I. First, Clarify What Is Actually Being Recovered

Before comparing models, one fact has to be stated plainly: there is no prompt inside a video file.

An MP4 holds pixels, a frame sequence and an audio track. It does not hold the text that generated it. So “reverse-prompting” means asking a multimodal model to look at the finished result and reconstruct a description that could reproduce it. Published walkthroughs describe the recovered artifact the same way: it is an approximation built from what the model sees, not the exact string the original creator submitted.

That framing yields two practical consequences:

This report therefore does not score “who guesses the original prompt best” — that target does not exist. It scores four measurable things: can the model read the video at all (input capability), is the resulting shot language professional, is the temporal and camera reconstruction accurate, and does the output land cleanly on a downstream generation model.

1.1 The five stages of a reverse-prompting pipeline

Video-MME’s benchmark scale is 900 videos, 254 hours, 2,700 QA pairs, spanning 11 seconds to 1 hour — see the arXiv paper. The “approximation, not the original string” framing and the recovery limits are from a reverse-engineering walkthrough.

II. The Dividing Line: Native Video, or Frames Only

Sort the nine candidates by how they accept video and the picture becomes immediate. The commonly missed point: both 2026 Western flagships sit outside this capability entirely.

Why Tier C is a structural problem. Sampling collapses three dimensions — time plus two of space — into a stack of flat images. Shot size, composition and lighting survive intact. Camera movement (is the lens moving or the subject?), acceleration, and cut points do not: those can only be pinned down by watching continuously. OpenAI’s model page lists GPT-5.5 input as text and images; Anthropic’s documentation likewise confirms Claude accepts text, images and PDFs only, with no video path and no native audio.

If your goal is to replicate a camera move, Tier C starts at a disadvantage. If your goal is to replicate a look — grade, composition, lighting character — sampling is acceptable. That is precisely why Section V routes by scenario.

One more detail worth internalising: every model degrades with video length, long video is the field’s shared weak point, and published scores move with frame-sampling settings. Treat any leaderboard number as a guide, not a guarantee.

III. Nine Models at a Glance

Pricing: Gemini 3.1 Pro figures are the ≤200K input tier; above 200K the rate moves to $4 / $18. Doubao Seed 2.1 Pro bills video analysis by duration, in 15-second units, per Volcano Engine’s documentation. Cost per video hour varies enormously across the field — from free to roughly $4.32/hour — and the third-party video-understanding ranking is useful for magnitude, though some of its scores inherit a predecessor or family estimate. Treat it as directional only.

The current leading group on VideoMME (sources disagree on exact figures — do not compare across sources): Doubao Seed 2.1 Pro 89.2, Qwen3.7-Plus 88.0, MiMo-V2.5 87.7, Kimi K2.5 87.4, MiniMax M3 85.4. The front of the video-understanding field is now held largely by Chinese vendors — and that capability is precisely the first gate for reverse-prompting.

IV. Model by Model

V. Choosing by Scenario

Reframe the question from “which model is best” to “which model for this shot” and the answers separate immediately. Each scenario below carries a recommendation and the prompt detail that decides success.

Scenario breakdowns and prompt checkpoints follow a reverse-prompting guide that gives layered templates for ads, UGC, cinematic and product scenes, and proposes the “model-agnostic core plus adapter” approach.

VI. After Recovery: Adapting to the Target Generator

This is where many recovery pipelines quietly fail. No matter how accurate the recovered text is, skipping the target model’s structure costs more quality than upgrading models would gain.

6.1 Three structures, three orders

6.2 Two hard rules

Rule one: one dominant camera move per shot. Google’s guidance recommends choosing one primary camera move per clip, and an internal test of 200 prompts puts the sweet spot at two to three camera modifiers with a single dominant move. Past three, the model averages between conflicting instructions and the output turns muddy.

Rule two: separate the creative core from the format adapter. Extract subject, action, environment, composition, camera, light, timing, style, audio and constraints into one model-agnostic list, then write an adapter layer per target. That way, a downstream model being discontinued or rebuilt does not invalidate your prompt library — which matters in 2026 specifically: the Sora consumer app shut down in April and the API was retired on September 24, so any Sora-specific syntax has to be migrated.

Source: a Veo and Kling camera-prompt comparison covering the five-part Veo order, the rationale for Kling’s camera-last order, and the 2–3 modifier finding.

VII. Scoring: Seven Weighted Dimensions plus a Sensitivity Check

Important disclosure: the scores in this section are a subjective quantification by this report, based on public documentation and third-party comparisons. This is not an official benchmark and is not a claim about any model’s general capability. It evaluates the specific task of reverse-prompting video, nothing more.

7.1 Dimensions and weights

7.2 Overall scores

7.3 Score breakdown

Legend: ✓ clearly leading on this dimension ｜ △ partially capable ｜ ✗ missing ｜ — not applicable. Sub-scores are subjective integers on a 0–10 scale.

7.4 Dimension winners

7.5 Weight sensitivity check

The default weights assume video input plus temporal fidelity matter 37% combined. Strengthen that assumption — video input to 25%, temporal fidelity to 22%, for 47% combined, while reducing cost and Chinese — and what happens?

The conclusion is clean: the more weight you place on genuinely reading motion, the more the four native-video models gain and the two frame-sampling flagships lose (−0.4 each). Conversely, if your recovery target is really a look rather than a movement, raise cost and Chinese weighting and Doubao and GLM climb quickly. There is no absolute winner — only a weighting choice.

7.6 Head-to-head

VIII. Decision Framework

IX. Scope Limits and Risk Notes

X. Key Findings

1. The gate for this task is “is there a video path”, not “how smart is the model”. The two strongest writing models of 2026 both lack the single most important input for reverse-prompting. That is structural, not a version lag.
2. The leading group has been redrawn by Chinese vendors and Google. ByteDance posts the highest public video-understanding figure with the billing model best matched to the task; Google retains the deepest shot language and longest video access; Alibaba and Moonshot hold the middle with two-hour ceilings and open weights.
3. Cheap and capable are not in conflict here. Doubao’s duration billing brings an hour of analysis to roughly ¥0.48 while holding the top benchmark — the opposite of the usual “cheaper means worse” intuition on AI workloads.
4. The durable output of reverse-prompting is the vocabulary, not any single prompt. Camera terms, lighting descriptors and shot structures transfer across projects; one recovered prompt does not. Filing recovered fragments by purpose is where the compounding value sits.
5. Cross-model adaptation is systematically underrated. Veo camera-first, Kling camera-last, Seedance multi-shot in one prompt — the same description in a different order is a different result. A good adapter layer beats a more expensive model.
6. Reverse-prompting is turning from a trick into a pipeline stage. At roughly 25 seconds for a ten-minute clip at high precision, it now behaves like upstream material processing rather than a one-off inspiration tool.

XI. Related Reading

This report is part of DeepVideo’s ongoing content research. Related technical and market reports are collected here:

• The full DeepVideo report series — benchmarks, version comparisons and third-party tool reviews in one place
• DeepVideo Video Translation: User Guide (2026) — end-to-end operations from project setup to subtitle editing and final export
• Global Short-Video Platforms 2026: Creator Opportunity by Country & Platform — the demand-side view of where a recovered prompt’s output eventually has to land
• Open-weight options referenced above — InternVL is a further self-hostable multimodal family worth evaluating if the Qwen and Kimi routes do not fit your stack