feat(cost): support exact custom model pricing (#2131)

* feat(cost): support exact custom model pricing * fix(cost): address custom pricing review feedback
2026-08-24 02:34:15 -05:00 · 2026-08-16 15:55:02 +08:00
parent c30578819e
commit 09eba26d30
19 changed files with 1582 additions and 68 deletions
@@ -0,0 +1,37 @@
+import { test } from 'bun:test'
+import { spawnSync } from 'node:child_process'
+import { dirname, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+const FIXTURE_TIMEOUT_MS = 60_000
+
+test(
+  'SDK result and maxBudgetUsd use positive and zero custom prices',
+  () => {
+    const fixture = resolve(
+      repoRoot,
+      'src/test/fixtures/queryEngineCustomPricingBudget.fixture.ts',
+    )
+    const result = spawnSync(process.execPath, [fixture], {
+      cwd: repoRoot,
+      encoding: 'utf8',
+      timeout: FIXTURE_TIMEOUT_MS,
+      env: { ...process.env, FORCE_COLOR: '0' },
+    })
+
+    if (result.error) throw result.error
+    if (result.status !== 0) {
+      throw new Error(
+        [
+          `Fixture exited with status ${result.status ?? 'unknown'}.`,
+          result.stdout.trim(),
+          result.stderr.trim(),
+        ]
+          .filter(Boolean)
+          .join('\n\n'),
+      )
+    }
+  },
+  { timeout: FIXTURE_TIMEOUT_MS + 5_000 },
+)
@@ -0,0 +1,36 @@
+import { test } from 'bun:test'
+import { spawnSync } from 'node:child_process'
+import { dirname, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), '../../..')
+const FIXTURE_TIMEOUT_MS = 60_000
+
+test(
+  '/model and /fast price strings use the exact custom price',
+  () => {
+    const fixture = resolve(
+      repoRoot,
+      'src/test/fixtures/customPricingDisplay.fixture.tsx',
+    )
+    const result = spawnSync(process.execPath, [fixture], {
+      cwd: repoRoot,
+      encoding: 'utf8',
+      timeout: FIXTURE_TIMEOUT_MS,
+      env: { ...process.env, FORCE_COLOR: '0' },
+    })
+    if (result.error) throw result.error
+    if (result.status !== 0) {
+      throw new Error(
+        [
+          `Fixture exited with status ${result.status ?? 'unknown'}.`,
+          result.stdout.trim(),
+          result.stderr.trim(),
+        ]
+          .filter(Boolean)
+          .join('\n\n'),
+      )
+    }
+  },
+  { timeout: FIXTURE_TIMEOUT_MS + 5_000 },
+)
@@ -4,6 +4,7 @@ import { useState } from 'react';
 import type { CommandResultDisplay, LocalJSXCommandContext } from '../../commands.js';
 import { Dialog } from '../../components/design-system/Dialog.js';
 import { FastIcon, getFastIconString } from '../../components/FastIcon.js';
+import { useSettings } from '../../hooks/useSettings.js';
 import { Box, Link, Text } from '../../ink.js';
 import { useKeybindings } from '../../keybindings/useKeybinding.js';
 import { type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, logEvent } from '../../services/analytics/index.js';
@@ -11,7 +12,8 @@ import { type AppState, useAppState, useSetAppState } from '../../state/AppState
 import type { LocalJSXCommandOnDone } from '../../types/command.js';
 import { clearFastModeCooldown, FAST_MODE_MODEL_DISPLAY, getFastModeModel, getFastModeRuntimeState, getFastModeUnavailableReason, isFastModeEnabled, isFastModeSupportedByModel, prefetchFastModeStatus } from '../../utils/fastMode.js';
 import { formatDuration } from '../../utils/format.js';
