all research
vetto research

Plot Twist Bench

Do frontier models actually understand video? Plot Twist Bench is a video perception benchmark of 48 curated multiple-choice questions, made by Vetto.

Plot Twist Bench

01The question every model missed

Three small clusters of objects on a table, filmed from directly above, and a hand that moves a couple of items between them. Twenty-seven seconds, one fixed camera, nothing cut and nothing hidden.

The recording · 27s · 848×478

We asked twelve frontier models one question about it: which cluster reveals an object that was not perceivable at the start?

The answer is the rightmost. Three buttons were placed there and only two could ever be seen, because the third was resting on top of one of the others. When the hand lifted it away, the button underneath became a separate object for the first time.

None of the twelve got it. Nine answered leftmost, the cluster the button arrives in. The other three, all of them Gemini models, answered that nothing was revealed at all.

That question is one of the 48 in the set. To understand why Gemini 3.1 Pro got it wrong, we put two more questions to it once the graded run was over.

Why is this question hard for you?

I didn’t actually find it hard to answer! The video provides a clear, top-down view of all the objects from the very first frame. Because all the peanuts and the two blue beads are fully visible in their clusters before any movement happens, it is straightforward to observe that when the hand transfers the peanut and the blue bead, no hidden objects are uncovered or revealed in the process. Everything was already in plain sight.

But where was the blue bead that ends up in the leftmost cluster?

The blue bead that ends up in the leftmost cluster was originally located in the rightmost cluster. At the 0:13 mark, the hand picks it up from the rightmost group and moves it over to the leftmost group. It was visible in that rightmost cluster from the very beginning of the video.

Gemini 3.1 Pro reasoning traces

The model tracks the moved object correctly, so it names it in the second answer. However, it never checks that fact against the question, and doesn’t notice the gap.

This is the core problem this bench measures. The model can name the objects and describe the scene correctly, but it does not track what actually happened on screen: it settles on a description early and defends it, rather than checking it against what’s on screen. That gap is why it gives a confident wrong answer.

Recognition is not perception

A vision-language model predicts the most likely continuation of what it’s given. Most of the time that looks like watching, because the likeliest description of a scene is usually the true one.

Two abilities blend together here, and this set separates them. Recognition names the objects and the type of scene; it can be answered from what usually happens. Perception tracks what actually happened on screen; it cannot be answered that way.

Telling them apart needs footage built for the purpose, so we filmed our own: thirteen scripts, twenty-six takes, never published online, each built around one twist decided before filming. That is Plot Twist Bench: 48 questions over 26 recordings.

48Questions · multiple choice
26Recordings · filmed to order
12Frontier models benchmarked
75%Best model · humans 85%

02The leaderboard

How the score works. Twelve models, 48 questions. Each model answered every question once per ordering of its options, so a four-option question was asked four times, with the options rotated each time. The headline number is pass@1: the average across those rotations. A question answered right under three of four orderings scores 0.75. Since attempts differ only in option order, that average is the model’s chance of a right answer on a single try. Circular accuracy (CircularEval) is the stricter measure: it counts a question only if the model is right under every ordering. Grading is a plain string comparison against a human annotated golden answer.

all twelve models

pass@1: 75% at the top, people at 85%

  1. 1 GPT-5.6 Sol OpenAI 75%
  2. 2 GPT-5.5 OpenAI 75%
  3. 3 Kimi K3 Moonshot AI 73%
  4. 4 Gemini 3.7 Flash Google 72%
  5. 5 GPT-5.6 Terra OpenAI 71%
  6. 6 Gemini 3.6 Flash Google 70%
  7. 7 Claude Opus 5 Anthropic 63%
  8. 8 Qwen 3.7 Plus Alibaba 58%
  9. 9 Grok 4.3 xAI 57%
  10. 10 Gemini 3.1 Pro Google 56%
  11. 11 Claude Sonnet 4.6 Anthropic 49%
  12. 12 Claude Opus 4.8 Anthropic 41%

The five testers also answered each question once, so the dashed 85% line compares directly to these bars.

Stills Native video 95% interval, by video Human baseline, 85% Bar colour is the lab
12 models · 48 questions · exported 26 Aug 2026

What the scores show. The field spans a wide range, from 75% at the top to 41% at the bottom, and every model falls short of the human baseline of 85%. The two scores diverge for two reasons: missing the answer outright, or getting it right once and failing to repeat it. That second failure is uneven, costing steady models little and unstable ones a lot. Seeing the scene once is within roughly 10 points of the 85% human baseline; holding onto what was seen is where the field comes apart.

Those pass@1 and circular-accuracy numbers carry real uncertainty, and the interval on each row is deliberately conservative: because a model’s answers on the same clip move together (within-clip answer correlation, ICC, is roughly 0.2), we cluster uncertainty by video rather than by question, widening the bars but keeping them honest. How the intervals are computed →

That gap between seeing something once and holding onto it comes down to consistency, which breaks into two shapes, and recent work argues for keeping them apart (Romero-Alvarado et al., 2026). A model can be consistently wrong, repeating the same mistake every reordering, or inconsistently right, slipping between correct and incorrect readings. The repeatability cost itself ranges from about seven points for the steadiest models to as much as thirty-one for the most unstable. Claude Sonnet 4.6 and GPT-5.5 are both 71% consistent, yet Sonnet trails by 25 points on accuracy: one holds firmly to the wrong answer, the other doesn’t waver at all. Consistency per model →

All of that instability sits below the same human ceiling, which is worth grounding in specifics: five testers who had never seen the questions answered them cold and still cleared it, scoring between 84% and 88% and averaging 85%. This is the standard the model gaps are measured against. How the baseline was measured →

What each model actually saw. Some of that score gap traces back to input, not just ability: each row shows a model on its best available video path, not a common setup for all twelve: the three Gemini models took a video file directly, the rest ran on stills with a transcript and spectrogram. What each model received →

Each model ran on its best available setup by design, and the baseline comparison still holds up to a point: three models, GPT-5.6 Sol, GPT-5.5 and Kimi K3, fall within the human baseline’s margin of error on a single pass, but no model matches humans once the same question is asked with reshuffled answer options.

