top of page
Version 2026 Evaluation
The Numbers
Behind the Gap.
Internal benchmarks on NVIDIA Blackwell B300 and H200, side by side with the leading video models.
Benchmark categories led
29 of 39
Realised inference efficiency
≈67×
Cheaper per minute, like for like
12 to 21×
Output seconds per second, six layers
0.5
Describe the Change
Borrowed From the Brain.
Built in Mathematics.
Shape Before Detail
Like human vision, it reads the overall shape of things first, through topology, and keeps every fine detail alongside.
Fire Together, Wire Together
Sights and sounds that keep appearing together grow linked, the Hebbian way. Contradictions weaken the link, and correlation is never mistaken for cause.
Memory Like Ours
Episodic memory for moments, semantic memory for facts and procedural memory for skills, merged into one working state.
Learn From One Example
One exemplar is enough to recognise the same thing in a new style, the way a child knows a letter in any font.
Out of Sight, Still There
Objects persist when they are hidden, and a moment of surprise marks the start of a new event.
Recall From a Fragment
Holographic memory recalls the whole from a partial cue, and degrades gracefully instead of failing.
Method
How We Measured.
Wishtales GSD
Measured internally on NVIDIA Blackwell B300 at FP8 and on NVIDIA H200, using the same model and the same prompts. Render times are wall clock, end to end, for all six layers.
What B300 Adds
Headroom for 4K at 50 fps across the full pass, and the Video Reasoning Layer’s extra quality passes.
Leading Video Models
Seedance, Veo, Kling, Runway and Sora, from public documentation, hosted medians and list pricing. Price comparisons use like for like premium tiers with native audio only.
What Comes Next
H200 render times and the full 39 category breakdown will be published with the next evaluation.
bottom of page