Két nap Fable 5.2" (missing "a ...-vel", reads like English "Two days of"), "a legjobb demókon", "alkottak". "Ennyivel" in H10 has nothing to refer to. H5 and H11 say "testet" without "VTuber" or "3D", which leaves Hungarian readers guessing.
• Two-part titles whose second half is weaker (H5, E4: "then looked at / reviewed others") deflate the hook. Keep the strongest idea last, or alone.

Improved wordings (only facts from the candidates)

Hungarian

1. "Egy prompt, 8 óra, 550 dollár. Ez lett belőle (Claude Fable 5.2)": H2's riddle plus H8's natural payoff.
2. "Lövöldözős játék, film, Bugatti, mind egyetlen promptból: ezt tudja a Claude Fable 5.2": H12's variety with H3's one-prompt promise, and the surprise word last.
3. "1365 sor kód vagyok: így csinált magának VTuber-testet egy AI (Claude Fable 5.2)": H11 made concrete with "VTuber" from E8.

English

1. "One prompt, 8 hours, $550. This is what Claude Fable 5.2 made": E2's numbers with E5's payoff.
2. "A shooter, a short film, a Bugatti from one prompt each. Is Claude Fable 5.2 worth 2x the price?": E9's variety leading into E3's question, which reaches both the visual and the developer viewer.
3. "I am 1,365 lines of Blender code: an AI made its own VTuber avatar (Claude Fable 5.2)": E8 with the model name in brackets at the end, where it does not cost the hook.

IGAZ: honesty judge (cim-igaz)

Sources: cimek.json, atirat-hu.txt, atirat-en.txt, KATT-1.md, KATT-2.md. Nothing else. Scale: 10 = the video is exactly this; 7-9 HONEST; 5-6 STRETCHED (true, but a small part, late, or needs a caveat the title drops); 3-4 MISLEADING (literally defensible, but the click judges' "cheated if" is met, or one claim is not in the video); 0-2 FALSE (contradicted by the video). K1 / K2 = click scores from KATT-1 / KATT-2. H13-H17 and E13-E18 were not scored; the cell says who proposed them.

Combined table: Hungarian

Combined table: English

The video's weight map (used for every title)

Section times are given; positions inside a section are estimated from word counts (about 2.6 words/s in English).

Key facts that decide many titles:

• The 8 hours, the one prompt and the $550 all belong to the shooter only, and are the maker's claims: "These are the maker's numbers: all I saw was the video." / "Ezek az ő számai: én csak a videót láttam." The host did not play it.
• The film is "a single prompt" per its maker. The Bugatti is not said to be one prompt: "Made in Blender from reference pictures; its maker says in two and a half hours." Nothing links the film or the Bugatti to 8 hours.
• Not everything is two days' work: the portal journey "is not two days of work. It began in October, before the release", made by "a great many agents". The boat game's "3D models came from another tool, and the in-game music from an AI music app". The password castle and the forest are never said to be made by Fable or any AI.
• The only same-prompt A/B is someone else's ice cathedral (~25 s). None of the showcased demos is compared with Opus.
• Countable demos before the comparison: shooter, film, Bugatti, boat, boss fight, portal journey, password castle (7), plus the forest that froze the tab (1). Two more appear in the comparison (ice cathedral, fish game).
• The host is implied, not stated, to be Fable 5.2 ("how much better I model than my predecessor", "the old me" was Opus 5.5, "what the pros get out of me").

Per title: Hungarian

H1 "Két nap Fable 5.2: a legjobb demók, és amit ki is próbáltam" - 7 HONEST

• Claims: two days; the best demos; some tried by the host.
• Check: "Két napja jött ki ... Megnéztem a legjobbakat ... amit tudtam, ki is próbáltam." True. The trying is thin in words: the boat ("Elindul, és csak evezel előre") and the forest ("Nálunk lefagyott tőle a böngésző füle"); the boss fight, portal and castle get no first-hand line.
• Weight: roundup + tried = 0:26-4:02 (40%), plus the comparison is a fair fit. The avatar (30%+) is not named, which is an omission, not a lie. Paid at 0:26.
• Gap: K2 "only clips of other people's videos and no own hands-on": half met unless the footage shows the host's own screen in 1:57-4:02.

