Terminal Bench 4.0 has been released, with GLM-5.3 achieving performance comparable to Fable 5 within the margin of error. The benchmark's announcement emphasizes rapid iteration to keep pace with new model releases and combat benchmark saturation. The post also raises the question of more cost-effective alternatives for benchmarking coding agents, given that large benchmarks require 5–10B tokens.