Grok 4.7 Stateless Agents: Replay Encrypted Reasoning

xAI shipped Grok 4.7 on September 21, and the launch thread filled up fast with benchmark screenshots: 38.0% on Terminal-Bench 4.0 (up from 20.3% for Grok 4.6), 64.0% on EEBench, same $2 / $6 per million token price as 4.6. Those numbers are all xAI's own runs, so treat them as a claim rather than a verdict. The part that matters more if you actually build agents is buried in the developer docs: on the Responses API, Grok 4.7 now returns reasoning.encrypted_content on every response, whether you ask for it or not.

That one change decides whether a multi-step agent stays coherent. If you manage conversation history yourself (and most production agents do, because they need retries, audit logs and zero server-side storage), the model's thinking from step 3 is gone by step 4 unless you send the encrypted reasoning items back. Drop them and the agent re-derives its plan every turn: more tokens, worse answers, and a cache that keeps missing.