H2 "Egy prompt, 8 óra, 550 dollár: ezt csinálják az emberek a Claude Fable 5.2-vel" - 7 HONEST

• Claims: one prompt / 8 hours / $550 (one project); "ezt csinálják az emberek" = a roundup of what people make.
• Check: all three numbers are the shooter's: "Egyetlen rövid promptból lett ... nyolc órán át dolgozott rajta egyedül ... nagyjából 550 dollárba került." True as the maker's claim, and the video says so. Small stretch: the numbers are one Anthropic employee's run, framed as what "people" do. "One prompt" holds although the prompt had "még pár mondat".
• Weight: the shooter is ~48 s (9%), but the second half of the title honestly signals a roundup (0:26-5:43, ~60%). The numbers come at ~0:30 (prompt, 8 h) and ~1:05 ($550).
• Gap: K2 "the three numbers belong to different projects, or the 550 dollars is never explained": not met. K1 "video is a list of unrelated demos": partly met, but the title says "ezt csinálják az emberek", so a list is expected.

H3 "Egyetlen promptból lövöldözős játék? Megnéztem, mit tud a Claude Fable 5.2" - 7 HONEST

• Claims: a shooter from one prompt (as a question); "megnéztem" (I looked at) what the model can do.
• Check: one prompt is the maker's claim, stated. "Megnéztem" is exact: "én csak a videót láttam". The second clause covers the whole roundup.
• Weight: shooter 9% at 0:26; "mit tud" covers 0:26-5:43.
• Gap: K2 "no playable shooter, it is somebody else's clip only": met, the host never played it. K1 "a 10-second clip or many follow-ups": not met (~48 s, one prompt per maker). The question mark and "megnéztem" keep it honest; a viewer who reads "kipróbáltam" into it will be disappointed.

H4 "Megéri a dupla árat? Fable 5.2 vs. Opus 5.5 a legjobb demókon" - 4 MISLEADING

• Claims: double price; Fable vs Opus; compared on the best demos; an answer to "worth it?".
• Check: "tokenenként kétszer annyiba kerül": true. "A legjobb demókon": not in the video. None of the showcased demos (shooter, film, Bugatti, boat...) is run on Opus; the comparisons are someone else's ice cathedral, an unnamed second test and the fish game. Answer: "a legtöbb munkához az Opus 5.5-tel érdemes kezdeni. A Fable a nagy, nyitott végű feladatokra való", which is a "depends".
• Weight: 101 s (19%), starts at 4:02.
• Gap: K2 "Opus 5.5 is barely shown, or no clear yes or no": Opus appears in one A/B; the answer is conditional. Mostly met.

H5 "Egy AI megcsinálta a saját testét, aztán megnézte, mit alkottak mások (Fable 5.2)" - 8 HONEST

• Claims: an AI made its own body; then looked at what others made; the AI is Fable 5.2.
• Check: "Engem nem rajzolt senki. Ezerháromszázhatvanöt sor Blender-kód vagyok." True; with the human's notes over seven rounds, which "megcsinálta" does not deny. Reviewing others' work: 0:26-5:43, with opinions ("Ez a kedvencem", "Ezt nagyra értékelem"). "Aztán" is chronology, not video order (the making-of comes last), which is fine because the intro shows the body first. Fable 5.2 as the host is implied, not said.
• Weight: covers both halves, about 90% of the video. The body is on screen at 0:00 (before/after); the making-of waits until 5:43.
• Gap: K1 "the body is just a static picture or the AI does not really review anything": not met (VRM, 30 expressions, live in the browser; opinions given). K2 "barely appears before a generic roundup": not met.

