We estimate that Claude Opus 4.6 has a 50%-time-horizon of around 14.5 hours (95% CI of 6 hrs to 98 hrs) on software tasks. While this is the highest point estimate we’ve reported, this measurement is extremely noisy because our current task suite is nearly saturated.

Seems like a lot of people are taking this as gospel—when we say the measurement is extremely noisy, we really mean it. Concretely, if the task distribution we're using here was just a tiny bit different, we could've measured a time horizon of 8 hours, or 20 hours.

2026 ai bubble is peaking itself right now: someone just dropped a new benchmark on claude skills and the result of their paper is actually insane the paper explicitly states smaller models with skills beat larger models without them a smaller model like claude 4.5 haiku equipped with high quality skills smokes a raw state of the art opus 4.5 model by about 6 percent (27.7 vs 22.0) imagine getting sota level performance from a free model, its basically cheating, you just have to manually spoonfeed it a basic markdown file explaining how to do its job all of you opus guys are dumb, you can literally spam haiku with skills and get things shipped in 5x lesser time and 0 cost even wilder thing is that codex gpt 5.2 fails on the pareto frontier. codex burned massive compute and costs, just to get completely mogged by gemini 3 flash hitting maximum performance at a fraction of cash i can believe skill engineering is now a valid, mathematically proven substitute for compute over that, it says self generated skills provide zero benefit on average and show negative deltas on 16/84 tasks. if you give an agent more than 3 skills at once, it bloats its context and completely fails