Terminal Bench 4.0 just dropped, GLM-5.3 is at the same level as Fable 5, accounting for margin of error

Imo the best aspect in their announcement is their focus on rapidly iterating on TerminalBench to keep the pace up with new model releases to fight benchmark saturation.

On a similar note, what cheaper/smaller alternatives are there to benchmarking coding agents or your own harness? Large benchmarks like this take 5-10B tokens, which is not economically/computationally feasible for the vast majority of us.

I'd love to objectively measure how my skills/harness/tools/etc change token usage and success probability on general coding tasks, there has to be a way to do this to at least give an idea or general direction, without requiring billions of tokens for each run.