Anthropic Claude Opus 4.5 hits 4h49m METR long-horizon mark

METR’s new result for Claude Opus 4.5 is one of the clearest quantitative signals yet that long‑horizon agent performance is climbing fast. A 50% time horizon of 4 hours 49 minutes means that, on METR’s coding tasks, Opus can often stay coherent and productive over nearly five hours of human‑estimated work—more than double earlier Claude variants and ahead of prior leaders on the same benchmark. Commentators analyzing METR’s updated curves argue that effective horizons may now be doubling every few months rather than every half‑year, implying “workday‑length” autonomy sometime in 2026 if current trends hold. ([the-decoder.com](https://the-decoder.com/anthropics-claude-opus-4-5-can-tackle-some-tasks-lasting-nearly-five-hours/))

For the race to AGI, this is less about one model “being AGI” and more about the shape of the curve. If agents can reliably tackle multi‑hour projects with only light supervision, you start to see real substitution in software engineering, data work, and ops—and a lot more experimentation with stacked, self‑verifying agent systems. The catch is reliability: at an 80% success threshold, Opus 4.5’s horizon collapses to ~27 minutes, roughly on par with some OpenAI models, underscoring how brittle very long runs remain. That gap between “can sometimes do it” and “almost always does it” is exactly where safety, verification, and orchestration research will determine whether these systems become dependable teammates or just spectacular but flaky tools. ([ai-primer.com](https://ai-primer.com/en/engineer/reports/2025-12-20?utm_source=openai))