-import { formatModelPricing, getOpus46CostTier } from '../../utils/modelCost.js';
+import { getDefaultOpusModel } from '../../utils/model/model.js';
+import { getModelPricingString } from '../../utils/modelCost.js';
 import { updateSettingsForSource } from '../../utils/settings/settings.js';
 function applyFastMode(enable: boolean, setAppState: (f: (prev: AppState) => AppState) => void): void {
  clearFastModeCooldown();
@@ -47,6 +49,7 @@ export function FastModePicker(t0) {
  const model = useAppState(_temp);
  const initialFastMode = useAppState(_temp2);
  const setAppState = useSetAppState();
+  useSettings();
  const [enableFastMode, setEnableFastMode] = useState(initialFastMode ?? false);
  let t1;
  if ($[0] === Symbol.for("react.memo_cache_sentinel")) {
@@ -58,14 +61,11 @@ export function FastModePicker(t0) {
  const runtimeState = t1;
  const isCooldown = runtimeState.status === "cooldown";
  const isUnavailable = unavailableReason !== null;
-  let t2;
-  if ($[1] === Symbol.for("react.memo_cache_sentinel")) {
-    t2 = formatModelPricing(getOpus46CostTier(true));
-    $[1] = t2;
-  } else {
-    t2 = $[1];
-  }
-  const pricing = t2;
+  // Settings are reloadable while the process is running, so do not pin this
+  // external-store-derived price in the compiler memo cache.
+  const pricing = getModelPricingString(getDefaultOpusModel(), {
+    speed: 'fast'
+  }) ?? '';
  let t3;
  if ($[2] !== enableFastMode || $[3] !== isUnavailable || $[4] !== model || $[5] !== onDone || $[6] !== setAppState) {
    t3 = function handleConfirm() {
@@ -80,7 +80,10 @@ export function FastModePicker(t0) {
      if (enableFastMode) {
        const fastIcon = getFastIconString(enableFastMode);
        const modelUpdated = !isFastModeSupportedByModel(model) ? ` · model set to ${FAST_MODE_MODEL_DISPLAY}` : "";
-        onDone(`${fastIcon} Fast mode ON${modelUpdated} · ${pricing}`);
+        const confirmationPricing = getModelPricingString(getDefaultOpusModel(), {
+          speed: 'fast'
+        }) ?? '';
+        onDone(`${fastIcon} Fast mode ON${modelUpdated} · ${confirmationPricing}`);
      } else {
        setAppState(_temp3);
        onDone("Fast mode OFF");
@@ -178,17 +181,11 @@ export function FastModePicker(t0) {
  } else {
    t9 = $[21];
  }
-  let t10;
-  if ($[22] !== enableFastMode || $[23] !== unavailableReason) {
-    t10 = unavailableReason ? <Box marginLeft={2}><Text color="error">{unavailableReason}</Text></Box> : <><Box flexDirection="column" gap={0} marginLeft={2}><Box flexDirection="row" gap={2}><Text bold={true}>Fast mode</Text><Text color={enableFastMode ? "fastMode" : undefined} bold={enableFastMode}>{enableFastMode ? "ON " : "OFF"}</Text><Text dimColor={true}>{pricing}</Text></Box></Box>{isCooldown && runtimeState.status === "cooldown" && <Box marginLeft={2}><Text color="warning">{runtimeState.reason === "overloaded" ? "Fast mode overloaded and is temporarily unavailable" : "You've hit your fast limit"}{" \xB7 resets in "}{formatDuration(runtimeState.resetAt - Date.now(), {
-            hideTrailingZeros: true
-          })}</Text></Box>}</>;
-    $[22] = enableFastMode;
-    $[23] = unavailableReason;
-    $[24] = t10;
-  } else {
-    t10 = $[24];
-  }
+  // Pricing comes from the reloadable settings cache, so keep the rendered
+  // value out of the compiler memo cache as well as the confirmation callback.
