OpenAI follows Anthropic's anti-distillation measures: GPT-6.2's chain of thought is also locked down.
动察 Beating AI News Flash: OpenAI has begun locking down the hidden chain of thought for GPT-6.2 Astra and Sol to prevent other models from stealing the reasoning process via API to train their own models. This mechanism is highly similar to Anthropic's previously introduced anti-distillation measures.
At the end of October, The Information, citing people familiar with the matter, reported that DeepSeek and Moonshot AI extensively used outputs from GPT-6 Astra when training new models. Previously, domestic AI companies had extensively distilled Claude, but Anthropic continuously strengthened its blocking measures, prompting some vendors to shift their targets to OpenAI.
Distillation is the process of having one model learn from the outputs of a stronger model to acquire similar capabilities at lower cost. If the full chain of thought can be obtained, even how the model breaks down problems and arrives at answers step by step can be learned—far more valuable than only seeing the final answer.
Anthropic added a "context lock" to Claude's encrypted chain of thought in June this year. OpenAI now employs a similar mechanism: in multi-turn conversations, each segment of reasoning is bound to the historical messages preceding its generation. If a developer modifies old prompts, tool configurations, or conversation content, the API may discard reasoning items after the point of modification, preventing attackers from tampering with context to induce the model into leaking internal thinking.
OpenAI's self-test results show that in attacks attempting to extract the previous round's hidden reasoning, GPT-6 Astra had an average attack success rate as high as 54%, GPT-6.2 Astra dropped to 0.2%, and Sol stood at 0.5%.
Normal developers are also affected. Modifying old messages or compressing chat history may cause the Agent to lose previous reasoning results, increasing token consumption and response latency. OpenAI therefore recommends preserving the original conversation history as much as possible and using the officially provided context compression interface.