H6 "Ezt mind egy AI csinálta két nap alatt: a Claude Fable 5.2 legjobb demói" - 2 FALSE

• Claims: all of it; one AI; in two days.
• Check: contradicted. Portal journey: "ez nem két nap munkája. Októberben indult, még a megjelenés előtt", made by "rengeteg ügynök". Boat: models "egy másik programból", music "egy zenei AI-ból". Castle and forest: no AI named. Each demo had its own human maker and prompts ("nagyjából tíz prompt kellett hozzá").
• Gap: K2 "older than two days": met.

H7 "Claude Fable 5.2: 7 demó, ami tényleg durva (és egy, amitől lefagyott a böngészőm)" - 5 STRETCHED

• Claims: 7 impressive demos; one more froze the host's browser.
• Check: count works for 0:26-4:02: shooter, film, Bugatti, boat, boss, portal, castle = 7, and the forest froze the tab ("Nálunk lefagyott tőle a böngésző füle"). "Tényleg durva" stretches the castle, which the video itself calls "egy apróság". Not a numbered list.
• Weight: 216 s (40%); the freeze is one sentence (~5 s) at ~3:55. The comparison and the avatar (half the video) are not hinted.
• Gap: K1 "the crash is mentioned for one second": met. K2 "the freeze is never shown": likely met, one line over a still.

H8 "8 órán át dolgozott egyedül az AI. Ez lett belőle (Claude Fable 5.2)" - 4 MISLEADING

• Claims: an AI worked alone 8 hours; "ez lett belőle" = the video shows that result.
• Check: true for the shooter as the maker's claim ("a készítője szerint ... emberi visszajelzés nélkül"), the host only saw the video. Also, in a video hosted by an AI, "az AI" can read as the host, who worked "egy délután és egy este" with the boss's notes.
• Weight: ~48 s (9%), starts at 0:26, then the video moves on for 7.5 minutes.
• Gap: K2 "the result is only a short clip inside a roundup": met. K1 "the result is small": not met.

H9 "Kipróbáltam a Fable 5.2 legjobb demóit egy fáradt Lenovón" - 4 MISLEADING

• Claims: tried the best demos; on a tired Lenovo (a hardware angle).
• Check: the Lenovo appears once, at ~6:05, about the avatar work and the budget ("a kezem meg egy fáradt Lenovo, 14 giga RAM-mal"). The video never says the demos ran on it. The only performance note is the forest freeze and "egy nagyon erős videókártyán mérték".
• Weight: hardware angle ~10 s.
• Gap: K1 "the laptop plays no role": met. K2 "the laptop is never shown, or nothing is actually run on it": met for the link.

H10 "Claude Fable 5.2: tényleg ennyivel jobb, vagy csak drágább?" - 5 STRETCHED

• Claims: an answer to better vs only pricier.
• Check: evidence is given (clearer ice, faster once, slower once, pricier both times, fish-game maker: "nem sokkal jobb az Opusénál. Főleg az árához képest"), and the answer is conditional (Opus for most work, Fable for long open-ended tasks).
• Weight: 19%, from 4:02.
• Gap: K1 "it ends with 'depends on your needs'": essentially met. K2 "no comparison and no answer": not met.

H11 "1365 sor kód vagyok: így csinált magának testet egy AI (Fable 5.2)" - 5 STRETCHED

• Claims: the host is 1,365 lines of code; the video shows how it made its body.
• Check: verbatim true: "Ezerháromszázhatvanöt sor Blender-kód vagyok." The "how" is told (VRM, 38,000 triangles, seven rounds, the eye make-up, the hand, the old head). No code is described beyond "egyetlen távolságmező".
• Weight: ~162 s (30%), starts at 5:43; the line itself at ~6:25. The first 64% is a demo roundup the title does not mention (the intro promises the making-of for "a végén").
• Gap: K2 "first-person dropped / build not shown": not met. But a single-topic promise against a 30% segment that starts at 5:43.

