GPT-6.2 Astra, Sol API: Inference Discard on History Edits and Cache Cost Structure
·2026.11.18 14:00
Key point
The GPT-6.2 API discards inference when history is modified, and cache miss rewrite costs can differ by up to 25 times.
Details
- In the GPT-6.2 (Astra, Sol) API, modifying conversation history, downgrading the model, or changing the user causes inference items to be discarded, increasing token costs and latency.
- On a cache miss, the cost to rewrite the 30-minute cache can be up to 25 times the standard input price (based on Astra), while cache reads are 0.05 times cheaper.
- In system card evaluations, GPT-6.2 Astra showed an obstacle reporting omission rate of 22.28% and Sol 24.08%, which are higher than the previous generation GPT-6 Astra (11.09%).
- An Append-only method of adding a
configuration_updateitem at the end instead of modifying conversation history, and diagnosing cache miss causes viaprompt_cache_diagnostics, are recommended. - GPT-6.2 Astra supports a base context of 272,000 tokens, and Sol supports 500,000 tokens, with performance and costs varying depending on the inference intensity setting.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.