+  const t10 = unavailableReason ? <Box marginLeft={2}><Text color="error">{unavailableReason}</Text></Box> : <><Box flexDirection="column" gap={0} marginLeft={2}><Box flexDirection="row" gap={2}><Text bold={true}>Fast mode</Text><Text color={enableFastMode ? "fastMode" : undefined} bold={enableFastMode}>{enableFastMode ? "ON " : "OFF"}</Text><Text dimColor={true}>{pricing}</Text></Box></Box>{isCooldown && runtimeState.status === "cooldown" && <Box marginLeft={2}><Text color="warning">{runtimeState.reason === "overloaded" ? "Fast mode overloaded and is temporarily unavailable" : "You've hit your fast limit"}{" \xB7 resets in "}{formatDuration(runtimeState.resetAt - Date.now(), {
+          hideTrailingZeros: true
+        })}</Text></Box>}</>;
  let t11;
  if ($[25] === Symbol.for("react.memo_cache_sentinel")) {
    t11 = <Text dimColor={true}>Learn more:{" "}<Link url="https://code.claude.com/docs/en/fast-mode">https://code.claude.com/docs/en/fast-mode</Link></Text>;
@@ -223,7 +220,7 @@ function _temp2(s_0) {
 function _temp(s) {
  return s.mainLoopModel;
 }
-async function handleFastModeShortcut(enable: boolean, getAppState: () => AppState, setAppState: (f: (prev: AppState) => AppState) => void): Promise<string> {
+export async function handleFastModeShortcut(enable: boolean, getAppState: () => AppState, setAppState: (f: (prev: AppState) => AppState) => void): Promise<string> {
  const unavailableReason = getFastModeUnavailableReason();
  if (unavailableReason) {
    return `Fast mode unavailable: ${unavailableReason}`;
@@ -239,7 +236,9 @@ async function handleFastModeShortcut(enable: boolean, getAppState: () => AppSta
  if (enable) {
    const fastIcon = getFastIconString(true);
    const modelUpdated = !isFastModeSupportedByModel(mainLoopModel) ? ` · model set to ${FAST_MODE_MODEL_DISPLAY}` : '';
-    const pricing = formatModelPricing(getOpus46CostTier(true));
+    const pricing = getModelPricingString(getDefaultOpusModel(), {
+      speed: 'fast'
+    }) ?? '';
    return `${fastIcon} Fast mode ON${modelUpdated} · ${pricing}`;
  } else {
    return `Fast mode OFF`;
@@ -0,0 +1,143 @@
+import { afterEach, beforeEach, expect, test } from 'bun:test'
+import { mkdtempSync, rmSync, writeFileSync } from 'fs'
+import { tmpdir } from 'os'
+import { join } from 'path'
+import {
+  getAllowedSettingSources,
+  getFlagSettingsInline,
+  getFlagSettingsPath,
+  setAllowedSettingSources,
+  setFlagSettingsInline,
+  setFlagSettingsPath,
+} from './bootstrap/state.js'
+import {
+  acquireSharedMutationLock,
+  releaseSharedMutationLock,
+} from './test/sharedMutationLock.js'
+import {
+  addToTotalSessionCost,
+  formatTotalCost,
+  getModelUsage,
+  getTotalCost,
+  hasUnknownModelCost,
+  resetCostState,
+} from './cost-tracker.js'
+import { resetSettingsCache } from './utils/settings/settingsCache.js'
+import {
+  calculateCostFromTokens,
+  calculateUSDCost,
+} from './utils/modelCost.js'
+
+let tempDir: string
+let originalSources: ReturnType<typeof getAllowedSettingSources>
+let originalFlagPath: string | undefined
+let originalFlagInline: Record<string, unknown> | null
+
+beforeEach(async () => {
+  await acquireSharedMutationLock('cost-tracker.customPricing.test.ts')
+  tempDir = mkdtempSync(join(tmpdir(), 'openclaude-cost-pricing-'))