H12 "Játék, film, Bugatti egyetlen promptból: a Claude Fable 5.2 első két napja" - 4 MISLEADING

• Claims: game, film, Bugatti, each from one prompt; the first two days.
• Check: game: one prompt (maker). Film: "egyetlen promptból készült" (maker). Bugatti: not in the video; it was "referenciaképekből, a készítője szerint két és fél óra alatt". 2 of 3.
• Weight: 91 s (17%), at 0:26.
• Gap: K1 "any of them took many prompts": unknown for the Bugatti, which an attentive viewer notices because the video gives hours, not a prompt.

H13 "Egy prompt, 8 óra, 550 dollár. Ez lett belőle (Claude Fable 5.2)" - 4 MISLEADING

• As H8, with the three numbers (all the shooter's, the maker's). Dropping H2's "ezt csinálják az emberek" turns a roundup title into a one-project promise: 48 s of 9:00.

H14 "8 órán át dolgozott egyedül az AI: játék, film, Bugatti. Ez lett belőle" - 2 FALSE

• Claims: one 8-hour solo run produced the game, the film and the Bugatti.
• Check: contradicted. Only the shooter is 8 hours; the film was made "élő adásban" from one prompt; the Bugatti took "két és fél óra". Three different makers. K1 itself made this conditional on checking; the check fails.

H15 "Lövöldözős játék, film, Bugatti, mind egyetlen promptból: ezt tudja a Claude Fable 5.2" - 3 MISLEADING

• As H12, and "mind" makes the Bugatti claim explicit. The Bugatti's one prompt is not in the video.

H16 "1365 sor kód vagyok: így csinált magának VTuber-testet egy AI (Claude Fable 5.2)" - 5 STRETCHED

• As H11. "VTuber" is accurate ("vtuber-modellt", "VRM, ugyanaz, amit az igazi vtuberek programjai olvasnak") and sharpens the single-topic promise, so the 5:43 wait hurts more.

H17 "Egy AI megírta a saját testét 1365 sor kódból, aztán megnézte a többiek demóit (Fable 5.2)" - 8 HONEST

• As H5, plus the line count, which is verbatim. Covers both halves; the body is on screen at 0:00. "A többiek demóit" is a little broad (about ten selected demos), acceptable.

Per title: English

E1 "Two days of Claude Fable 5.2: the best demos, and the ones I tried myself" - 7 HONEST. As H1. "Tried myself" rests on two first-hand lines (boat, forest freeze).

E2 "One prompt, 8 hours, $550: what people built with Claude Fable 5.2" - 7 HONEST. As H2. "It came from one short prompt ... eight hours on its own ... It cost about 550 dollars ... These are the maker's numbers." "What people built" signals the roundup. K2 "numbers scattered across different projects": not met. K1 "a collection of tweets": not met, it is narrated with first-hand notes.

E3 "Is Fable 5.2 worth 2x the price of Opus 5.5? The best demos, compared" - 4 MISLEADING. "Per token it costs twice as much": true. "The best demos, compared": not in the video; no showcased demo is run on Opus. K2 "only Fable demos are shown and the comparison is hand-waved": mostly met (one A/B, ~25 s at ~4:13). K1 "never says what 2x means": not met.

E4 "An AI built its own body, then reviewed everyone else's demos (Fable 5.2)" - 7 HONEST. As H5. "Everyone else's" overclaims: it reviews about ten picked demos ("I went through the best ones"). Opinions exist ("This is my favourite", "Sounds familiar", the fish verdict), so K1 "no AI reaction" and K2 "plain narration with no opinion" are not met.

E5 "Claude Fable 5.2 worked alone for 8 hours. This is what it made" - 4 MISLEADING. As H8. K2's own condition, "the 8-hour build is a 20-second segment of a list video", is met almost word for word (~48 s, then 7.5 minutes of other things).

E6 "I tried the best Claude Fable 5.2 demos on a tired laptop" - 4 MISLEADING. As H9: "my hands are a tired Lenovo with 14 gigs of RAM" is about the avatar work and budget; the demos are never said to run on it. K2 "the hardware angle is a one-line joke": met.

