We Ran the Same Six Prompts Through Kling, Veo, and Runway. Here Is the Scored Result.
Six briefs. Three tools. The same prompt, word for word, in every tool: a serialized-drama scare scene, an everyday life scene, a photo-to-life emotional hook, a sci-fi art portrait, a cute-creature atlas clip, and a food-hall promo. Every run is logged — credits before and after, retries, wall-clock time, export specs, watermark status verified frame by frame. This is the scored evidence article behind the comparison frame in our Sora Migration Center (Piece 1).
Plans tested: Kling Standard ($6.99 first month, then $8.80/mo; 751 credits shown), Google Veo 3.1 Quality via Flow on Google AI Pro ($19.99/mo list; 1,000 Flow credits/mo), Runway Standard ($15/mo; 625 credits/mo, Gen-4.5 text-to-video). Every clip below took exactly one generation, zero retries — all three tools are past the “does it work at all” bar. What separates them is fidelity, format, cleanliness, and cost.
Watch the actual outputs
These are the real exported clips from the 2026-10-23 runs — the same files our scores and frame checks come from. Playback copies on this page are compressed for the web; the export specs, costs, and watermark status in each caption are the logged originals.
T1 — Drama scare scene (《午夜信号》Ep1)
T2 — Life scene (rainy-night delivery rider)
T3 — Photo-to-life hook
T4 — Alien Muse Vela (sci-fi art)
Runway Gen-4.5Not run. Gen-4.5 text-to-video locks 16:9 (see T1/T2/T6); running it horizontally would not test the vertical brief. Its absence is itself a logged result.
T5 — Star Pet BoBo (creature atlas)
Runway Gen-4.5Not run — same 16:9 lock as T4.
T6 — Food-hall promo (generative control)
The evidence, brief by brief
| Brief | Kling 4.0 | Veo 3.1 Quality | Runway Gen-4.5 |
|---|---|---|---|
| T1 — Drama scare scene (《午夜信号》Ep1) | Only tool that fully rendered the scripted beat: second figure in the CCTV grid + heroine turns to an empty store. Mirrored gibberish in the timestamp overlay. | Beat partial: distant figure on the aisle feed, not “a person behind her.” Best face of the three; zero text artifacts. | Beat not verifiable from stills; gibberish signage (“2P XSC”); output 16:9, not vertical. |
| T2 — Life scene (rainy-night delivery rider) | Hand-over beat complete; steadiest composition under the awning. | Beat complete with the fullest arc (wave → receive); strongest neon mood; ~30 min wall-clock, almost all queue; 720p export with slight black bars. | Beat complete; fastest run of the day (~3.5 min); Western-street casting read (prompt did not specify — interpretation difference, logged as such). |
| T3 — Photo-to-life hook | Strongest: clear closed-eye blink frames, smile building to a laugh, lip movement reads as speech. | Continuous smile animation; micro-expression detail sampled weaker than Kling. | Moves (closed-lip smile → open smile); most subtle; no clear blink frame sampled. |
| T4 — Alien Muse Vela (sci-fi art) | Blink-stardust confirmed in dynamic QA (stardust tear-trails, close-up) — but framing pushed to an extreme close-up; the brief’s mid-shot composition (observatory, gown) is lost in that frame. | Blink-stardust confirmed (golden particle burst, profile). Mid-shot composition complete — but the tool added pointed ears the character roster never specified. | Not run. Gen-4.5 text-to-video locks 16:9 (see T1/T2/T6); running it horizontally would not test the vertical brief. Its absence is itself a logged result. |
| T5 — Star Pet BoBo (creature atlas) | Spin + bubbles confirmed in dynamic QA; design stable through the spin. Landing squish does not occur (clip ends airborne) — 2 of 3 motion beats. | Bubbles + rise confirmed; spin unclear; no landing squish — and a consistency failure still frames hid: mid-clip, BoBo’s arms morph into multiple octopus-like tentacles. For a serialized atlas, that is a fail. | Not run — same 16:9 lock as T4. |
| T6 — Food-hall promo (generative control) | Strongest single-frame food appeal: ramen close-up with chashu, ajitama, scallion, nori, steam — plus a soup-dumpling chopstick shot. | Steam dumplings + bamboo steamer, strong neon, good appetite appeal. | Dumplings + ramen in one frame, cinematic. All three stayed clean on text — because signage stayed out of focus (a prompt technique worth copying). |
Cross-cutting facts (all frame- or ffprobe-verified)
- Format: Kling and Veo delivered true vertical (1080×1920 / 720×1280–1080×1920). Runway Gen-4.5 text-to-video was locked to 1280×720 landscape in every run — the 9:16 control was not selectable. Runway’s image-to-video path (Gen-4 Turbo) does offer 9:16 (found in T2): on Runway, vertical is a different workflow, not a setting.
