A long conversation often mixes routine formatting with difficult analysis. For GPT-6 Astra, the interesting question is whether effort can change for the next response while the stable conversation prefix remains reusable. That is a narrower claim than promising that a higher setting always produces a better answer.
OpenAI documents a configuration_update input item for changing reasoning effort between responses. Keep the request-level reasoning.effort unchanged, and place the update before the next user message. The selected effort persists until another update overrides it. This feature is currently limited to Astra in standard, single-agent mode. Change reasoning mid-conversation
Suppose a conversation began at request-level effort low to summarize a synthetic migration plan. The follow-up asks for a harder dependency analysis. This illustrative HTTP Responses request body is untested; resp_previous represents the actual previous response ID:
{
"model": "gpt-6-astra",
"reasoning": { "effort": "low" },
"previous_response_id": "resp_previous",
"input": [
{
"type": "configuration_update",
"reasoning": { "effort": "high" }
},
{
"role": "user",
"content": "Find dependency cycles in the proposed migration and explain which steps must change."
}
]
}
The response’s reasoning.effort still reports the request-level value. Seeing low there is therefore not, by itself, evidence that the update failed. Preserve update items through previous_response_id, or retain their original positions when replaying history. The API rejects consecutive update items. Automatic compaction, automatic truncation, and standalone /responses/compact are incompatible with these histories; the guide describes explicit compaction_trigger and a fresh update after compaction. Configuration and history constraints
For evaluation, I would freeze a small set of synthetic migration graphs with known cycles. Compare a fixed-effort baseline with a policy that raises effort for the cycle-analysis follow-up and lowers it for subsequent formatting. Score detected cycles, false positives, unsupported claims, and whether the proposed ordering actually resolves each cycle. Measure latency and token usage alongside those task checks; a longer answer should not earn a higher score merely for its length.
Prefix preservation enables cache reuse under the usual caching conditions. It does not establish a cache hit or a cost saving on this workload. Keep the source graphs, instructions, and tool definitions stable across repeated trials, and distinguish intended effort, request-level metadata, usage, and measured correctness in the experiment record. Astra also does not support effort none; it should not be used as the routine-work baseline. Astra model guidance
CometAPI’s Astra model page identifies the model, and its Responses reference warns that supported request and response fields vary by model. These sources do not verify configuration_update, its persistence, or its usage reporting through that route. No provider experiment was run for this post. Adopt the adaptive policy only after compatibility tests and the graph evaluation pass; retain the fixed-effort baseline when the result is inconclusive.