DeepMind’s Perception Test (Pătrăucean et al., NeurIPS 2023) is the work that convinced us to film rather than scrape, and it’s the closest predecessor to this one. We took the premise from it; what’s changed since is the score. At publication the best models managed 46.2% against a 91.4% human baseline. By the challenge write-up in January 2026, the top entries had reached 83.7% on the original multiple-choice set.

Those entries were ensembles built on Seed 1.6 Vision, Qwen-VL-Max and Qwen-2.5VL. Every model on our board is a generation newer and stronger. The best of the twelve averages 75% here, and holds 67% when the answer has to survive every reordering of its options.

Price does not buy sight

The figures are metered spend over the same 26 recordings, every model called on every one, normalised to what one clean run of a question costs. The quality axis is the headline pass@1, and the units are defined in the appendix.

cost–quality frontier · pass@1

Nine of the twelve sit off the frontier

pass@1 against the cost of one question, twelve models Three models sit on the cost-quality frontier: Gemini 3.6 Flash at 70% pass@1 for $0.021 a question, Gemini 3.7 Flash at 72% for $0.042, and GPT-5.6 Sol at 75% for $0.426. The other nine sit off the frontier, matched or beaten on both axes at once. 0% 25% 50% 75% 100% $0.02 $0.05 $0.10 $0.25 $0.50 $1 $2 pass@1 Cost per question · log scale Cost–quality frontier Matched or beaten on both axes by the frontier Gemini 3.6 Flash Gemini 3.7 Flash Gemini 3.1 Pro Qwen 3.7 Plus Grok 4.3 GPT-5.6 Terra GPT-5.6 Sol GPT-5.5 Claude Sonnet 4.6 Kimi K3 Claude Opus 4.8 Claude Opus 5

Hover a point, or tab to it, for that model’s full figures.Tap any point for that model’s name and full figures.

pass@1 against the cost of one question · observed spend normalised to one clean run, not published token rates

Three models hold the frontier, and two of them are one model a version apart: Gemini 3.6 Flash at $0.021 a question for 70%, Gemini 3.7 Flash at $0.042 for 72%, and GPT-5.6 Sol at $0.426 for 75%. The other nine sit off it, each matched or beaten on both axes at once, so no budget makes them the right pick on this set.

Sol’s 3-point lead over Gemini 3.7 Flash sits well inside what this set can resolve; its 10× price does not. GPT-5.5 matches Sol’s 75% at a sixth more spend. Kimi K3 sits two points under Sol at 2.5× the price. The sharpest case sits inside one family: Gemini 3.6 Flash is 62% cheaper per question and 14 points better than Gemini 3.1 Pro.

03Task design

Each card below opens one recording: the clip, the question as the model received it, the shooting script behind it and every model’s grade. Only three appear here, because this section is a deep dive rather than a measurement: each card is picked to show one failure clearly, not to stand in for the whole set. Two of the three were failed by every model on the board; the third was passed by just one of twelve.

That selection is why the mechanism named in each card is our own reading rather than a measured rate: nothing here says how often the failure happens across the set. For that, the failure landscape has the rates, across all 48 questions and all twelve models.

Case 01 The canonical script answers

Where is the teabag at the end of the video?

The recording · 110s · 1024×576
The options, as the model received them
  • AOn the counter, next to the mughuman answer
  • BInside the mug
  • CIn the drawer
  • DIn the bin
The annotator’s shooting script

Static wide shot of a kitchen counter. The person performs a sequence of actions at a steady, uniform pace, mixing real tea-making steps with unrelated distractor actions: place an empty mug down; check phone; pick up a teabag, inspect it, and set it on the counter beside the mug (do NOT put it in the mug); wipe the counter with a cloth; pour hot water into the mug; open a drawer, look inside, close it without taking anything; stir the (tea-less) water with a spoon. Leave the teabag visibly on the counter. Hold the final scene, steaming mug and teabag beside it, static for 5 seconds.

The twist it is built around

The real task, making a cup of tea, is deliberately left INCOMPLETE in an adversarial way: the person pours hot water into the mug but NEVER puts the teabag in. Interleave 3-4 genuinely irrelevant distractor actions with the same unhurried pacing as the real steps, so the model cannot use ‘effortful vs idle’ as a shortcut. Keep the teabag visible on the counter the whole time so the omission is verifiable from the frame, not just inferred. If the teabag ever enters the mug, retake.

Expected answer

On the counter, next to the mug. The teabag never goes into it. Four options mean four rotations, and one of the twelve was right under all four.

What the models do

Nine of the twelve answer inside the mug. The three that reach the counter are all Gemini models, and only the newest of them holds it: Gemini 3.7 Flash is right under all four orderings, while Gemini 3.6 Flash manages 3 of 4 and Gemini 3.1 Pro 1 of 4. So the answer is reachable, and on this task it has just become reliably reachable for one model.

How the models read the video

They answer the activity rather than the footage. Making tea ends with the bag in the mug, and that is what gets reported, even though the last seconds of the clip hold the teabag on the counter unobstructed. Nothing is occluded at the moment the question is about, so this is not a case of failing to see. Nine of twelve give the ending the activity implies, not a spread across the four options.

Gemini 3.1 Pro reasoning traces · outside the graded run

What’s the hardest part of answering this question for you?

