Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesIf an LLM evaluation returns an old output after its input changes, first check whether your evaluation harness or application returned a cached completed result. That is not the same as provider-side prompt caching: OpenAI describes prompt caching as reusing intermediate key-value computation for a matching rendered input prefix, not replaying a finished evaluation result. Without logs or code from your system, the exact cause cannot be identified from the symptom alone.
Identify what was reused
There are at least two distinct cache layers to investigate. A provider prompt cache can reuse computation for a compatible request prefix; an application or evaluation harness can instead return a previously completed model or grader result. OpenAI’s prompt-caching documentation describes the former. It does not establish that prompt caching caused a stale completed evaluation result.
As an Amazon Associate I earn from qualifying purchases.
| Cache layer | What is reused | What to inspect |
|---|---|---|
| Provider prompt-prefix cache | Intermediate key-value (KV) computation for a matching rendered prefix, rather than the completed output itself. | Rendered prefix, compatible request settings, and provider-side cache diagnostics or usage where available. |
| Application or evaluation-result cache | A stored completed model or grader result, if the implementation uses such a cache. | Application or harness cache key, cache-hit records, evaluation/run identifiers, and the input and configuration recorded with the stored result. |
The table describes the conceptual distinction; whether your particular harness or application has a result cache, and how it keys or invalidates entries, must be confirmed in that system.
Recommended Free Tools
Trace a stale completed evaluation result
Start with one evaluation item whose input changed. Capture the old and new input, the expected and actual outputs, evaluation or run identifiers, and timestamps. Then determine whether the old output was returned before a model or grader call. If so, investigate the harness or application cache path; if a provider request was made, inspect that request and its cache behavior as a separate question.
#1 Best Overall
Inspect the result-cache key and provenance
Log the exact key used for lookup and the values from which it was constructed. Check whether it includes the changed input or a stable digest of it, and whether it accounts for other output-affecting changes: prompt or template revision, model and configuration, dataset or example, grader, and tool or retrieval versions. These are general engineering checks, not a universal cache-key schema prescribed by OpenAI’s Create eval API reference.
- Look for input fields omitted from the key, stale normalization, and accidental reuse across dataset rows.
- Check whether prompt or configuration version fields are missing or unchanged when the underlying content changed.
- Verify that mutable references or serialization behavior are not causing distinct inputs to collapse to the same key.
- Compare the provenance stored with the cached result against the current evaluation item and configuration.
A cache hit is useful only when the key represents the inputs and configuration that determine the result. Preserve enough provenance alongside each stored result to establish what produced it; do not infer the key design of a particular product from the symptom alone.
Check provider prompt caching separately
OpenAI prompt caching depends on a matching rendered prefix and compatible settings. Compare the complete rendered request prefix—not just the user-visible question—and the settings that can affect compatibility. OpenAI lists model, tools and tool ordering, output format or schema, reasoning effort, verbosity, and context management among relevant factors in its prompt-caching documentation.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →OpenAI’s prompt-cache diagnostics page describes a tool that compares a current request with an earlier response to investigate why an expected prefix was not reused. Its diagnostic reasons include input_changed, tools_changed, text_format_changed, reasoning_effort_changed, verbosity_changed, and context_compacted. For input_changed, the page notes that changed or reordered earlier input—including timestamps or request IDs in instructions—can affect reuse. It advises keeping dynamic content after reusable prefix content and its breakpoint.
These diagnostics concern expected prompt-prefix reuse. A matching prompt prefix explains reuse of computation; it does not, by itself, explain why an application might return an old completed evaluation output.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Use diagnostics and retention details in context
Use provider cache diagnostics and usage information, where supported, to examine prompt-prefix reuse. Use application or harness logs and cache-hit records to trace completed-result reuse. The two kinds of evidence answer different questions, so a provider cache diagnostic alone cannot establish that a result cache returned an old evaluation.
OpenAI describes prompt-cache retention as model-dependent. Its current documentation lists 30m as the supported minimum-lifetime setting or default for GPT-5.6 and later, and describes retention options for earlier models. These are provider prompt-cache retention details, not time-to-live guidance for an evaluation-result cache; confirm the documented behavior for the model you use.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
