Anthropic has Claude grade its own conversations for about 10 tokens
After each reply, Anthropic adds one hidden turn and Claude returns a JSON safety flag in ~10 tokens, under 2% of compute. The KV cache makes it cheap.

Anthropic’s Safeguards team published how Claude monitors its own conversations for misuse. After each reply, the system appends a short hidden instruction asking Claude to reflect on the safety of the conversation, and Claude answers with a small JSON object. Most of the time that’s just {"flagged": false}. When it flags something, it adds a reason, a category, a subpolicy, whether it refused, and a 1 to 10 priority for human review.
Here is how they describe the cost:
When Claude finishes a response, the entire conversation is already in the KV cache. Each assessment therefore only requires processing a short instruction and generating ~10 additional tokens on average. Over our production API traffic distribution, we estimate that self-assessment adds less than 2% to our inference compute.
The alternative they compare against, a separate prompted classifier with the policies in its prompt, has to read the whole conversation again and reason about it, which they say takes many times more compute per conversation. A footnote says Anthropic doesn’t bill the self-assessment cost to customers. On Claude Fable 5, at a 1% false-positive rate, self-assessment flags 78% of the conversations that stronger classifier flags, and 94% of the high-priority ones.
They’re also clear about where it’s weak. Used as a blocking filter against an automated attacker on 24 offensive cyber tasks, it cut attack success from 79% to 54%. The prompted classifier got it to 25%. So it runs offline, as a wide first pass that feeds higher-precision review. And they cite a study showing models rate an action as less risky when it appears in their own earlier turn, which is the obvious worry with any model grading itself.
What I take from it as someone building agents: the cheapest check you can run on an agent’s transcript is one more turn on the same cached context, with a fixed JSON schema and a very short answer. On Anthropic’s API you pay the cache-read price for the history (0.1x the base input price on most models as of today, lower on the newest ones) plus a few output tokens. That’s a very different bill from the agent_summary side query I wrote about yesterday, which kept resending whole transcripts for a status line. Same machinery, and whether it’s cheap depends on whether the cache actually hits. I covered that in the cost per task post.
I’d use this pattern for flags that a human reviews later, never as the only gate in front of an action that can’t be undone. Anthropic’s own blocking numbers are the reason.
Via X, where it was trending on October 30. The post itself is dated October 16, 2026.