fix(core): price OpenAI prompt-cache input at the documented cached rate (#13104)

[santhiprakash]

Description

calculateOpenAICost() in core/llm/utils/calculateRequestCost.ts bills every prompt token at the standard input rate and ignores usage.promptTokensDetails.cached_tokens, so any OpenAI request that hits the prompt cache (the common case for stable, long system prompts) is over-reported by the cache-read discount. The Anthropic branch in the same file already handles cache tokens; the OpenAI branch does not.

This PR:

1. Adds cachedInput to the OpenAI pricing table for the model families where OpenAI documents a prompt-cache rate (gpt-4o, gpt-4o-mini).
2. Splits usage.promptTokens into uncached and cached portions and bills each at its own rate, mirroring the Anthropic branch.
3. Clamps cachedTokens to promptTokens to defend against malformed usage payloads that report more cached tokens than total prompt tokens.
4. Keeps the original behavior for models without a documented cachedInput rate: the cached portion is treated as uncached and billed at the standard input rate (the API would never return a cached_tokens value for them in the first place).

Tests

Added regression cases in core/llm/utils/calculateRequestCost.vitest.ts:

• gpt-4o with 700/1000 prompt tokens cached
• gpt-4o with all 1000 prompt tokens cached
• gpt-4o with explicit zero cached tokens (no cache rows in breakdown)
• gpt-4o-mini with 800/1000 prompt tokens cached
• gpt-4 with cached tokens reported but no cachedInput rate defined (full prompt billed at the standard input rate)
• gpt-4o with cached tokens > prompt tokens (clamped to promptTokens)

The fix was verified by extracting calculateOpenAICost into a standalone Node script and running 11 cases against hand-computed expected costs (all pass). I could not run the vitest suite in this environment (no node_modules at the workspace root); the upstream vitest suite should be re-run by the maintainer.

Checklist

• I've read the contributing guide
• The relevant tests have been updated

I have read the CLA Document and I hereby sign the CLA

🤖 Generated with Hermes Agent