fix(cost): Gemini 2.5 cache reads are 10% of input, not 25% (2.5x overcharge)

[xyzs996]

Ticket

None — found while diffing this registry against Google's published rates.

Component/Service

• Packages

Type of Change

• Bug fix

Deployment Notes

• No special deployment steps required

Context

All three Gemini 2.5 models price a cache read at 25% of input. Google publishes 10% for
every one of them, so cached input is billed at 2.5x.

25% is not an arbitrary number — it is the ratio for Gemini 2.0 Flash ($0.10/M in,
$0.025/M cached). 2.0 Flash was shut down on 2026-06-01. The 2.5 rows are still carrying its
multiplier.

Google's current rates (ai.google.dev/gemini-api/docs/pricing):

model	input	context-caching	ratio	this repo
gemini-2.5-pro	$1.25/M (≤200k)	$0.125/M	10%	0.25
gemini-2.5-pro	$2.50/M (>200k)	$0.25/M	10%	—
gemini-2.5-flash	$0.30/M	$0.03/M	10%	0.25
gemini-2.5-flash-lite	$0.10/M	$0.01/M	10%	0.25
gemini-2.5-flash-lite (audio)	$0.30/M	$0.03/M	10%	0.0833333
gemini-2.0-flash (deprecated)	$0.10/M	$0.025/M	25%	—

The repo already disagrees with itself in two places. All four Gemini 3.x models use 0.1:

gemini-3/endpoints.ts:16                     cachedInput: 0.1, // $0.2/1M = 10% of input
gemini-3-flash-preview/endpoints.ts:16       cachedInput: 0.1, // $0.05/1M = 10% of input
gemini-3.1-pro-preview/endpoints.ts:17       cachedInput: 0.1, // $0.20/1M = 10% of input
gemini-3.1-flash-lite-preview/endpoints.ts:16 cachedInput: 0.1, // $0.025/1M = 10% of input

And inside gemini-2.5-flash/endpoints.ts the audio line one block above already reads 10% while
the cacheMultipliers block below it reads 25%:

audio: {
  input: 0.000001,
  cachedInputMultiplier: 0.1,   // $0.1/1M cached audio (10% of input)   <- 10%
},
cacheMultipliers: {
  cachedInput: 0.25,                                                     // <- 25%

What this costs a user

Cache reads dominate agent traffic, so the error lands almost entirely on the token type that
carries the bill. Per 1M cache-read tokens:

model	billed today	should be
gemini-2.5-pro	$0.3125	$0.1250
gemini-2.5-flash	$0.0750	$0.0300
gemini-2.5-flash-lite	$0.0250	$0.0100

Blended at a coding agent's measured token mix (95.6% cache read / 4.1% cache miss / 0.3%
output, over 8.04B tokens), the reported cost per 1M tokens is 90% / 89% / 97% too high for
Pro / Flash / Flash-Lite respectively — the bill roughly doubles.

Changes

Six cachedInput values and two cachedInputMultiplier values, google-ai-studio and vertex
blocks in each file:

gemini-2.5-pro/endpoints.ts:16,63          cachedInput 0.25 -> 0.1
gemini-2.5-flash/endpoints.ts:20,66        cachedInput 0.25 -> 0.1
gemini-2.5-flash-lite/endpoints.ts:20,66   cachedInput 0.25 -> 0.1
gemini-2.5-flash-lite/endpoints.ts:16,62   cachedInputMultiplier 0.0833333 -> 0.1

Snapshot regenerated with npx jest __tests__/cost/registrySnapshots.test.ts -u.
npx jest __tests__/cost passes: 13 suites, 190 tests, 8 snapshots.

Extra Notes

cachedInput: 0.25 elsewhere is correct and is not touched. It appears 29 more times under
openai/gpt-4.1, openai/o3, openai/o4, openai/gpt-5.1 and xai/ — OpenAI and xAI really do
discount cached input to 25%. That is exactly why a stale 25% on a Gemini row does not look wrong
on inspection.

What I verified and what I did not. I verified the four rates above against the Gemini API
pricing page. I could not get the Vertex AI pricing page to render its tables, so the three
vertex lines rest on this repo's own convention — every Gemini 3.x model here carries identical
cachedInput on the google-ai-studio and vertex blocks. If Vertex actually differs, the
vertex half of this diff should be dropped and the studio half kept.

One thing I left alone. gemini-2.5-pro's threshold: 200000 tier has input: 0.0000025 but
no cacheMultipliers block at all, so there is no cached-input rate for prompts over 200k. Google
charges $0.25/M there (10% of $2.50). Adding the block would change behaviour for that tier rather
than correct a wrong number, so I did not fold it into this PR — flagging it in case you want it.

[vercel]

@xyzs996 is attempting to deploy a commit to the Helicone Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

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