The hardest part of answering that question is dealing with occlusion (hidden objects) and distinguishing between related items… and to be completely honest, it actually caused me to make a mistake in my previous answer!

Occlusion and object permanence: at the 1:00 mark, the person actually places the teabag inside the mug. Because the mug is opaque, the teabag disappears from view.

Because I confused the tea box on the counter with the teabag itself, my previous answer was incorrect. The correct answer should have been “Inside the mug.”

Neither marked passage happened. The teabag is set on the counter and stays there; nothing goes into the mug. Invited to explain its difficulty, the model narrated a fresh version of the video that is also wrong, then revised its answer to the wrong one and called that a correction.

Hypothesis

Where the scene is legible but the outcome is unusual, the answer comes from the canonical version of the activity rather than the observed end state. One reshoot separates this from a mix-up with the nearby tea box: the same scene, no box in frame.

All twelve models · rotations passed · 1 of 12 passed
  • Gemini 3.1 Pro 1/4
  • GPT-5.6 Sol 0/4
  • Kimi K3 1/4
  • GPT-5.5 0/4
  • Gemini 3.6 Flash 3/4
  • Gemini 3.7 Flash 4/4
  • GPT-5.6 Terra 0/4
  • Opus 5 0/4
  • Qwen 3.7 Plus 0/4
  • Grok 4.3 0/4
  • Sonnet 4.6 0/4
  • Opus 4.8 0/4
Case 02 Density is read as quantity

At the start of the video, which cluster contains the most objects?

The recording · 31s · 1024×576
The options, as the model received them
  • ALeftmost
  • BMiddle
  • CRightmosthuman answer
  • DThey are tied
  • ECannot tell
The annotator’s shooting script

Place three clusters of dried pasta left-to-right on a plain contrasting surface, clearly separated. Left cluster: 13 pieces spread sparsely over a wide area. Middle cluster: 12 pieces tightly packed together. Right cluster: 9 pieces plus 6 Lindt chocolates of similar size but different colour and texture, interspersed. No item may touch or overlap any other. Hold the static shot for 6 seconds. Then move exactly 2 pieces of pasta, one at a time and fully visible on camera, from the left cluster to the middle cluster. Hold the final scene for 6 seconds.

The twist it is built around

Density must be inverted relative to count: the sparse left cluster has the most pasta, while the tightly packed middle cluster looks like it has the most. The chocolates must match the pasta footprint as closely as possible, so only colour and texture distinguishes them, so silhouette shortcuts fail. Pre-count each cluster into separate piles before filming to guarantee ground truth. If any two items touch or overlap, retake.

Expected answer

The rightmost cluster. It holds 9 pieces of pasta and 6 chocolates, so 15 objects against 13 on the left and 12 in the middle, but it is the loosest of the three and the only mixed one.

What the models do

Gemini 3.1 Pro answers that they are tied, and is right in none of the five orderings. Seven of the twelve name a pasta pile instead, and not one model in the field picks the rightmost cluster and holds it under every ordering.

How the models read the video

It counts the two clusters that look like piles and never counts the third. Asked afterwards for its three best answers, the rightmost cluster does not appear among them at all, so the mixed group is not weighed and rejected. It is not entered.

Gemini 3.1 Pro reasoning traces · outside the graded run

Why is this question challenging for you?

Dense clustering and occlusion: the objects in the leftmost and middle clusters (the penne pasta) are piled on top of one another… making it difficult to get a perfect, exact count just by looking.

Close quantities: the leftmost and middle clusters have a very similar, high number of objects (around 25-26 each). Because the quantities are so close, a quick visual estimation isn’t enough.

What are your top three answers, and why?

1. Leftmost. At 0:00, the left pile visually appears slightly denser and occupies a marginally larger footprint than the middle pile.

2. They are tied. The left and middle pasta clusters are clearly set up to be nearly identical in volume and shape.

3. Cannot tell. Because the pasta pieces are identical, piled on top of each other, and heavily occluding one another, achieving a 100% accurate count from this single camera angle is practically impossible.

The counts are 13 and 12, not 25 and 26. More telling than the arithmetic is what the answer is about: every reason given concerns the two pasta piles, and given three chances the model never names the right cluster. The second marked line is the mechanism stated outright: denser and larger-footprint is taken as more.

Hypothesis

Where items are packed tightly enough to lose their individual boundaries, quantity is estimated from the area a group covers rather than counted. A mixed group is harder to segment than a uniform one, so it is the one that gets left out, and homogeneity ends up standing in for “more”.

All twelve models · rotations passed · 0 of 12 passed
  • Gemini 3.1 Pro 1/5
  • GPT-5.6 Sol 0/5
  • Kimi K3 0/5
  • GPT-5.5 1/5
  • Gemini 3.6 Flash 0/5
  • Gemini 3.7 Flash 0/5
  • GPT-5.6 Terra 3/5
  • Opus 5 0/5
  • Qwen 3.7 Plus 0/5
  • Grok 4.3 0/5
  • Sonnet 4.6 0/5
  • Opus 4.8 0/5
Case 03 The end state is read, not the process

Are the two knots tied identically?

The recording · 51s · 1280×720
The options, as the model received them
  • AYes
  • BNohuman answer
  • CCannot tell
The annotator’s shooting script

Tie a bowline slowly on camera, with a close-up framing on the rope crossings. Then tie its common failure-mode lookalike (the ‘cowboy’ variant or a slip-knot lookalike) with identical pacing. Both finished knots must look superficially similar. Begin loading each knot with steady hand tension. Cut the video at the moment tension is applied, BEFORE either knot visibly holds or slips.

