Grok 4.7 delayed at the last minute: Musk says training over-penalized response length

Beating AI News Flash: Musk confirms Grok 4.7 will take a few more days. On September 2, he said "release in 10 days," originally pointing to around September 12. Now, as the launch approaches, xAI has found that the model wraps up prematurely when encountering high-difficulty tasks, abandoning even tasks it could have completed, and its answer checking is not strict enough.

Musk suspects the problem lies in the reinforcement learning stage. During training, the penalty for response length may have been too heavy, pushing the model toward shorter, faster-ending answers, and the reasoning needed for complex tasks was compressed along with it. However, he himself added "possibly" and "or similar reasons," indicating that xAI has not yet fully confirmed the root cause.