### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

We ran a clean cost reconciliation of litellm against OpenAI for `openai/gpt-5.6-sol`, comparing the `StandardLoggingPayload` litellm emits (captured via the `s3_v2` callback into a mock S3) against what OpenAI actually billed for the same requests on platform.openai.com/usage

litellm under-counts cache-read tokens, and because those tokens fall back to the fresh-input / cache-write rates, spend is overstated by roughly 8.5%

| | litellm (StandardLoggingPayload) | OpenAI (billed) | Δ | | --- | --- | --- | --- | | requests | 40 | 40 | 0 | | input tokens | 73,550 | 73,550 | 0 | | output tokens | 20,480 | 20,480 | 0 | | cached tokens | 49,563 | 65,304 | -15,741 | | cost | $0.939487 | $0.865846 | +$0.073641 (+8.5%) |

Request count, input tokens, and output tokens reconcile exactly, so litellm is receiving usage from OpenAI on every call. The only field that diverges is the cache-read count: OpenAI billed 65,304 tokens at the cached-input rate, litellm recognized 49,563, about 24% fewer

Those 15,741 tokens are the entire cost gap. OpenAI priced t…