GPT-6 Astra nails 72-hour agent tasks with 63% on-time rate, crushing Fable 5.1's 4%
maksym_andr · x · 2026-10-12
An analysis of 1,991 agent runs measuring how precisely models follow requested work durations.
- GPT-6 Astra: 63% on-time rate (runtime within 0.95–1.05x of the request), only 1.18x average deviation. Asked to work for 72 hours, it finished exactly at 72h.
- GPT-5.6 Sol: 39% on-time, 1.77x deviation.
- Fable 5.1: just 4% on-time, 2.86x deviation — it gave up after 3 hours on a 72-hour task.
The author notes precise duration-following only emerges on long tasks (>2 hours), and muses that early stopping to sync with the user might actually be the better behavior.
Related event: AgentTime Benchmark: GPT-6 Astra Hits 63% On-Time Rate in 72-Hour Tasks(2 posts)→
More from coding & agent
- 10 open-source projects extending AI from chatbots to docs, browsers and memory — Shruti_0810 · 2026-10-12
- LangChain founder lays out three models for enterprise agent identity — hwchase17 · 2026-10-12
- jax-graft: an AI-built JAX backend runs JAX on Apple Silicon GPUs — twiecki · 2026-10-12
- Ditch shadcn and Tailwind patterns to avoid the AI-slop web look — michalmalewicz · 2026-10-12
- Student struggles with agentic coding: Claude Code specs + Antigravity still miss details — 3ATAE · 2026-10-12
- All of science embedded and free: 200M papers searchable by AI agents, no API key — pbaylies · 2026-10-12