[Bug]: Long-context tier starts one token early — getPricingTier is inclusive, Google and OpenAI document their thresholds as exclusive #5788

What happened?

getPricingTier treats every long-context threshold as inclusive, but two of the three vendors that have thresholds in the registry document theirs as exclusive. A request whose prompt lands exactly on the threshold is billed at the long-context tier when the vendor bills it at the base tier — 2x input and 1.5x output on that request.

packages/cost/models/calculate-cost.ts:105

// Find the highest threshold that the value meets
for (let i = 0; i < preprocessedPricing.length; i++) {
  if (value >= preprocessedPricing[i].threshold) {

The comparison is >= for all vendors. What the vendors actually say:

Sources, read 2026-08-22: https://ai.google.dev/gemini-api/docs/pricing (Gemini 2.5 Pro, Standard, paid tier) and https://developers.openai.com/api/docs/models/gpt-5.4.

xAI is in the table on purpose: it is the reason this is not a one-character fix. Flipping >= to > globally would make xAI wrong in the opposite direction. The convention genuinely differs per vendor, so the tier needs to carry it.

Which rows are affected

From your own registry snapshot (packages/__tests__/cost/__snapshots__/registrySnapshots.test.ts.snap), the non-zero thresholds are:

• threshold: 200000 — google/gemini-2.5-pro, google/gemini-3, google/gemini-3-pro-image, google/gemini-3.1-pro-preview (documented exclusive → currently overcharged at exactly 200,000)
• threshold: 272000 — openai/gpt-5.4 (documented exclusive → currently overcharged at exactly 272,000)
• threshold: 200000 — anthropic/claude-4.5-sonnet, anthropic/claude-sonnet-4, anthropic/claude-sonnet-4-5-20250929. I am not claiming a direction for these: Anthropic's current pricing page documents 4.6-and-later as standard pricing across the full 1M window and no longer states the ≤/> wording for the 4.5-era tier, so I could not verify the boundary from a live source. Worth checking separately.

Repro

A google/gemini-2.5-pro request with exactly 200,000 prompt tokens and no cache:

• expected (Google's page): 200000 * 0.00000125 = $0.25
• actual (getPricingTier picks the 200000 tier): 200000 * 0.0000025 = $0.50

One token less (199,999) gives $0.2499…, so the jump lands one token early.

Suggested fix

Let the tier carry the convention instead of hard-coding one, e.g.

{ threshold: 200000, thresholdExclusive: true, input: 0.0000025, output: 0.000015 }

const t = preprocessedPricing[i];
if (t.thresholdExclusive ? value > t.threshold : value >= t.threshold) {

Default false keeps today's behaviour for every existing row; only the Google and OpenAI tiers above need the flag. The alternative — redefining threshold as "first token count billed at this tier" and storing 200001 / 272001 — is a data-only change with no code change, but it makes the field mean something different from its comment, so I'd lean to the flag.

Happy to send the PR for either shape if you tell me which you prefer.

Note

Same function as #5766 (cache-write tier selection), different line — that one is about which tier cache writes use, this one is about where the tier starts. They don't overlap.

I keep a table of which vendors are inclusive vs exclusive at the threshold, each with the quote from the vendor's own page and the date it was read, here: https://xyzs996.github.io/llm-api-pricing/prices.html#same-number-opposite-answer. It is where the wording above comes from.