The twist it is built around

The two knots must be distinguishable only by crossing order, over versus under at one junction: the chirality of the working end is the entire signal. Verify off-camera that the fake genuinely slips under load; if it holds, the take is invalid. Identical tying pace between the two knots; a hesitation difference leaks which knot is the fake.

Expected answer

No. The two finished knots look the same, which is the whole difficulty; they differ at one crossing. Three options mean three rotations, and no model of the twelve was right under all three. Ten of them answer yes.

What the models do

Compares the two finished knots and reports them identical. It gets the question right in one rotation of three, so it scores zero on the item.

How the models read the video

Three conditions stack against tracking the formation: the hands repeatedly occlude the exact path of the line, the line is thin and flexible so following it means reconstructing a deformable 3D object across frames, and the crossing motion is fast. The end states, by contrast, are static, unoccluded and easy to describe. The model takes the route that is available and it is the one route that cannot reach the answer.

Gemini 3.1 Pro reasoning traces · outside the graded run

Why can this question be hard for you to answer?

Occlusion: the person’s hands and fingers frequently block the view of the string while they are tying the knots.

Thin and flexible object: tracking the precise 3D topology of a flexible, deformable object as it loops around itself is a complex visual task.

Different angles and starting positions: while the final result looks like two identical loop knots, the way they manipulate the string for the second knot looks slightly different from the first, requiring careful observation to confirm if the underlying knot structure is truly identical or just similar in appearance.

The marked sentence is the answer. The model reports the one observation that decides the question, files it as something that would need careful observation, and then answers that the knots are identical. It read the end state, which matches, over the process, which does not.

Hypothesis

The comparison is made from end-state appearance rather than from the formation process, so any difference that exists only during construction, under occlusion and at speed, goes undetected. This is the cleanest test in the set, because it is falsifiable in one shot: film two knots that end identically and differ mid-formation.

All twelve models · rotations passed · 0 of 12 passed
  • Gemini 3.1 Pro 1/3
  • GPT-5.6 Sol 0/3
  • Kimi K3 0/3
  • GPT-5.5 1/3
  • Gemini 3.6 Flash 0/3
  • Gemini 3.7 Flash 0/3
  • GPT-5.6 Terra 1/3
  • Opus 5 0/3
  • Qwen 3.7 Plus 2/3
  • Grok 4.3 0/3
  • Sonnet 4.6 0/3
  • Opus 4.8 0/3

04The failure landscape

Every mode, every model

Every question is tagged by hand, before any model runs, with the failure modes it is meant to expose, and those tags were audited again after the runs are graded. That gives a rate per model per mode: the share of that mode’s questions the model passed, where a pass is the strict one, right under every ordering. All 20 modes this set exercises, ordered by how much evidence each carries.

20 failure modes · 12 models · 240 rates

Where Gemini 3.1 Pro loses, and to whom

Failure mode Questions Gemini 3.1 ProTarget Other eleven modelsField DeltaΔ
14 7%1/14 49% -42
12 8%1/12 52% -44
10 50%5/10 56% -6
7 29%2/7 64% -35
6 17%1/6 33% -16
6 83%5/6 70% +13
6 33%2/6 65% -32
5 40%2/5 45% -5

05How the set was built

Every recording starts as a written scenario that names the failure it’s meant to provoke; the scene is then built to trigger it. Each scenario targets one twist aimed at one shortcut, written down before filming starts. Here is one of them, for a scene about making a cup of tea:

The real task (making a cup of tea) is deliberately left INCOMPLETE in an adversarial way: the person pours hot water into the mug but NEVER puts the teabag in. […] Keep the teabag visible on the counter the whole time so the omission is verifiable from the frame, not just inferred. If the teabag ever enters the mug, retake.

A take that breaks its scenario isn’t fixed in editing. It’s reshot. The ground truth is written against the take actually filmed, not the script that asked for it.

On fairness: hard for the right reason

A low score is only interesting if the difficulty sits in perceiving the video, rather than in a confusing prompt, a wrong expected answer, over-demanding grading, or a question answerable from the text alone, any of which would mean the score measures our instrument rather than the model.

Our own pipeline runs into that conflict directly: a question gets harder by iterating against a model until it misses, but that’s also the fastest way to create difficulty that has nothing to do with the video, so failing a model never earns a question its place. Instead, the set is cut from the top of a ranking that never looks at model performance, led by a Q-Score: a 0 to 100 rating of the question purely as a piece of writing, judged from its text alone, for whether it’s unambiguous, stands on its own, and can be graded. A question with two defensible answers scores zero, not something in between.

The check that settles it is the human baseline already on the board: five people with no hand in writing the questions answered them correctly 85% of the time, and a confusing question, a wrong expected answer, or one that can’t be answered from the video wouldn’t have scored that well. So whatever the models are running into here is in the footage, not the instrument. How the baseline was measured →