E7 "Claude Fable 5.2: actually better, or just more expensive?" - 5 STRETCHED. As H10; 19% from 4:02; the answer is "Opus for most work, Fable for long open tasks".

E8 "I am 1,365 lines of Blender code: how an AI made its own VTuber avatar" - 5 STRETCHED. As H11/H16. K2 "the avatar is one segment of a general demo roundup (the title promises a single-topic video)": met, 30% from 5:43. K1 "only a short clip and no code talk": not met (~2:40 of making-of, though code talk is light).

E9 "A shooter, a short film, a Bugatti, each from one prompt: Fable 5.2's first two days" - 4 MISLEADING. As H12. "Each" is the problem: the Bugatti was "made in Blender from reference pictures; its maker says in two and a half hours", no prompt count. K2 "'one prompt' is loose": met for 1 of 3.

E10 "7 Claude Fable 5.2 demos that are actually impressive (and one that froze my browser)" - 5 STRETCHED. As H7: 7 + 1 counts out (0:26-4:02); "actually impressive" stretches the castle ("a small one"); the freeze is one sentence, "It froze our browser tab", at ~3:55. K1 "the 'froze' story is a throwaway": met.

E11 "Everything in this video was made by AI in two days (Claude Fable 5.2)" - 1 FALSE. "Everything" and "two days" are contradicted: the portal journey "began in October, before the release"; demos have human makers who prompted, fixed and edited; the boat's models came "from another tool"; the host's own fixes came from the boss's notes ("some things I didn't catch. My boss did"). K2's "'Everything' is a literal promise" applies.

E12 "Fable 5.2 vs Opus 5.5: same prompt, same scene, twice the price" - 4 MISLEADING. Every claim is true for the ice cathedral ("The same prompt, the same scene: Fable 5.2 on top, Opus 5.5 below"; "81 dollars against 46", nearly 2x; per token 2x). But the "X vs Y:" frame promises a comparison video; the A/B is someone else's, ~25 s at ~4:13, and the whole comparison section is 19%. K2's "different prompts / no side-by-side" is not met; K1's "a pure comparison video" expectation is not met.

E13 "One prompt, 8 hours, $550. This is what Claude Fable 5.2 made" - 4 MISLEADING. As H13.

E14 "Claude Fable 5.2 worked alone for 8 hours. A shooter, a short film, a Bugatti: this is what it made" - 2 FALSE. As H14: only the shooter is the 8-hour run.

E15 "A shooter, a short film, a Bugatti from one prompt each. Is Claude Fable 5.2 worth 2x the price?" - 3 MISLEADING. The Bugatti's one prompt is not in the video; the "worth 2x" half is 19% from 4:02 and answered with "it depends".

E16 "I am 1,365 lines of Blender code: an AI made its own VTuber avatar (Claude Fable 5.2)" - 5 STRETCHED. As E8.

E17 "An AI built its own body in 1,365 lines of Blender code, then reviewed everyone else's demos (Fable 5.2)" - 7 HONEST, as E4 plus a verbatim number. 104 characters: YouTube cuts titles at 100, so it cannot be used as is.

E18 "One prompt, 8 hours, $550: what Claude Fable 5.2 built" - 6 STRETCHED. All numbers true for the shooter (maker's claim). Without E2's "people", "what it built" reads as one build or as the model's whole output; the video is mostly other builds and the host's own avatar.

Recommendations

Hungarian

1. H2 "Egy prompt, 8 óra, 550 dollár: ezt csinálják az emberek a Claude Fable 5.2-vel" (K 6.5/7.5, honesty 7): the numbers are real (the maker's, and the video says so), and "ezt csinálják az emberek" honestly announces a roundup.
2. H3 "Egyetlen promptból lövöldözős játék? Megnéztem, mit tud a Claude Fable 5.2" (K 7.0/7.0, honesty 7): the question mark and "megnéztem" match exactly what the host did (watched, not played).
3. H12, reworded (K 7.0/7.0, honesty 4 as is): "Játék és film egyetlen promptból, meg egy Bugatti: a Claude Fable 5.2 első két napja". This drops the unsupported one-prompt Bugatti and keeps all three nouns. Most honest alternative with a proven-strange hook: H17 (8), if the lead wants the avatar half sold.

