UniWorld-View walked onto the WorldScore board and swept it
The Aggregate Digest — Friday, July 24, 2026
UniWorld-View walked onto the WorldScore board and swept it. The newcomer took the overall WorldScore crown at 80.81 — edging WorldScape-0.2(MoE)'s 80.68 by a hair — and carried six of its sub-boards with it: Camera Control (97.72), Photometric Consistency (94.11), Object Control (88.98), Content Alignment (86.61), Static (85.53), and Motion Accuracy (77.66). Most of those crowns changed hands by fractions, but the breadth is the story: one model, one day, nearly the whole board.
The frontier labs spent the day trading blows on the hard exams. Claude Fable 5 posted the loudest single number, taking the EnigmaEval (SEAL) crown at 39.28 — well above the 23.82 that led it yesterday — though it should enjoy the lead while it lasts: GPT-5.6 Sol sits at 37.12 on a board only five models deep. Fable 5 also opened the brand-new SealedBench at 71.0, a private, continuously rotated frontier-capability eval whose questions never sit still. There too GPT-5.6 Sol shadows it, tied for second at 68.0 with Kimi K3.
GPT-5.6 Sol had crowns of its own. It ran away with Vellum's ARC-AGI-2 at 92.5, seven and a half points clear of GPT-5.5, and took Vellum HumanEval at 96.2. It placed third on Vellum GPQA at 94.6. The honest counterpoint: on Kaggle FACTS Grounding it managed only 16th of 45 — grounding is still a step behind its raw reasoning.
Grok had a quieter, pointed day. Grok 4.5 nudged Claude Fable 5 off the TrackingAI IQ crown, 98.04 to 97.30, and Grok 4.20 opened the day's other new board — REFUTE, an epistemic-honesty test scoring models on accuracy and calibration — at 74.5. The top there is a photo finish: Claude Opus 4.7 and 4.6 trail by barely a point, at 73.4 and 73.2.
Elsewhere: Kimi K3 kept frontier company on the reasoning and agentic boards, placing third on Agents' Last Exam and fourth on Humanity's Last Exam. No new models entered the catalog today.