+  originalSources = [...getAllowedSettingSources()]
+  originalFlagPath = getFlagSettingsPath()
+  originalFlagInline = getFlagSettingsInline()
+  setAllowedSettingSources(['flagSettings'])
+  setFlagSettingsInline(null)
+  resetSettingsCache()
+  resetCostState()
+})
+
+afterEach(() => {
+  try {
+    resetCostState()
+    setAllowedSettingSources(originalSources)
+    setFlagSettingsPath(originalFlagPath)
+    setFlagSettingsInline(originalFlagInline)
+    resetSettingsCache()
+    rmSync(tempDir, { recursive: true, force: true })
+  } finally {
+    releaseSharedMutationLock()
+  }
+})
+
+test('raw helper, per-model display, and session total share one custom price', () => {
+  const model = 'provider/model:v1?profile=paid'
+  const settingsPath = join(tempDir, 'settings.json')
+  writeFileSync(
+    settingsPath,
+    `${JSON.stringify({
+      modelPricing: {
+        [model]: {
+          inputTokens: 1,
+          outputTokens: 2,
+          promptCacheReadTokens: 3,
+          promptCacheWriteTokens: 4,
+          webSearchRequests: 5,
+        },
+      },
+    })}\n`,
+    'utf8',
+  )
+  setFlagSettingsPath(settingsPath)
+  resetSettingsCache()
+
+  const usage = {
+    input_tokens: 1_000_000,
+    output_tokens: 2_000_000,
+    cache_read_input_tokens: 3_000_000,
+    cache_creation_input_tokens: 4_000_000,
+    server_tool_use: { web_search_requests: 2 },
+  } as Parameters<typeof calculateUSDCost>[1]
+  const cost = calculateUSDCost(model, usage)
+
+  expect(cost).toBe(40)
+  expect(
+    calculateCostFromTokens(model, {
+      inputTokens: 1_000_000,
+      outputTokens: 2_000_000,
+      cacheReadInputTokens: 3_000_000,
+      cacheCreationInputTokens: 4_000_000,
+    }),
+  ).toBe(30)
+  expect(addToTotalSessionCost(cost, usage, model)).toBe(40)
+  expect(getTotalCost()).toBe(40)
+  expect(getModelUsage()[model]).toMatchObject({
+    inputTokens: 1_000_000,
+    outputTokens: 2_000_000,
+    cacheReadInputTokens: 3_000_000,
+    cacheCreationInputTokens: 4_000_000,
+    webSearchRequests: 2,
+    costUSD: 40,
+  })
+  const display = formatTotalCost()
+  expect(display).toContain('Total cost:            $40.00')
+  expect(display).toContain('provider/model:v1?profile=paid:')
+  expect(display).toContain('($40.00)')
+  expect(hasUnknownModelCost()).toBe(false)
+})
+
+test('omitted web-search price uses the documented $0.01 request default', () => {
+  const model = 'provider/model-with-default-web-price'
+  const settingsPath = join(tempDir, 'settings.json')
+  writeFileSync(
+    settingsPath,
+    `${JSON.stringify({
+      modelPricing: {
+        [model]: {
+          inputTokens: 0,
+          outputTokens: 0,
+          promptCacheReadTokens: 0,
+          promptCacheWriteTokens: 0,
+        },
+      },
+    })}\n`,
+    'utf8',
+  )
+  setFlagSettingsPath(settingsPath)
+  resetSettingsCache()
+
+  const usage = {
+    input_tokens: 0,
+    output_tokens: 0,
+    server_tool_use: { web_search_requests: 3 },
+  } as Parameters<typeof calculateUSDCost>[1]
+  expect(calculateUSDCost(model, usage)).toBeCloseTo(0.03)
+})
@@ -0,0 +1,229 @@
+import { mock } from 'bun:test'
+import assert from 'node:assert/strict'
+import { mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { PassThrough } from 'node:stream'
