Implications of Large-Scale Test-Time Compute (5 minute read) (opens in new tab)

Benchmark grids hide the real story because LLM capability is now a function of test-time compute, illustrated by GPT-5.5 looking only marginally better than GPT-5.4 on max-compute cyber evals but substantially stronger once tokens, cost, or latency are controlled on the x-axis. The performance plateau is now empirically very far out and stronger models push the plateau further, so single-scalar benchmark scores will only get less informative each release.