InferenceBench trend: frontier models' inference speedups grew from 3x to 12x in a year
maksym_andr · x · 2026-11-09
InferenceBench's historical trend from Sep 2025 to Oct 2026 shows frontier models' inference-optimization speedup factors growing from 3x to 12x, with a doubling time of 6.3 months (×1.117/month, R²=0.95). The leaderboard has Claude Opus 5.5 (Max) at 12.08x geometric-mean speedup, followed by Claude Fable 5.1 (9.83x) and Claude Opus 5 (8.90x); GPT-6 Astra (Ultra) hits 7.90x and GLM-5.2 (Max) 7.00x — all far above default vLLM/SGLang configs.
More from Infra
- US local elections on data center curbs could raise AI compute costs — pstAsiatech · 2026-11-09
- Local AI rig runs DeepSeek 4.1 + video cloning pipeline on RTX Pro 6000s — dee_hw · 2026-11-09
- Oracle's AI alliances with OpenAI, Nvidia and Google span the stack from chips to chatbots — thealexbanks · 2026-11-09
- Akamai's AI bet: $11B of $14B dealbook tied to Anthropic, a 2028 story at 2.5x sales — pdamodaran · 2026-11-09
- Apollo's planned $75B AI chip buying would rival hyperscalers, eclipse all neoclouds — pstAsiatech · 2026-11-09
- Anthropic Pays $3.4B for 1.7GW of Gas Generators, Largest AI Lab Power Buy — Exponential View (Azeem Azhar) · 2026-11-09