+import { stripVTControlCharacters as stripAnsi } from 'node:util'
+import React from 'react'
+import { KeybindingSetup } from '../../keybindings/KeybindingProviderSetup.js'
+import {
+  getAllowedSettingSources,
+  getFlagSettingsInline,
+  getFlagSettingsPath,
+  resetModelStringsForTestingOnly,
+  setAllowedSettingSources,
+  setFlagSettingsInline,
+  setFlagSettingsPath,
+} from '../../bootstrap/state.js'
+import { resetSettingsCache } from '../../utils/settings/settingsCache.js'
+
+const originalOpenClaudeConfigDir = process.env.OPENCLAUDE_CONFIG_DIR
+const fixtureDir = mkdtempSync(join(tmpdir(), 'openclaude-pricing-display-'))
+const userConfigDir = join(fixtureDir, 'user-config')
+const pricingSettingsPath = join(fixtureDir, 'pricing-settings.json')
+process.env.OPENCLAUDE_CONFIG_DIR = userConfigDir
+
+mock.module('../../utils/model/providers.js', () => ({
+  getAPIProvider: () => 'firstParty',
+  getAPIProviderForStatsig: () => 'firstParty',
+  isFirstPartyAnthropicBaseUrl: () => true,
+  isFirstPartyAnthropicProvider: () => true,
+  isCustomAnthropicProvider: () => false,
+  isGithubNativeAnthropicMode: () => false,
+  usesAnthropicAccountFlow: () => true,
+}))
+mock.module('../../utils/auth.js', () => ({
+  getSubscriptionType: () => null,
+  isClaudeAISubscriber: () => false,
+  isMaxSubscriber: () => false,
+  isProSubscriber: () => false,
+  isTeamPremiumSubscriber: () => false,
+}))
+mock.module('../../utils/fastMode.js', () => ({
+  clearFastModeCooldown: () => {},
+  FAST_MODE_MODEL_DISPLAY: 'Opus 4.8',
+  getFastModeModel: () => 'opus',
+  getFastModeRuntimeState: () => ({ status: 'active' }),
+  getFastModeUnavailableReason: () => null,
+  isFastModeEnabled: () => true,
+  isFastModeSupportedByModel: () => true,
+  prefetchFastModeStatus: async () => {},
+}))
+
+const originalSources = [...getAllowedSettingSources()]
+const originalFlagPath = getFlagSettingsPath()
+const originalFlagInline = getFlagSettingsInline()
+
+function writePricing(opusInput: number, opusOutput: number): void {
+  writeFileSync(
+    pricingSettingsPath,
+    `${JSON.stringify({
+      modelPricing: {
+        'claude-sonnet-4-6': {
+          inputTokens: 9,
+          outputTokens: 10,
+          promptCacheReadTokens: 0,
+          promptCacheWriteTokens: 0,
+          webSearchRequests: 0,
+        },
+        'claude-opus-4-8': {
+          inputTokens: opusInput,
+          outputTokens: opusOutput,
+          promptCacheReadTokens: 0,
+          promptCacheWriteTokens: 0,
+          webSearchRequests: 0,
+        },
+      },
+    })}\n`,
+    'utf8',
+  )
+}
+
+async function waitFor(
+  check: () => boolean,
+  description: string,
+  timeoutMs = 5_000,
+): Promise<void> {
+  const deadline = Date.now() + timeoutMs
+  while (!check()) {
+    if (Date.now() >= deadline) {
+      throw new Error(`Timed out waiting for ${description}`)
+    }
+    await Bun.sleep(5)
+  }
+}
+
+let cleanupRender: (() => void) | undefined
+
+try {
+  writePricing(7, 8)
+  setAllowedSettingSources(['flagSettings'])
+  setFlagSettingsPath(pricingSettingsPath)
+  setFlagSettingsInline(null)
+  resetSettingsCache()
+  resetModelStringsForTestingOnly()
+
+  const {
+    getDefaultOptionForUser,
