Independent visual research / Frame referenceFrame notes 05 Sep 2026 · Moving review pending
Motion fieldnotesOpenAI film study
About this study

New source reference / Three close examples

The details between states.

A staged title, a 34 ms caret return, and a logo that keeps settling. Three brief, silent excerpts from OpenAI’s GPT-5.4 Thinking tutorial, with the existing fieldnotes playback controls.

Official source / OpenAI

Interrupting and Adding Details in GPT-5.4 Thinking

Watch the complete official video ↗

Published March 5, 2026. The full source remains on the official channel; only the brief excerpts below are included here.

Source video stream
25.725 seconds · 771 native frames
Included excerpts
4.937 seconds total · three non-overlapping windows
Timing
Native 33/34 ms timestamps; no frame-rate resampling
Download
Reference notes and timing JSON

Text revealsFrame-inspected reference

A title assembled in uneven steps

Source 0.000–1.101s · frames 0–32 · 1920 × 1080 native crop

Interrupting and Adding Details in GPT-5.4 Thinking · OpenAI · 00:00–00:01 Official source ↗
33 source frames · 29.970 fps · silent

Frame controls load this short excerpt into the browser for reliable seeking, even without server byte-range support.

Click a sampled state to seek. These anchors are not exact event onsets.

Observed. A white first frame is followed by whole-word and partial-word states. Interrupting appears at 0.033s; the second line begins at 0.434s; the complete title is present at 0.834s. Each line keeps its left origin instead of recentering as text is added.

Choreography. Lead-in → grouped words → smaller steps through the model name and Thinking → hold. The visible increment sizes and dwell times vary. No insertion caret was resolved in the title.

Adaptation. For original presentation copy, set line breaks first, then reveal deliberate groups. Treat any new duration, font, or spacing as an adaptation choice, not a recovered OpenAI preset.

Limit. This excerpt ends at 1.101s after the title has settled; it does not include the remaining title hold or the cut at 2.636s. Consecutive stills establish these sampled states, not a measured easing function or perceived playback smoothness. No sound claim is made.

Inspection: Two source-first consecutive still-preview traversals of all 771 source frames at one-third scale, followed by selected native-resolution regions. The independent record was sealed before comparison. Informed rechecks corrected the caret and late-logo notes. This is not a completed moving-video or listening review, and not an every-native-pixel claim. Native-cadence encode; stepping uses decoded frame timestamps, not a 30 fps assumption.

Context: Source, timing and inspection limits

Interface feedbackFrame-inspected reference

The one-frame caret return

Source 8.775–9.509s · frames 263–284 · 900 × 280 native crop

Interrupting and Adding Details in GPT-5.4 Thinking · OpenAI · 00:08–00:09 Official source ↗
22 source frames · 29.970 fps · silent

Frame controls load this short excerpt into the browser for reliable seeking, even without server byte-range support.

Click a sampled state to seek. These anchors are not exact event onsets.

Observed. The insertion bar appears for source frame 269 alone, from 8.975 to 9.009s. At frame 270, Follow up replaces Ask anything, the plus becomes lighter, and the caret is absent again. It returns at frame 279, 9.309s.

Choreography. Quiet field → one-frame caret → coordinated placeholder/plus change → nine-frame caret absence → visible caret. The stop, microphone, and Thinking controls remain in the crop. The primary notes initially missed the 34 ms return; both informed rechecks recovered it.

Adaptation. When editing a real interface capture, check the frames immediately around a state change, including the preceding blink cycle. Preserve actual feedback rather than adding a click ring or inventing a keystroke.

Limit. Native bottom-window crop [540,740,1440,1020) from the 1920×1080 source. The extra footer/background context keeps browser playback controls away from the composer; the upper conversation is outside this excerpt. The visible bar does not establish focus loss, mouse versus keyboard input, implementation latency, or sound.

Consecutive-frame check · 5 frames

The caret is absent, returns for f269 alone, then disappears at the placeholder replacement. These source crops preserve native pixels; the page may display them at a different size.

Scroll the strip horizontally; click any frame to seek.

8.909–9.042s · every decoded frame in this short window, not the entire film.

Inspection: Two source-first consecutive still-preview traversals of all 771 source frames at one-third scale, followed by selected native-resolution regions. The independent record was sealed before comparison. Informed rechecks corrected the caret and late-logo notes. This is not a completed moving-video or listening review, and not an every-native-pixel claim. Native-cadence encode; stepping uses decoded frame timestamps, not a 30 fps assumption.

Context: Source, timing and inspection limits

Logo choreographyFrame-inspected reference

Drift, remove, morph—then allow the settle

Source 22.623–25.725s · frames 678–770 · 560 × 270 native crop

Interrupting and Adding Details in GPT-5.4 Thinking · OpenAI · 00:22–00:25 Official source ↗
93 source frames · 29.970 fps · silent

Frame controls load this short excerpt into the browser for reliable seeking, even without server byte-range support.

Click a sampled state to seek. These anchors are not exact event onsets.

Observed. The wordmark replaces the interface at 22.756s. A small rightward drift precedes right-to-left character removal, beginning at 23.690s. At 24.024s, the final C is replaced directly by a six-loop mark, which grows and changes its internal geometry into the knot.

Choreography. Cut → initial hold → rightward anticipation → stepped letter removal with continuing translation → loop growth and deformation → small late contour refinements. The frame 744/745 pair still changes; ending the action at the apparent settle near 24.825s would be premature.

Adaptation. For an original mark, separate anticipation, text removal, shape change, and settling. Keep enough ending context to inspect the last small adjustments. Do not reuse OpenAI's mark as a production asset or label a proposed easing curve as measured.

Limit. Native center crop [680,410,1240,680), including four preceding interface frames and the final source frame. The exact subpixel drift onset and final perceptual settle remain unresolved. Small decoded-pixel differences are not automatically intentional animation; no exact motionless-tail duration is claimed.

Inspection: Two source-first consecutive still-preview traversals of all 771 source frames at one-third scale, followed by selected native-resolution regions. The independent record was sealed before comparison. Informed rechecks corrected the caret and late-logo notes. This is not a completed moving-video or listening review, and not an every-native-pixel claim. Native-cadence encode; stepping uses decoded frame timestamps, not a 30 fps assumption.

Context: Source, timing and inspection limits

Reference frame

Enlarged reference frame

Watch this moment on the official channel ↗