fix(cost): tier Anthropic cached-input reads by the long-context threshold

[eeshsaxena]

Problem

getThresholdValueFunction returns, per provider, the token count used to pick which pricing tier applies to a given cost field. The anthropic branch handles inputCost and outputCost but has no cachedInputCost case, so that field falls through to default: return 0. Every other provider that tiers (vertex, xai, google-ai-studio) has a cachedInputCost case.

Because the threshold value is always 0 for Anthropic cached reads, getPricingTier always returns the base tier for them, even on a request past the long-context threshold.

Concretely, claude-4.5-sonnet:anthropic prices:

threshold 0:      input 0.000003, cacheMultipliers.cachedInput 0.1
threshold 200000: input 0.000006  (cacheMultipliers inherited)

A cached read is priced as cachedInput * tier.input * cacheMultipliers.cachedInput. Above 200k prompt tokens the input rate doubles, so the cached-read price should double too (0.1 * 0.000006). Instead it stays at the base tier (0.1 * 0.000003), so cached reads on long-context Anthropic requests are billed at half the intended rate. Input and output already tier correctly, so within one request the cached-read rate is inconsistent with the input rate it is derived from.

Fix

Add the cachedInputCost case to the anthropic threshold function, returning the same total-prompt value as inputCost/outputCost. Cached reads now select the same tier as input and output.

Tests

packages/__tests__/cost/anthropic-cache-tier.test.ts checks that a >200k request prices cached reads at the tier-1 rate (0.06 for 100k cached tokens, where the base tier would give 0.03), and that a sub-threshold request still uses the base tier.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[vercel]

@eeshsaxena is attempting to deploy a commit to the Helicone Team on Vercel.

A member of the Team first needs to authorize it.