+    getMaxOpus46_1MOption,
+    getMaxSonnet46_1MOption,
+    getOpus46_1MOption,
+    getSonnet46_1MOption,
+  } = await import('../../utils/model/modelOptions.js')
+
+  assert.match(getDefaultOptionForUser().description, /\$9\/\$10 per Mtok/)
+  assert.match(getSonnet46_1MOption().description, /\$9\/\$10 per Mtok/)
+  assert.match(getMaxSonnet46_1MOption().description, /\$9\/\$10 per Mtok/)
+  assert.match(getOpus46_1MOption(true).description, /\$7\/\$8 per Mtok/)
+  assert.match(getMaxOpus46_1MOption(true).description, /\$7\/\$8 per Mtok/)
+
+  const { FastModePicker, handleFastModeShortcut } = await import(
+    '../../commands/fast/fast.js'
+  )
+  const { AppStateProvider, getDefaultAppState, useAppStateStore } =
+    await import('../../state/AppState.js')
+  const { createRoot } = await import('../../ink.js')
+  let appState = getDefaultAppState()
+  const shortcut = await handleFastModeShortcut(
+    true,
+    () => appState,
+    update => {
+      appState = update(appState)
+    },
+  )
+  assert.match(shortcut, /\$7\/\$8 per Mtok/)
+  const isolatedUserSettings = JSON.parse(
+    readFileSync(join(userConfigDir, 'settings.json'), 'utf8'),
+  ) as Record<string, unknown>
+  assert.equal(isolatedUserSettings.fastMode, true)
+
+  let output = ''
+  const stdout = new PassThrough()
+  const stdin = new PassThrough() as PassThrough & {
+    isTTY: boolean
+    ref: () => PassThrough
+    setRawMode: (mode: boolean) => void
+    unref: () => PassThrough
+  }
+  stdin.isTTY = true
+  stdin.ref = () => stdin
+  stdin.setRawMode = () => {}
+  stdin.unref = () => stdin
+  ;(stdout as unknown as { columns: number }).columns = 120
+  stdout.on('data', chunk => {
+    output += chunk.toString()
+  })
+  const completions: string[] = []
+  const onPickerDone = (message: string) => completions.push(message)
+  let triggerSettingsReload: (() => void) | undefined
+  const PickerWithSettingsReload = () => {
+    const store = useAppStateStore()
+    triggerSettingsReload = () => {
+      store.setState(previous => ({
+        ...previous,
+        settings: { ...previous.settings },
+      }))
+    }
+    return <FastModePicker onDone={onPickerDone} unavailableReason={null} />
+  }
+  const picker = () => (
+    <AppStateProvider initialState={appState}>
+      <KeybindingSetup>
+        <PickerWithSettingsReload />
+      </KeybindingSetup>
+    </AppStateProvider>
+  )
+  const instance = await createRoot({
+    stdin: stdin as unknown as NodeJS.ReadStream,
+    stdout: stdout as unknown as NodeJS.WriteStream,
+    exitOnCtrlC: false,
+    patchConsole: false,
+  })
+  cleanupRender = () => {
+    instance.unmount()
+    stdin.end()
+    stdout.end()
+  }
+  instance.render(picker())
+  await waitFor(
+    () => /\$7\/\$8 per Mtok/.test(stripAnsi(output)),
+    'the initial fast-mode pricing render',
+  )
+  assert.match(stripAnsi(output), /\$7\/\$8 per Mtok/)
+
+  writePricing(17, 18)
+  resetSettingsCache()
+  output = ''
+  assert.ok(triggerSettingsReload)
+  triggerSettingsReload()
+  await waitFor(
+    () => /\$17\/\$18 per Mtok/.test(stripAnsi(output)),
+    'the reloaded fast-mode pricing render',
+  )
+  assert.match(stripAnsi(output), /\$17\/\$18 per Mtok/)
+
+  stdin.write('\r')
+  await waitFor(