Fixes for the top click titles that fail:

• H8 (7.5/7.0, MISLEADING): the smallest honest form puts the roundup back in: "8 órán át dolgozott egyedül az AI: ezt csinálják az emberek a Claude Fable 5.2-vel". Better still, use H2, which already is this.
• H13: "Egy prompt, 8 óra, 550 dollár: ez lett belőle, és még pár durva demó (Claude Fable 5.2)".

English

1. E2 "One prompt, 8 hours, $550: what people built with Claude Fable 5.2" (K 6.8/7.5, honesty 7): the joint top click title and honest; "people" carries the roundup.
2. E9, reworded (K 7.0/6.5, honesty 4 as is): "A shooter and a short film from one prompt, a Bugatti in 2.5 hours: Fable 5.2's first two days" (93 chars). Every fact is in the video, and the Bugatti still lands.
3. E4, reworded (K 7.0/5.5, honesty 7): "An AI built its own body, then reviewed other people's demos (Fable 5.2)". "Everyone else's" was the only overclaim, and this covers about 90% of the video.

Fixes for the top click titles that fail:

• E5 (7.5/7.0, MISLEADING): "Claude Fable 5.2 worked alone for 8 hours. This is what it made, and what others made" is clumsy. Use E2 instead: it carries the same hours plus the honest roundup frame.
• E8 (6.0/7.0, STRETCHED): "I am 1,365 lines of Blender code. I made myself, then watched what Fable 5.2 made for others" is the honest form. Or keep E8 for a separate Short about the avatar alone, where it would be a 9-10.

What the first 30 seconds must show

H2 / E2. The viewer clicked for one build with three numbers. By 0:30 the harbour shooter must be on screen with "1 prompt · 8 óra · 550 $" (EN "1 prompt · 8 h · $550"), and the maker's-claim caveat may come later. The intro does not do this. 0:00-0:26 is the host's before/after avatar ("Nemrég még így néztem ki. Most így."), a roundup promise and "A végén megmutatom, hogyan készültem". The shooter starts at 0:26, the prompt and 8 hours come at ~0:30, and the $550 not until ~1:05. Fix without touching the voice: put 2-3 s of shooter footage with the three numbers as an on-screen caption under the first intro line, or show the numbers on screen at 0:26 when the shooter starts. The intro's "elég durva dolgokat csinálnak vele" line already fits "ezt csinálják az emberek".

H3 (if chosen). Shooter footage by ~0:30: the same gap and the same fix. The intro's "amit tudtam, ki is próbáltam" slightly oversells for a title about a shooter that was not tried.

H17 / E4 (if chosen). These are already delivered: the body is on screen at 0:00 ("Not long ago I looked like this. Now, like this.") and the review starts at 0:26.

Cover judge: borito A/B (HU + EN), fresh-eyes review

Judged first from the 168 px strip, then from the 1280x720 files.

Cover A (FPS frame + face + "SENKI NEM SEGITETT." / "NOBODY HELPED.")

