0.5.2: cache tip quotes the session's own model (Flash is 50x, not 30x)

CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## 0.5.2 - 2026-09-18
+
+### Changed
+
+- **The cache tip quotes the session's own model.** "A stable prompt prefix
+  bills at 1/N of the miss rate" used the smallest discount on the whole
+  card, which is Pro's 30x. V4.1 Flash, the default model, discounts a hit
+  50x, so every Flash session was told the payoff was 40% smaller than it
+  is. `cacheDiscountOf(models)` takes the smallest discount among the models
+  that actually billed; the rate-table footnote still quotes the card-wide
+  floor, which is right there as an "at least".
+- README examples price `deepseek-flash`, the model most sessions now run.
+
 ## 0.5.1 - 2026-09-18
 
 ### Fixed

README.md

@@ -314,7 +314,7 @@ In JavaScript, skip the fetch and call the same module directly:
 import { costOf, tariffAt } from '@dshworks/dsh-meter/core'
 
 const tokens = { miss: 188_542, hit: 1_204_880, out: 9_310 }
-costOf(tokens, 'deepseek-v4-pro', tariffAt(Date.now()), 'cny')
+costOf(tokens, 'deepseek-flash', tariffAt(Date.now()), 'cny')
 ```
 
 `lib/core.js` imports nothing. Rate card, tariff clock, and cost fold are
@@ -333,7 +333,7 @@ running:
 curl -s https://dsh.works/dsh-meter/pricing.json | jq -r '
   ((now + 8*3600) | gmtime) as $b
   | .timeOfUse.scheduleBeijing[$b[6]][$b[3]] as $t
-  | "\($t) · v4-pro out $\(.models["deepseek-v4-pro"].rates[$t].usd.out)/1M"'
+  | "\($t) · flash out $\(.models["deepseek-flash"].rates[$t].usd.out)/1M"'
 ```
 
 Shift by `8*3600` first, then break down **once**, and take both indices

README.zh.md

@@ -175,7 +175,7 @@ JavaScript 里可以跳过这次 fetch，直接调同一个模块：
 import { costOf, tariffAt } from '@dshworks/dsh-meter/core'
 
 const tokens = { miss: 188_542, hit: 1_204_880, out: 9_310 }
-costOf(tokens, 'deepseek-v4-pro', tariffAt(Date.now()), 'cny')
+costOf(tokens, 'deepseek-flash', tariffAt(Date.now()), 'cny')
 ```
 
 `lib/core.js` 不依赖任何包。价目表、档位时钟、成本折叠都是纯函数，内部不读时钟，所以历史会话按它当时真实的档位回算。

lib/core.js

@@ -188,6 +188,21 @@ export const CACHE_DISCOUNT = Math.floor(
       Object.values(byTariff[tariff]).map((rate) => rate.miss / rate.hit)))),
 )
 
+/**
+ * The same ratio for the models one session actually billed, so the cache
+ * tip speaks about the model in front of the user. On the live card Flash
+ * discounts a hit 50x and Pro 30x; quoting the card-wide 30 to a Flash
+ * session understated the payoff by 40%.
+ * @param {string[]} models - billed model ids.
+ * @returns {number} the smallest live discount among them, or CACHE_DISCOUNT when none is on the live card.
+ */
+export function cacheDiscountOf(models) {
+  const ratios = models.filter(model => RATES[model] !== undefined).flatMap(model =>
+    TARIFFS.filter(tariff => tariff !== 'flat' && RATES[model][tariff] !== undefined).flatMap(tariff =>
+      Object.values(RATES[model][tariff]).map(rate => rate.miss / rate.hit)))
+  return ratios.length === 0 ? CACHE_DISCOUNT : Math.floor(Math.min(...ratios))
+}
+
 /** Currency symbols for the two published tables. */
 export const CURRENCY_SYMBOL = { usd: '$', cny: '¥' }
 

package.json

@@ -1,7 +1,7 @@
 {
   "name": "@dshworks/dsh-meter",
   "description": "The DeepSeek time-of-use meter for dsh: what this session cost, which tariff is running, when it flips, and the account balance behind it — one line under the composer.",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "license": "MIT",
   "homepage": "https://dsh.works/dsh-meter/",
   "publishConfig": {

src/ui.js

@@ -495,7 +495,7 @@ function MeterCard({ value, balance, currency, now, t, position }) {
   if (value.tokens.hit > 0) {
     notes.push(t('meter.cacheSaved', { amount: formatMoney(saved, currency), percent: hitShare }))
   } else if (promptTokens > 0) {
-    notes.push(t('meter.cacheCold', { factor: CACHE_DISCOUNT }))
+    notes.push(t('meter.cacheCold', { factor: cacheDiscountOf(value.models.map(entry => entry.model)) }))
   }
   if (clock.tariff === 'flat') {
     notes.push(t('meter.ifNewRates', {

tests/tariff.spec.mjs

@@ -1,7 +1,7 @@
 import { describe, expect, it } from 'vitest'
 import {
   CACHE_DISCOUNT, CURRENCY_SYMBOL, PEAK_WINDOWS_UTC, RATES, TIME_OF_USE_FROM,
-  WEEKEND_OFFPEAK_FROM, bucketCostOf, costOf, formatCountdown, formatMoney,
+  WEEKEND_OFFPEAK_FROM, bucketCostOf, cacheDiscountOf, costOf, formatCountdown, formatMoney,
   formatTokens, isBeijingWeekend, nextTariffChange, tariffAt, tariffSchedule,
 } from '../lib/core.js'
 
@@ -263,3 +263,22 @@ describe('display helpers', () => {
     expect(formatCountdown(-1)).toBe('0s')
   })
 })
+
+describe('the cache tip speaks about the session\'s own model', () => {
+  const cardWide = CACHE_DISCOUNT
+
+  it('quotes Flash its own discount, not the card-wide minimum', () => {
+    // ¥1 miss against ¥0.02 hit, $0.15 against $0.003: 50x in both tables.
+    expect(cacheDiscountOf(['deepseek-flash'])).toBe(50)
+    expect(cardWide).toBe(30)
+  })
+
+  it('takes the smaller discount when a session mixed models', () => {
+    expect(cacheDiscountOf(['deepseek-flash', 'deepseek-v4-pro'])).toBe(30)
+  })
+
+  it('falls back to the card-wide figure for a model the live card does not price', () => {
+    expect(cacheDiscountOf(['mystery-model'])).toBe(cardWide)
+    expect(cacheDiscountOf([])).toBe(cardWide)
+  })
+})