+    () => /\$17\/\$18 per Mtok/.test(completions.at(-1) ?? ''),
+    'the fast-mode confirmation',
+  )
+  assert.match(completions.at(-1) ?? '', /\$17\/\$18 per Mtok/)
+
+  cleanupRender()
+  cleanupRender = undefined
+} finally {
+  cleanupRender?.()
+  mock.restore()
+  resetModelStringsForTestingOnly()
+  setAllowedSettingSources(originalSources)
+  setFlagSettingsPath(originalFlagPath)
+  setFlagSettingsInline(originalFlagInline)
+  resetSettingsCache()
+  if (originalOpenClaudeConfigDir === undefined) {
+    delete process.env.OPENCLAUDE_CONFIG_DIR
+  } else {
+    process.env.OPENCLAUDE_CONFIG_DIR = originalOpenClaudeConfigDir
+  }
+  rmSync(fixtureDir, { recursive: true, force: true })
+}
@@ -0,0 +1,234 @@
+import { mock } from 'bun:test'
+import assert from 'node:assert/strict'
+import { mkdtempSync, rmSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import {
+  getAllowedSettingSources,
+  getFlagSettingsInline,
+  getFlagSettingsPath,
+  isSessionPersistenceDisabled,
+  setAllowedSettingSources,
+  setFlagSettingsInline,
+  setFlagSettingsPath,
+  setSessionPersistenceDisabled,
+} from '../../bootstrap/state.js'
+import type { SDKMessage } from '../../entrypoints/agentSdkTypes.js'
+import type { AppState } from '../../state/AppState.js'
+import type { Message } from '../../types/message.js'
+import {
+  addToTotalSessionCost,
+  formatTotalCost,
+  resetCostState,
+} from '../../cost-tracker.js'
+import {
+  calculateCostFromTokens,
+  calculateUSDCost,
+} from '../../utils/modelCost.js'
+import { resetSettingsCache } from '../../utils/settings/settingsCache.js'
+
+const paidModel = 'provider/paid-model'
+const freeModel = 'nvidia/free-model'
+const usage = {
+  input_tokens: 10_000,
+  output_tokens: 0,
+  cache_read_input_tokens: 0,
+  cache_creation_input_tokens: 0,
+}
+
+let requestedModel = paidModel
+
+mock.module('../../utils/processUserInput/processUserInput.js', () => ({
+  processUserInput: mock(async () => ({
+    messages: [
+      {
+        type: 'user',
+        message: { role: 'user', content: 'run' },
+        isMeta: false,
+        uuid: `user-${Math.random()}`,
+        timestamp: new Date().toISOString(),
+      } as Message,
+    ],
+    shouldQuery: true,
+    allowedTools: [],
+    model: requestedModel,
+    resultText: undefined,
+  })),
+}))
+mock.module('../../utils/queryContext.js', () => ({
+  fetchSystemPromptParts: mock(async () => ({
+    defaultSystemPrompt: [],
+    userContext: {},
+    systemContext: {},
+  })),
+}))
+mock.module('../../utils/messages/systemInit.js', () => ({
+  buildSystemInitMessage: mock((input: { model: string }) => ({
+    type: 'system',
+    subtype: 'init',
+    session_id: 'test-session',
+    tools: [],
+    mcp_servers: [],
+    model: input.model,
+    permissionMode: 'default',
+    apiKeySource: 'none',
+    cwd: process.cwd(),
+  })),
+  sdkCompatToolName: (name: string) => name,
+}))
+mock.module('../../commands.js', () => ({
+  REMOTE_SAFE_COMMANDS: new Set<string>(),
+  builtInCommandNames: new Set<string>(),
+  clearCommandsCache: () => {},
+  findCommand: () => undefined,
+  getCommand: () => undefined,
+  getCommandName: (command: { name?: string }) => command.name ?? '',
+  getCommands: () => [],
+  getMcpSkillCommands: () => [],