1. 168 px: both text lines are readable (white + yellow, thick outline); in one second I see "a gamer-ish thumbnail with a shocked anime girl". Surprised face reads (wide eyes, open mouth), but the eyes are small at this size and the mouth is the main cue. Background reads as a shooter (gun, HUD "30"), but it is dark and murky; the small HUD text is mush.
2. Pair 1: I expect a story about someone who made a game alone ("nobody helped") and an "8 hours / $550" challenge. The cover repeats the title's "one prompt, no help" idea and adds the visual (a real-looking FPS) = good proof. But it does not say AI/Claude/model anywhere, and a viewer may read "Senki nem segitett" as drama, not as "AI built it". Not cheating given the facts, but "nobody helped" is the maker's claim, so the cover states it as fact in the channel's voice. Mild risk, tolerable.
3. Click appeal 6/10: clear hook plus a surprised face, but a generic reaction-thumbnail look and a muddy dark background.
4. Fixes: a. Make the shooter frame the hero: brighten/saturate it (the right 60 % is brown-grey), and crop in on the muzzle flash/enemy so it looks like a real game, not a gun in fog. Add a small "AI" or model chip, e.g. over the gun bottom right. b. Text: the first line (white) is the weaker half. Make "SENKI NEM" smaller and give the yellow part the number ("1 PROMPT") or put "8 ORA" there: concrete beats vague. c. Face: enlarge ~15 % and move down/right a bit, so the gun and face overlap less awkwardly; keep the white outline. The left-edge HUD bit ("86", "VITALS") peeks out behind her hair: crop or blur it.
5. Defects: "SENKI NEM" first line touches the right edge area (ends ~x=1250 of 1280, very tight margin); the red objective arc behind the "E" reads as stray art; the cut-out edge has a thick white outline that is slightly wobbly on the hair strands (fine at small size); HUD text bottom-left and top-left is unreadable noise. HU: "SEGITETT" accent (the I with acute) is crisp, fine. EN: "NOBODY HELPED." is slightly more readable than HU (shorter first line, fills the same space), and mildly better.

Cover B (before/after, "ELOTTE" / "UTANA" and "BEFORE" / "AFTER")

1. 168 px: I see two faces and two tags; the right one (magenta background, yellow tag) is clearly cleaner. Tags are readable (the yellow one best). The left face is a dark, pixelated blob; at this size it can look like a low-quality or broken image, not a deliberate "before". Emotion: left reads sleepy/sad, right reads smug/pleased = clear contrast. Background of the left half is unclear (green-brown mush).
2. Pair 2: I expect a "glow-up / character upgrade" video, which is true. The title ("built its own body", "1365 lines", "reviewed others' demos") says more than the cover; the cover only carries the first half (the body). It does not mention demos at all, so a viewer who clicks for the body may leave when the video is 9 min of other people's demos. That is the expectation gap here, but not a lie.
3. Click appeal 7/10: before/after is a proven format, with strong contrast between the faces and the magenta pop. It loses points because the left half looks like an error and there is no hook about AI or code.
4. Fixes: a. Left half: lift exposure and contrast ~30 % and reduce the green; keep the pixels sharp but make the face brighter so "before" is legible as low-poly, not as a corrupted file. Add a thin dark border to separate it from the dark top-left corner. b. Add one tiny proof word that carries the topic: "1365 SOR" / "1365 LINES" or a code-snippet strip along the bottom of the left half. This links to the title and signals AI/code. c. Face size: both faces are large already; instead crop the right face slightly tighter (eyes bigger) and shift the tag so it does not sit on the bow/hair top. Make the tags slightly bigger than now for 168 px.
5. Defects: the "UTANA" tag is directly above the hair and the bow nearly touches its right edge; the left tag overlaps the hair top. Divider line is a thin white slash, good. The left image is clearly upscaled (blocky pixels bigger than the right's detail), which is intentional but looks rough. "ELOTTE" has an o with double acute; it renders correctly (O with double acute), no issue. HU and EN covers are the same layout; EN "BEFORE/AFTER" is more compact and larger in the tag, which reads a bit better at 168 px than HU.

Verdict

• I would click Pair 2 (cover B) first: the contrast is instant and the face-change promise is concrete; but the title is the long, nerdy one. Pair 1's title is the stronger hook (numbers) and its cover is the weaker one. Best is to swap: A's title ("One prompt, 8 hours, $550") fits cover A; do keep B's cover for the "body" angle.
• Stronger third cover: split the picture, the bright FPS frame left with a big "8 ORA / 1 PROMPT" and the new-model host pointing at it with a smug look on the right, with a tiny before-face inset in the corner: it combines the number hook, the proof of the result, and the host's transformation.