The Growing Pains of Frontier Models: When Leaderboards Stop Separating and What to Measure Next

In three linesAnalysis of 34 frontier models (2024-2026) showing reasoning and coding capabilities cooperate (r=+0.72) but vary by lab. DeepSeek shifted from reasoning-rich to coding-first (+11.2→-4.7); Google maintains balance; Anthropic oscillates. SWE-bench saturating while HLE and instruction-following remain discriminative. Seven falsifiable predictions for next 12 months with interactive dashboard.

Summary generated by Claude — human-verified