Day two of Claude Opus 5, and the debutant is still collecting — two more crowns today, both lifted from its own stablemate, Claude Fable 5
The Aggregate Digest — Sunday, July 26, 2026
Day two of Claude Opus 5, and the debutant is still collecting — two more crowns today, both lifted from its own stablemate, Claude Fable 5. On Multi-turn Debate (Lechmazur), Opus 5 (High) now rates 1763.4 to Fable 5 (High)'s 1756.8, a seven-point Bradley-Terry edge on a 42-model board. On EQ-Bench Longform Writing it posted 86.3 against Fable's 83.0. Add those to Saturday's thirty lead changes and the new model has taken 32 boards in its first two days — while its overall ELO climbed from 2882 at debut to 2963, still one notch below Fable 5 on the main leaderboard.
The rest of Opus 5's day is a study in honest margins. Its Benchmarks.bio rows hold #1 on VariantBench (49.72%) and BioSecBench-Surveillance (59.8%), but EpiBench shows the ceiling: fourth at 35.22%, nearly ten points behind GPT-5.5's 44.97%. Design Arena tells the same story — #2 on UI Components (1387 Elo) and #4 on Website (1331), both looking up at Kimi K3 (1408 and 1375).
One benchmark joined the corpus: SpatialGen-Bench (Text Answering), a spatial-cognition suite from the ProVisE project — perception, understanding, reasoning and interaction tasks across 14 spatial task types, answered in text. It opens as one of the tightest two-horse races in the corpus: Claude Fable 5 at 81.26%, GPT-5.6 Sol at 80.98% — 0.28 points apart — and then an eighteen-point cliff down to Kimi K3 at 63.36% and the rest of the 25-model field.
Elsewhere. ByteDance's Seed 2.1 Pro took NYT Connections Older Models outright at 97.8%, more than eight points clear of GLM-5.2 (max reasoning) at 89.5, on a puzzle board where the lead had sat in the low nineties. And on FoodTruckBench, GPT-5.6 Sol's $53,229 net worth is good for second, still well short of GPT-5.5's $61,408.