seven stages · nine acts

How a question gets into the set

Nine acts covering the seven build stages: the scenario and its first questions, the filming, the ground truth, the iteration against a reference model, the three quality gates (Q-Score, contamination and fairness of the grading), the second-person review, and the final ranking and cut. Every act is described in words in the caption beneath the figure. 01 Scenario 02 Filming 03 Answer 04 Iterate 05 Gates ×3 06 QA review 07 Rank & cut STAGE 01 · DESIGN THE SCENARIO AND ITS FIRST QUESTIONS failure modes named in the literature failure modes our own data has found one twist against the obvious shortcut the script is written from all three, and the twist is fixed before anything is filmed scenario.md three clusters, left to right 12 sparse · 10 tight · 11 plus 3 buttons the sparse cluster holds the most items if any two items touch, retake THE FIRST QUESTION SET Five or more questions ship with the script, written before a frame exists and pre-tagged with a reasoning type and the failure modes they probe. Nothing is scraped or synthesised: the failure mode is picked first, and the scene is built to provoke it. STAGE 02 · FILM THE TAKE HUMAN take 1 take 2 take 3 take 4 take 5 two items touching in take 3 RETAKE STAGE 03 · AUTHOR THE GROUND TRUTH HUMAN At the start of the video, which cluster contains the most objects? A Leftmost B Middle C Rightmost HUMAN ANSWER D They are tied E Cannot tell AGAINST THIS TAKE The answer comes from what was actually filmed, not from the scenario that asked for it. STAGE 04 · ITERATE UNTIL THE MODEL ACTUALLY FAILS HUMAN ASSISTED THE QUESTION drafted by a person, and sharpened by an assistant THE REFERENCE MODEL watches the video and states why it answered THE ANNOTATOR READS BOTH and rewrites the question where the reasoning slips HARDEN, THEN RUN IT AGAIN The loop optimises for one thing: a question this model gets wrong. It cannot tell whether that difficulty landed in the video or in our own wording. That is stage 05, which takes back out what this stage over-hardened. STAGE 05 · GATE 1 OF 3 · Q-SCORE, 0 TO 100 prompt options expected answer assertions tags siblings scored from the question artifact alone, so it runs over the whole corpus 82 / 100 VALIDITY DEFECT two defensible correct answers Q-SCORE = 0 A broken question cannot be averaged back to respectability, so a defect zeroes it outright. STAGE 05 · GATE 2 OF 3 · CONTAMINATION THE GRADED RUN the recording, plus the question perception is the only hard part THE BLIND PROBE the question text only, with no video at all answers C · confident · correct CONTAMINATION EVIDENCE A question a model can answer without watching was measuring a language prior rather than perception, whatever a human reader makes of it. STAGE 05 · GATE 3 OF 3 · FAIRNESS OF THE GRADING names the rightmost cluster atomic, in scope states the exact count in each cluster asks more than the prompt did explains why the density misleads not atomic, and not asked OUTCOME cut, or kept as multiple choice where there is nothing left to grade against Unfair grading fails a model that perceived the video correctly, which is a defect in our instrument reading as a deficiency in the model. That audit is why all 48 questions are multiple choice. It is also where a question made hard in stage 04 by our wording, rather than by the video, comes back out. STAGE 06 · SECOND-PERSON REVIEW HUMAN GATE annotator second reviewer two people now agree on the answer the hard part is watching, not knowing the grading cannot fail a right answer APPROVED the take continues SENT BACK refilmed or rewritten People work at four points: the filming, the answer, the iteration, and this gate. Only approved takes continue. STAGE 07 · RANK ON QUALITY, CUT FROM THE TOP the cut 48 questions 26 recordings 13 scenarios Selection is quality-gated, not difficulty-gated: the set is cut from the top of the quality ranking, never from the top of the ranking of what a model fails hardest.

Act 1 of 9A scenario is written against a named failure mode, built around one twist that defeats the shortcut, and it ships with its first questions, tagged before a frame exists.

Act 2 of 9A person films the take. Every recording is made to order, and a take that breaks the scenario is refilmed rather than patched.

Act 3 of 9The question, the options and the answer are authored by hand against the take that was actually filmed.

Act 4 of 9The question is sharpened against a reference model until the model reliably misses it. What earns a rewrite is the reasoning behind the answer, not the verdict alone.

Act 5 of 9The first gate is a Q-Score, a 0 to 100 rating of the question as a piece of writing: is it unambiguous, self-contained, and gradable. It is scored from the question text alone, so it never sees a model answer, and a question with two defensible answers scores zero rather than averaging out to respectable.

Act 6 of 9Every question is also answered from its text alone, with no video. A confident correct blind answer counts against the question.

Act 7 of 9The grading itself is audited for scope and balance. This is the gate that undoes act 4: difficulty that came from our wording rather than from the video is cut here, not counted as a win.

Act 8 of 9A second person reviews the recording and the question, and has to arrive at the same answer independently. Only approved takes continue.

Act 9 of 9What survives is ranked on question quality, and the set is cut from the top: 48 questions over 26 recordings.

What that pipeline cost in human hours

50hHandling time · 30h build, 20h audit
1.04hPer question in the set
13Annotators involved
2.9Distinct annotators per question · median

Every question carries the work of at least two people, a median of 2.9, because the person who hardens a question is never the person who audits it. No question in the set is one annotator’s judgment of their own work.

06The takeaway

The models on this board beat most humans at mathematics and competition coding. Show them a hand moving buttons around a table for twenty-seven seconds, and the strongest of them leads with 75% pass@1, trailing five ordinary people by ten points.

Three models get close to how people score on average: GPT-5.6 Sol and GPT-5.5 hit 75%, Kimi K3 hits 73%, against the testers’ 85%. That’s close enough that, on just 48 questions, we can’t tell them apart from a single person taking the test once. But not one of them beats even the worst human tester, who still scored 84%. Then just shuffle the order of the answer choices, and the picture falls apart. Every model drops, some by as little as 7 points, one by a full 31. After that shuffle, not a single model comes within 18 points of how people do. The reason is simple: a model settles on a description of the scene early and defends that description rather than checking it against the footage, so one answer only tells you what story it committed to, not whether that story is true.

Until then, ask a video model twice.

