probability: 26%

rationale:

Core Mechanism Is Inert The primary mechanism that would drive frontier models to diverge in their behavior is learning from different live deployment experiences. However, as of mid-2026, no frontier lab updates model weights from live sessions at scale. Current frontier systems rely on retrieval, context injection, and discrete post-training releases rather than continuous online learning. Without live, continuous weight updates reflecting idiosyncratic user interactions, the main theoretical driver for behavioral divergence is essentially inert during this window.

Evidence Strongly Points Toward Convergence The specific metric in question—chance-adjusted probabilistic agreement on model errors (CAPA)—has empirically been shown to increase (indicating greater error overlap) as model capabilities rise arxiv.orgarxiv.org. Independent research on correlated errors in large language models corroborates this, finding that even after controlling for provider, architecture, and size, more accurate models show higher error correlation arxiv.org. This suggests a strong structural trajectory toward convergence. While a separate epistemic-diversity study (arXiv 2510.04226) found that newer models generate more diverse claims, this measured within-model claim diversity across prompts, not cross-model error overlap arxiv.orgarxiv.org. It is therefore a different construct and weak evidence for a drop in CAPA.

Ongoing Convergence Pressures Through June 2027, frontier labs will continue to face powerful convergence pressures. These include training on shared web and synthetic data distributions, cross-lab distillation, common reinforcement learning recipes, and optimizing for the same public benchmarks. The Artificial Analysis Intelligence Index itself serves as a common multi-evaluation target that clusters leading organizations tightly, incentivizing developers to optimize toward the same high-performing capability regions.

Uncertainties and Compositional Churn The most plausible pathways to a lower CAPA score are noise-driven rather than structural. The identity of the top five organizations on the index is unstable, and compositional churn—such as swapping in models with entirely different lineages or Chinese open-weight entrants—could mechanically alter pairwise agreement. Furthermore, the baseline CAPA for the August 2026 top five has not been formally measured, creating significant measurement uncertainty. However, given the strong directional evidence for convergence and the absence of live weight updating, these factors are not enough to push the probability of divergence above a modest baseline.