Are frontier LLMs now statistically indistinguishable on general benchmarks?

TRIATHLON-LLM tested frontier models and found a range of just 3 points (2.4%) across all top models. If the best models are within noise of each other on standard benchmarks, what should practitioners optimize for — cost, latency, context window, tool use? Is the model quality race effectively over for general tasks? Source: https://news.ycombinator.com/item?id=46336990