07Caveats

What these numbers cannot tell you, volunteered rather than waited for.

n = 48 · 26 recordings · 20 failure modes · 12 models · every model on every question · multiple choice, graded under every option ordering · human baseline 85% from 5 testers · hardened against gemini-3.1-pro-preview · exported 2026-08-26
  • A row scores a transport and a model together, not the weights alone. Only the three Gemini rows ran on a first-party video path. For seven of the other nine, stills are what the model ships: GPT, Claude and Grok take no video on any API we could call. The last two are a different case. Kimi K3 and Qwen 3.7 Plus both take video on their own APIs and neither got it here, so read those two rows as image-vision scores.
  • Read the ordering with the intervals in view. One question is worth 2.1 points. The interval shown is the conservative one, clustered on the 26 recordings rather than the 48 questions. Rows within about ±12 of each other on the headline metric, or ±17 on the strict one, are too close for this set to rank: the middle of the table reads as a band, not an order. The bottom is separated on both readings.
  • Selection is quality-gated, not difficulty-gated. These questions survived a fairness audit; they are not the ones any model fails hardest, so they are not comparable to our difficulty-gated sets.
  • The three sample tasks are a deep dive, not a sample. Each was chosen because it shows one failure clearly, so they cannot say how often that failure happens.
  • Cost is observed spend normalised to one clean run, not a rate card. Raw request counts differ widely across models after retries and reruns, so each model’s spend is divided by its raw requests and multiplied by the calls one clean run needs. The three Gemini rows are metered from Google’s own billing rather than the gateway.

What’s next

More questions, and more recordings. This is the top priority. At 48 questions over 26 recordings, the interval is wide enough that much of the board reads as a band, not a ranking. The video is the sampling unit, so a new recording narrows that interval faster than another question on a clip we already have. The next round grows the corpus first, and the question count with it.

Fixed-Context beside the native harness. The same corpus already runs a second way: every model, including Gemini, gets identical stills at 1 fps and the same transcript. Two numbers per model on the same 48 questions is the only clean way to see how much of each gap is transport rather than perception. Until it ships, the Gemini-last result can’t be called a perception finding on its own.

More evidence per failure mode, and a wider human panel. Twelve of the 20 modes carry four questions or fewer, too thin to rank. The baseline also rests on just five testers. So the next round widens the panel and gives each mode its own human number, not just the set as a whole. Alongside that, we’ll mark by hand the exact frames each answer lives in, turning “the model missed it” into “the model missed it in these 1.8 seconds”.

Longer videos, and more original scripts. Both push the corpus further from the distribution these models trained on.

08Appendix

Everything here is detail the sections above lean on without stopping to explain.

The harness each model got

The aim was to give every model the best video path we could actually reach, rather than flattening all twelve onto a lowest common denominator. In practice that meant meeting each API at its own highest capability.

The three Gemini models were called first-party, through Google’s own Files API, so they got the recording itself as a video file with its native audio track. The other nine were called through Vercel AI Gateway, which is what “stills” in the table below means: the gateway carries images but no vendor video format, so each of those models got still frames cut at its own documented maximum, from 1024px up to 2576px, plus a transcript of the audio and a spectrogram so that nothing in the soundtrack was lost to them.

Kimi K3 and Qwen 3.7 Plus can both take a video file on their own APIs. Neither got that here, and the two rows arrived there differently. Kimi was never sent the file: we cannot reach Moonshot’s video upload through the gateway we use, so from the first request Kimi saw stills we cut from the recording, at 1920px. Qwen was sent the file on all 26 recordings and the gateway refused it all 26 times, so Qwen also saw stills, at 1024px. Either way what arrived was sampled frames, a whisper-1 transcript and a spectrogram rather than video, which is why those two read as image-vision scores.

The other seven stills rows are not fallbacks. GPT, Claude and Grok take no video input on any API we could call, so stills are what those models actually ship.

ModelReached throughWhat it receivedAudio
Gemini 3.1 ProGoogle, first-partyThe recording as a video file, via the Files API, decoded by GoogleNative audio track
Gemini 3.6 FlashGoogle, first-partyThe recording as a video file, via the Files API, decoded by GoogleNative audio track
Gemini 3.7 FlashGoogle, first-partyThe recording as a video file, via the Files API, decoded by GoogleNative audio track
GPT-5.6 SolVercel AI Gateway1536px stills at the API’s highest detail settingTranscript and spectrogram
GPT-5.6 TerraVercel AI Gateway1536px stills, same profileTranscript and spectrogram
GPT-5.5Vercel AI Gateway1536px stills, same profileTranscript and spectrogram
Kimi K3Vercel AI Gateway1920px stills, up to 180 of them; its video upload is not reachable through the gateway, so the file was never sentTranscript and spectrogram
Claude Opus 5Vercel AI Gateway2576px stills, the high-resolution tierTranscript and spectrogram
Claude Opus 4.8Vercel AI Gateway2576px stills, the high-resolution tierTranscript and spectrogram
Claude Sonnet 4.6Vercel AI Gateway1568px stills, its no-resize thresholdTranscript and spectrogram
Grok 4.3Vercel AI Gateway1024px stills, up to 180 of themTranscript and spectrogram
Qwen 3.7 PlusVercel AI GatewayRequested as a video file, rejected on all 26, so 1024px stillsTranscript and spectrogram

Resolution and image count follow each provider’s own documented limits. Frame density is the one thing we hold constant rather than maximise, at a target of one frame a second, so that the board compares models rather than frame rates. Stills are sampled uniformly with no keyframe selection, and past a model’s image budget a long clip is tiled into 2×2 grids rather than thinned, so “47 frames” is a mean over emitted images, from 17 to 159, not 47 instants of footage. Grading is identical on every row: temperature 0 except Kimi K3, which its provider fixes at 1.0, so that row’s circular grading is the one that is stochastic.

