Executive summary
Solving a hard problem once is different from carrying a long undertaking through new evidence, interruptions and the consequences of earlier actions. Dawn of Durable Intelligence argues that capability takes intelligence and durability together, and measures what durability adds when the base model stays the same. Across 6,000 runs on GPT-5.5, Opus 4.6 and Gemini 3 Pro in five professional workloads, day-30 task fidelity climbs to 95.2%, against 63.9% for LangGraph Deep Agents, the strongest alternative tested. Better models make stronger contributions; durability keeps those contributions useful long enough to build on.
Findings
- Runs with zero violations rise from 18.4% to 97.6%, and outputs containing unsupported facts or reasoning fall from 8.2% to 0.7%.
- Across 384 work programmes capped at 42 days, agents run a median 109.6 hours before a person has to rebuild their work, up from 7.4. Accepted work rises from 16.2% to 24.3%, and the median human time each programme needs drops from 41.6 hours to 13.7.
- The same programmes use 82% fewer tokens, 2.5 billion per 42 days against 13.9 billion.
- Kept information can still support a wrong decision. A system that retained 94% of its information but left dependent decisions unrevised made 22 violations per thousand decisions, against 6 for the full system.
- The same continuity lets a system accumulate experience and adopt stronger models mid-project, carrying tested methods and their evidence into each generation.
Nenya brings these results to the agents you use, with work that holds up for thirty days and runs nearly fifteen times longer before you step in.