Benchmarks Studio · matched media comparison

One still. Two saved clips. Judge the motion before you know the model.

The same still bytes and the same recorded settings went to two Studio jobs. The model was the only intended variable. Both clips are the saved files, served through an opaque route; the labels, model ids, providers and exact settings stay hidden until you record a choice.

Task
Animate one supplied still v1
Inventory
studio-media
Compiled
2026-11-16 10:08 UTC
Source digest
d20cbf5
Sample
2 models · 2 saved clips · 1 pair · no repeated trials

This is one matched pair of saved jobs, not a leaderboard. It has no repeated trials, so it says nothing about variance. A vote recorded here is one unauthenticated label's preference, not the owner's verified judgement.

The input both jobs received

Preview of the supplied still: The line escapes the margin

The line escapes the margin 1536×864 image/png, 1,363,238 bytes

Source artwork credit: Woodhouse · gpt-image-2.6-flare · OpenAI Image API. Public visibility reviewed 2026-11-14. sha256 7758a17cd5e12dd1…

Both jobs attached these exact bytes (same sha256). The recorded prompt names paper grain and a warm lamp glow, which describe a different study; the attached still is the burnt-orange margin artwork. The conflict is published, not repaired.

Open the public source artwork

Shared recorded settings

recipe
animate-a-still
durationSeconds
5
aspect
from-image
budgetAdmission
atomic-v1

Exact prompt text is withheld because the Studio project is private. Its digest is fbba438461f4048e…. Public summary: A slow push-in on the supplied still; hold the palette, grain and marks steady; no new objects, text or camera shake. The exact prompt text is withheld because the Studio project is private. Its digest is recorded so the two jobs can be shown to share it.

One pair, labels hidden

Both clips are playable.The page shows each take's job latency, recorded Studio cost estimate and observed checks before the reveal. Those facts do not name a model. Pair order and the A/B side come from the record's hash rule, not from configuration order.

Pair 1

Two saved takes from the same still and the same recorded settings.

Pair 1
Take A5.18s clip · 17.1s saved-job latencyStudio estimate US$0.24 · provider charge not recordedFails: the clip holds the supplied still · Fails: no new objects or scene change
Take B5.04s clip · 141.2s saved-job latencyStudio estimate US$0.35 · provider charge not recordedHolds: the clip holds the supplied still · Holds: no new objects or scene change
Add a reason:

The server stores each observation as its own row. The short text field and the selected chips are saved with that row; a later observation adds another row and never replaces an earlier one. This browser keeps a random token in localStorage so it can show its own recorded observations. The token is not an account and is not an identity; only its SHA-256 hash appears in the public record.

Record a blind choice

  1. Play both clips. The poster is a frame from the saved file, not a separate still.
  2. Choose A is better, B is better, Tie or Can't tell. The server resolves the pair and appends a submission.
  3. The response reveals the model, provider, exact settings, dimensions, byte count, digest, latency and cost basis for each take.
  4. Reloading the page restores the reveal from the server for this browser token. Another browser never unblinds this page.
  5. After a recorded observation, Record another observation appends a second row. The earlier row is kept and both are counted.
  6. Use Reveal without choosing to see the labels first. That is recorded as unblinded and is never counted as a blind preference.

A self-declared display label, not a sign-in. It is stored with the observation; no email, account id or commenter data is written.

Observations are unauthenticated. Submissions are stored server-side, but this page does not sign anyone in. Anyone can submit under any reviewer label and any browser token. The record counts independent submissions, not verified people; a label called “owner” is not the owner's verified judgement. A random token in this browser's localStorage restores this browser's own observations and is published only as a hash.

The server vote record is being checked…

What this pair does not prove