- Watermark: Kling’s standard export carries a “KlingAI 4.0” watermark, bottom-right, frame-verified by us — an early run report claimed none; our re-check corrected it the same day (see /corrections). Whether Kling offers Standard subscribers a separate no-watermark download path is still unverified — until we check it in the logged-in UI, treat the watermark as present. Veo and Runway exports: no watermark.
- Garbled text: two of three tools invented signage/timestamps despite a “no text” instruction. Veo was the only tool with zero garbled on-screen text across the battery.
- Freeze check: dynamic QA (2026-10-24) found no freeze segment ≥1s in the T4/T5 clips reviewed.
Scores — protocol weights, evidence only
Weights (protocol v1 §4): output quality for the job 30% · cost per usable finished minute 25% · speed/reliability 15% · languages (EN/ES/ZH) 10% · rights/terms clarity 10% · editing control 10%. Our full methodology is published on /how-we-test.
Scored today (logged evidence exists):
| Dimension (weight) | Kling | Veo | Runway | Basis |
|---|---|---|---|---|
| Output quality for the job (30%) | 4 / 5 | 4 / 5 | 3 / 5 | Kling wins fidelity (T1 beat, T3 micro-expression, T6 food) but loses T4 framing; Veo wins cleanliness + faces but fails T5 character consistency and softens T1’s beat — a genuine tie at the headline, split on different strengths. Runway’s work is competent, but landscape output fails the vertical job it was briefed for. |
| Cost per usable finished minute (25%) | 5 / 5 | 3 / 5 | 2 / 5 | $9.54/min at intro price ($12.00 at renewal) vs $15.00 vs $17.28 — retries included (there were none). Runway’s minute is also landscape, worth less to a vertical workflow. Kling’s watermark is the hidden cost inside the cheapest number — see the commercial note below. |
| Speed / reliability (15%) | 4 / 5 | 3 / 5 | 4 / 5 | Average wall-clock across briefs: Kling ~8 min, Runway ~10 min (fastest single run: T2, ~3.5 min), Veo ~15 min with queue variance from ~9 to ~30 min. All three: zero retries on every brief. |
Not scored — and why (no score is published without a log behind it):
- Languages (10%) — NOT SCORED. Sprint A’s battery ran EN prompts only. An ES/ZH generative pass has not been run; the avatar battery (Sprint B) is where the frozen EN/ES/ZH language test lives.
- Rights / terms clarity (10%) — NOT SCORED, publish-gate item. Our logs record the tier used; the full terms quotes the protocol (§6) requires before publishing a rights score have not yet been archived. The one rights-relevant fact we have verified — Kling’s watermark — is reported above as an observed export fact, not a terms score.
- Editing control (10%) — NOT SCORED for generative tools. Sprint A tested generation, not re-editing. The one workflow-control finding we do have is Runway’s path split (text-to-video locked horizontal vs image-to-video offering vertical), reported as a fact above. The script-change re-render test belongs to the Sprint B avatar protocol.
Provisional Sprint A ranking — weighted over the scored 70% only, renormalized:
- Kling — 4.36 / 5 (quality 4, cost 5, speed 4)
- Veo — 3.43 / 5 (quality 4, cost 3, speed 3)
- Runway — 2.86 / 5 (quality 3, cost 2, speed 4)
This ranking becomes a full protocol score only when the three unscored dimensions are run and logged. If the rights pass confirms Kling’s watermark cannot be removed on Standard, the commercial ranking flips — Veo’s clean export is worth more than Kling’s price gap for client or brand work, exactly as Piece 1 states. How we handle affiliate relationships is on /disclosure.
Who should pick which (from the logs, not the marketing)
- Vertical drama / serialized shorts on a budget → Kling, if a watermark is acceptable for the channel or until its removal path is verified.
- Commercial or client-facing clips, faces that must stay clean, text-heavy scenes → Veo, and budget for queue time.
- Speed on simple scenes, or an image-to-video workflow → Runway, entering through image-to-video for anything vertical; do not build a vertical pipeline on its text-to-video mode.
Changelog
- v1 — 2026-10-27 draft (publish date at deploy): First scored publication of Sprint A (T1–T6 runs 2026-10-23, dynamic QA 2026-10-24). Three of six protocol dimensions scored from logs; languages, rights/terms, and editing control explicitly unscored pending their runs.
- v2 — 2026-10-30: The actual exported clips are now embedded above (“Watch the actual outputs”), and the homepage duel now plays the real T2 clips instead of a rendered comparison frame. No scores, claims, or findings changed.