Inference Optimization and Cost Trends

Claude Opus 5 launched at half Fable price (ECI 159 vs 161). Kimi K3 serving cost reality challenges 'open = cheap'. New techniques: OmniDelta token compression, Multi-Head Latent Control (90.7% reduction), JetSpec 9.64x speedup, KronQ 2-bit quantization. Etched hits $10.3B valuation for custom inference chips. AMD and Cerebras split inference. Inference chip financing trend continues.