Jack Clark, co-founder of Anthropic (AI safety company), stated in a May 4, 2026, newsletter that there is a greater than 60% likelihood that frontier AI models will autonomously engineer their own successors by the end of 2028. He noted that the progressive end-to-end automation of AI research could lead to a future that is exceptionally difficult to predict.

The forecast is based on significant growth in coding capabilities and long-duration task performance. The Claude Mythos Preview achieved a 93.9% success rate on SWE-bench, reaching a level where the benchmark is effectively saturated. Furthermore, in the METR time horizon (the maximum duration an AI agent can complete tasks with 50% reliability), performance extended to approximately 16 hours as of 2026.

Anthropic has integrated 'AI-driven R&D' as a priority area within the research agenda of The Anthropic Institute, which was established in March 2026. The company has also published research on 'Automated Alignment Researchers,' allowing AI to perform parts of alignment work autonomously, indicating an environment where AI can support the development of subsequent models.