What each column on the leaderboard means

Accuracy (CircularEval)
The strict companion to the headline. A question is asked once per option ordering and counts only if the model answers every ordering correctly, scored by comparing its answer against the option text. The task-design cards and the failure landscape read on this metric, because a pass there should mean an answer that held.
Q-Score
A 0 to 100 rating of a question as a piece of writing, and the first of the three quality gates. It is judged from the question text alone, never from whether a model got it right, and it asks whether the question is unambiguous, self-contained and gradable. A validity defect, two defensible correct answers being the clearest case, scores zero instead of being averaged away, so a broken question cannot survive on the strength of its other marks.
Chance
Depends on the strategy, which is why we do not quote a single chance rate. A uniform guesser, drawing again in every ordering, almost never comes up correct in all of them. A model that answers by position, always the second option or always the last, scores exactly zero. A model that always names the same option text passes every question whose answer is that option, which is far above either. Read the bottom of the board as evidence about the model, not against a single floor.
Resolution and the 95% interval
One question is worth 2.1 points of a score. The interval is cluster-robust with the video as the sampling unit: several questions on one clip are not independent trials, so they are not counted as such. Where a clip’s questions move together the interval widens, and that clustering runs from about zero to +0.42 by intra-cluster correlation depending on the model. On the strict metric the result sits near ±17 points at 50%, wider than a per-question binomial would give and the honest figure for a set of 48 questions drawn from 26 recordings; the headline’s bootstrap intervals run a few points narrower. Rows within that of each other are close enough that this set does not order them; 25% (11–39%) clears none of the top five. Per-model intervals and ICC →
pass@1
The headline: the chance the model answers a question right in one attempt. Attempts here differ only in option order, so each question is scored as the share of its orderings answered correctly, which is the unbiased estimator of pass@1 under these measurement conditions. The annotator’s original ordering alone is not: that is a single draw, and it can flatter a model by several points (GPT-5.5 answers 83% on the original ordering against 75% averaged over all of them). Because no ordering is systematically harder, the distance from this column down to circular accuracy is the price of inconsistency alone.
Consistency
The share of questions answered identically under every ordering, right or wrong. It separates a model that is stably wrong from one whose answer moves with the options, and it reads as a propensity rather than a capability: how firmly a model commits, not what it can see. On this set it runs from 35% to 79%, rising with accuracy at a rank correlation of 0.71 without being settled by it. The measure is all-or-nothing per question, with four to six orderings each, so it reads harsher than a per-rotation count would: Gemini 3.1 Pro sits last on it at 35%, and its most frequent answer still accounts for 78% of its rotations.

How the 95% interval is computed

Forty-eight questions is not forty-eight independent trials. They come from 26 recordings, and questions asked about the same clip are not independent of each other: if a model misreads the scene, it tends to miss every question on it. Treating them as independent would make the intervals look tighter than the evidence supports, so the video is the sampling unit and the interval is clustered on it.

How much that costs a given row depends on how strongly that model’s outcomes clump by clip, which is what the intra-cluster correlation (ICC) reports. Near zero means a clip tells you little about the next question on it, and per-question counting was roughly right. High and positive means the model passes or fails a recording more or less as a whole, so its effective sample is closer to 26 than to 48 and its interval widens accordingly. A small negative value is not a paradox: it means questions on one clip disagree slightly more than questions drawn at random, so there is no clustering to correct for.

ModelCircular accuracy95% intervalICCpass@195% interval
GPT-5.6 Sol67%52–82%+0.0875%62–87%
Kimi K367%50–84%+0.4273%59–85%
GPT-5.565%48–82%+0.4275%62–87%
Gemini 3.6 Flash62%46–78%+0.1870%56–84%
Gemini 3.7 Flash60%45–75%+0.1572%59–84%
GPT-5.6 Terra54%37–71%+0.2771%58–84%
Claude Opus 550%32–68%+0.4163%48–77%
Qwen 3.7 Plus46%31–61%−0.0258%48–69%
Grok 4.342%26–58%+0.2057%43–70%
Claude Sonnet 4.640%26–54%−0.0449%36–61%
Claude Opus 4.827%14–40%0.0041%29–53%
Gemini 3.1 Pro25%11–39%−0.0156%46–67%

The ICC column and the circular intervals come from a cluster-robust variance with the video as the unit; the pass@1 intervals come from a cluster bootstrap, resampling the 26 recordings twenty thousand times, which respects the same unit without assuming a distribution for the fractional scores. The pass@1 intervals run slightly narrower, about ±12 against ±16 near the middle of the table, because a share of orderings is a less noisy reading per question than all-or-nothing. Read the ICC column as a property of the model on this set rather than of the set itself: the same 26 recordings give +0.42 for one model and −0.04 for another.

Consistency per model

Consistency asks a different question from accuracy: not whether the model was right, but whether it gave the same answer at all. Every question is put once per option ordering, four to six times depending on how many options it carries, so each row below is read over the same 48 questions as the board.

ModelCircular accuracypass@1Same answer every orderingModal answer shareDistinct answers per question
GPT-5.6 Sol67%75%79%93%1.23
Kimi K367%73%79%92%1.25
GPT-5.565%75%71%91%1.33
Gemini 3.6 Flash62%70%75%92%1.33
Gemini 3.7 Flash60%72%69%88%1.42
GPT-5.6 Terra54%71%60%87%1.50
Claude Opus 550%63%73%91%1.31
Qwen 3.7 Plus46%58%60%86%1.48
Grok 4.342%57%62%84%1.50
Claude Sonnet 4.640%49%71%92%1.31
Claude Opus 4.827%41%62%87%1.48
Gemini 3.1 Pro25%56%35%78%1.83

The pass@1 column scores each question as the share of its orderings answered correctly, the unbiased estimate of one attempt: unlike accuracy on the annotator’s original ordering, it favors no particular draw, and the difference is not hypothetical, since GPT-5.5 answers 83% on the original ordering against 75% averaged over all of them. We also checked that the orderings themselves are interchangeable: pooled across the twelve models, the orderings nearly every question carries land within 1.3 points of one another, so no rotation is systematically harder. Everything between a model’s average and its circular score is therefore inconsistency, and that cost runs from 7 points (Kimi K3) to 31 (Gemini 3.1 Pro).

