Detailed Analysis
A new independent benchmark called TWIST (the acronym referring to its Rubik's cube-solving mechanic) offers a striking demonstration of how a frontier model handles a genuinely embodied, vision-only reasoning task. Rather than feeding the model a symbolic representation of the cube's state—a 54-character string or array denoting sticker colors—TWIST forces the model to interact with a live 3D cube purely through PNG screenshots and keystroke commands. If Claude Opus 5 wants to know what color is on a given face, it has to take a virtual photograph and interpret the pixels itself. This is a meaningfully harder problem than the text-based cube-solving demonstrations that have circulated before, because it removes the crutch of a pre-parsed board state and requires the model to build and maintain its own internal spatial model of a 3D object from 2D visual snapshots alone.
The results are notable both for what succeeded and for what the numbers reveal about how the model spent its effort. Given a 20-move scramble, Opus 5 solved the cube in 83 moves—far from the theoretical optimum of 19 moves calculated via Kociemba's algorithm, but a legitimate solve nonetheless. The process took 44 minutes of wall-clock time, but the breakdown is the real story: roughly 2,617 seconds were spent thinking, compared to just 25 seconds of actual cube manipulation. In other words, over 99% of the model's "work" was internal deliberation—reasoning about cube state, planning move sequences, and cross-checking its mental model against screenshots—while the physical execution of turns was almost instantaneous by comparison. The model took 73 screenshots to verify state as it went, consumed 240k tokens at maximum reasoning effort, and did not peek at any answer key. This pattern illustrates a broader characteristic of current reasoning-heavy models: they front-load enormous amounts of latent computation into planning and self-verification, treating the "acting" phase as almost trivial once a plan is formed.
The benchmark's design also anticipates an obvious critique of any Rubik's cube task: solving algorithms are ubiquitous in training data, so a model's apparent problem-solving prowess could be partly memorized retrieval rather than genuine spatial reasoning from scratch. The author acknowledges this directly and has built in a control specifically to address it—a "mirror cube" variant that shares the same twisting mechanism but has no color cues at all, requiring the solver to reason purely from the shape and geometry of pieces rather than pattern-matching against known color-based algorithms. This is a thoughtful methodological choice, since it decouples "the model knows Rubik's cube algorithms" from "the model can perceive and reason about 3D structure through vision alone." The fact that this control hasn't yet been run, combined with the single-sample (n=1) result, means the finding should be read as a compelling case study rather than a definitive claim about visual-spatial reasoning capability.
More broadly, this kind of benchmark reflects a growing trend in AI evaluation: moving away from clean, text-native tasks toward messier, perception-grounded challenges that better approximate how an embodied or agentic system would have to operate in the real world—inferring state from raw sensory input, acting on an environment, and verifying outcomes without privileged access to ground truth. The extreme skew toward "thinking" tokens over "acting" tokens also mirrors a broader shift in the field toward test-time compute scaling, where models like Opus 5 are increasingly evaluated not just on whether they get the right answer, but on how they allocate extended reasoning budgets to solve problems that require sustained internal simulation. The benchmark's creator open-sourcing the repository and inviting others to test additional cube sizes (2x2 through 6x6) and other models signals the kind of grassroots, community-driven benchmarking that has become an important complement to official model evaluations, especially for niche but revealing tests of multimodal reasoning that major labs may not prioritize themselves.
Read original article →