Will an AI model reach a 3 hour time horizon with 80% reliability during 2026?

Prediction market on metaculus. [METR](https://metr.org/) is an AI model evaluation and threat research organisation that studies AI capabilities and their ability to conduct AI research and self-improve. METR regularly tests released models for their threat capacity, which includes measuring their ability to complete realistic software and research tasks reliably. For each task, they measure how long it takes humans with relevant expertise to complete it, then see which tasks different models can solve. By fitting a curve of model success against human task duration, they define each model's "[time horizons](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/)": the longest task length (in human time) on which the model can succeed with around 50% or 80% probability. Their results show that these horizons have been growing exponentially over time. As of December 2025 the highest time horizon with an 80% chance of success is 31 minutes by [GPT-5.1-Codex-Max](https://evaluations.metr.org/gpt-5-1-codex-max-report/), a 5x improvement from 2024's best performer GPT-o1.

Resolves: 1/1/2027.