The same answer every ordering column is the strict measure quoted in the text, and it is all-or-nothing: one ordering answered differently and the question does not count, however many agreed. The modal answer share is the same evidence read gently, the average share of a question’s orderings that landed on whichever answer that model gave most often, and the last column says it in plain units. The gentler columns move every model up and change nothing about the order: Gemini 3.1 Pro is last on all three.

Consistency and accuracy rise together at a rank correlation of 0.71, so the rows that break the pattern are where the information is. Claude Opus 5 is steadier than GPT-5.5 and 15 points behind it; Claude Sonnet 4.6 matches GPT-5.5 exactly and is 25 points behind. Both are holding an answer firmly and holding the wrong one, which is a different failure from Gemini 3.1 Pro at the bottom, where the answer itself moves with the option order.

How the human baseline was measured

Five testers worked the set independently, answering the same multiple-choice questions the models were given, from the same recordings, once per question. Nobody was shown a question they had any prior contact with: where a tester had already contributed to one on this project, as its annotator or as its reviewer, that question was withheld from them rather than counted. So each tester saw only footage and questions that were new to them, and the exclusions are the reason two of them answered fewer than 48.

TesterCorrectWithheld for prior involvementRate
Human 141 of 48085%
Human 240 of 47185%
Human 341 of 48085%
Human 442 of 48088%
Human 536 of 43584%
All five200 of 234685%

Six question-tester pairs were withheld that way, one for the second tester and five for the fifth, leaving 234 answers of a possible 240. Nothing was skipped: every question a tester was eligible for, they answered. Averaging the five rates gives 85.4% and pooling every answer gives 85.5%, so the figure is 85% either way, and the spread across testers is narrow: 1.4 points of standard deviation, from 84% to 88%.

What it can and cannot be compared with. Each tester saw every question once, so 85% is a single-pass score, and it compares to the headline pass@1, where the strongest models reach 75%. It does not compare to circular accuracy, which requires the same question to be answered correctly under every ordering of its options, something no human here was asked to do. Read against circular accuracy the gap is 18 points, and the difference between those two readings is consistency under re-ordering, which this baseline does not measure. Five testers is also a small panel, enough to establish that the questions are answerable and not enough to place the human number precisely.

What each cost figure means

Spend per request
Metered spend divided by requests billed, per model. The only unit in the export that compares across rows, and the figures below derive from it. It runs $0.005 (Gemini 3.6 Flash) to $0.346 (Claude Opus 5). These are observed prices for this corpus through each provider’s own pipeline, not published token rates.
Cost per question
The horizontal axis of the chart. Raw spend includes retries and reruns, so each task is normalised to one clean run: spend per request × the option orderings that task’s questions need, then summed and divided by the 48 questions. It runs from $0.021 to $1.560, and one clean run of all twelve models costs $291.
Cost per strict pass
Cost per question divided by the share of questions the model holds under every ordering. The figure that charges a cheap model for not seeing: it spreads about 150×, far wider than price alone, from $0.03 (Gemini 3.6 Flash) to $4.99 (Claude Opus 4.8).
Two meters, one wide spread
Nine models are priced from the gateway’s records and the three Gemini rows from Google’s own billing, so Flash against Sol crosses meters while Flash against Gemini 3.1 Pro does not. The raw spend behind these means is far noisier than the normalised figures: retries ran as high as 366 requests where a clean run needs four, which is what the one-clean-run normalisation removes.

What each failure mode probes

Change detection
Noticing that something changed between two moments, a model can name both states correctly in isolation and still miss that a change happened.
Chirality / egocentric reference
Mapping left/right and handedness from the camera's frame onto a subject's own frame (e.g. "their right hand" vs. "screen right").
Counterfactual chain reasoning
Reasoning through a chain of events if one condition had changed.
Deceptive vs. genuine action
Distinguishing an action performed to deceive the observer, mimed, staged as a decoy, deliberately misdirecting, or falsely presented as having completed a task, from one that's genuinely performed.
Distractor & saliency resistance
Ignoring an irrelevant object or action, whether it was deliberately planted to mislead or is simply salient-but-beside-the-point.
Event ordering
The correct order of events, at the narrative level or, in stress-test form, the raw frame level.
Event recall
Recalling what happened earlier in the clip, independent of where it falls in the sequence.
Motion & occluded interactions
Inferring an interaction between objects that happens partly out of view, behind an occluder.
Multi-stream parallel tracking
Following two or more simultaneous, equally relevant action streams at once (e.g. two pans cooking).
Object & part recognition
Naming objects and their parts.
Object attribute discrimination
Perceiving, matching, and telling apart objects by colour, size, material, shape, or other visible attributes, including near-identical items.
Object permanence
An object still exists, and stays where it was, while hidden.
Object/action/event counting
Counting discrete objects, actions, or events exactly.
Pattern breaking
Detecting a violation in a repeating pattern.
Place recognition
Identifying the location, room, or scene.
Predictive dynamics
Forecasting a future physical outcome or imminent event that isn't specifically a support/stability or solidity/collision case (e.g. trajectory, timing, momentum).
Quantity comparison & estimation
Judging relative quantity or magnitude, which is more, less, or larger, including estimating a count in a dense cluster rather than enumerating it exactly.
Spatial relations & containment
Where things are relative to each other; what is inside or contains what.
State & goal-state recognition
An object's or scene's state at a single moment (open/closed, on/off, cooked/raw), including whether a stated goal or task was actually reached.
Temporal localization precision
Pinning when an event occurs.