Unverified50% confidenceDecision RuleExact time
The API performs two checks on each thinking or redacted_thinking block: the model must be able to read the block (otherwise it is silently discarded), and the content preceding the thinking block (the top-level system prompt, the tools array, and every message before that block) must have no changes whatsoever.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedDuring inference, a model may undergo server-side A/B routing, where the same API endpoint may invoke different weight versions or system prompt versions at different times, without users being able to perceive this.68% similarUnverifiedThe Responses API supports orchestrating multi-step reasoning and tool calls within a single API call, eliminating the need for the client to manually manage intermediate state.67% similarUnverifiedIn API timeout scenarios, non-interactive sessions and sub-agents now continue from partial responses, while responses containing only thinking content will be retried.66% similarUnverifiedThe API logs in the demo precisely recorded two tool calls—one read and one write—providing observability.66% similarUnverifiedDuring multi-turn tool calls, each tool execution result is appended to the conversation history as a new tool_result message, and the model reads the outputs of previous steps from the conversation context rather than retaining intermediate values in internal memory.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1022208API
curl https://kongchang.com/api/v1/knowledge/claims/1022208MCP
get_claim(id=1022208)