+  getSkillToolCommands: () => [],
+  getSlashCommandToolSkills: mock(async () => []),
+  hasCommand: () => false,
+  isCommandEnabled: () => true,
+}))
+mock.module('src/entrypoints/agentSdkTypes.js', () => ({ HOOK_EVENTS: [] }))
+
+async function* pricedQuery({
+  toolUseContext,
+}: {
+  toolUseContext: { options: { mainLoopModel: string } }
+}): AsyncGenerator<Message> {
+  const model = toolUseContext.options.mainLoopModel
+  const cost = calculateUSDCost(model, usage as never)
+  addToTotalSessionCost(cost, usage as never, model)
+  yield {
+    type: 'assistant',
+    message: {
+      id: `msg-${Math.random()}`,
+      type: 'message',
+      role: 'assistant',
+      model,
+      content: [{ type: 'text', text: 'done' }],
+      stop_reason: 'end_turn',
+      stop_sequence: null,
+      usage,
+    },
+    uuid: `assistant-${Math.random()}`,
+    timestamp: new Date().toISOString(),
+  } as Message
+}
+
+async function runCase(
+  QueryEngine: typeof import('../../QueryEngine.js').QueryEngine,
+  model: string,
+  maxBudgetUsd: number,
+): Promise<Extract<SDKMessage, { type: 'result' }>> {
+  requestedModel = model
+  let appState = {
+    fastMode: false,
+    toolPermissionContext: {
+      mode: 'default',
+      additionalWorkingDirectories: new Map(),
+      alwaysAllowRules: {},
+    },
+    fileHistory: {},
+    attribution: {},
+  } as unknown as AppState
+  const engine = new QueryEngine({
+    cwd: fixtureDir,
+    tools: [],
+    commands: [],
+    mcpClients: [],
+    agents: [],
+    canUseTool: async () => ({ behavior: 'allow' }),
+    getAppState: () => appState,
+    setAppState: update => {
+      appState = update(appState)
+    },
+    readFileCache: {} as never,
+    userSpecifiedModel: model,
+    thinkingConfig: { type: 'disabled' },
+    maxBudgetUsd,
+    query: pricedQuery as never,
+  })
+  const events: SDKMessage[] = []
+  for await (const event of engine.submitMessage('run')) events.push(event)
+  const result = events.findLast(
+    (event): event is Extract<SDKMessage, { type: 'result' }> =>
+      event.type === 'result',
+  )
+  assert.ok(result)
+  return result
+}
+
+const originalPersistence = isSessionPersistenceDisabled()
+const originalSources = [...getAllowedSettingSources()]
+const originalFlagPath = getFlagSettingsPath()
+const originalFlagInline = getFlagSettingsInline()
+const fixtureDir = mkdtempSync(join(tmpdir(), 'openclaude-query-budget-'))
+const settingsPath = join(fixtureDir, 'settings.json')
+
+try {
+  writeFileSync(
+    settingsPath,
+    `${JSON.stringify({
+      modelPricing: {
+        [paidModel]: {
+          inputTokens: 100,
+          outputTokens: 0,
+          promptCacheReadTokens: 0,
+          promptCacheWriteTokens: 0,
+          webSearchRequests: 0,
+        },
+        [freeModel]: {
+          inputTokens: 0,
+          outputTokens: 0,
+          promptCacheReadTokens: 0,
+          promptCacheWriteTokens: 0,
+          webSearchRequests: 0,
+        },
+      },
+    })}\n`,
+    'utf8',
+  )
+  setSessionPersistenceDisabled(true)
+  setAllowedSettingSources(['flagSettings'])
+  setFlagSettingsPath(settingsPath)
+  setFlagSettingsInline(null)
+  resetSettingsCache()
+
+  const { QueryEngine } = await import('../../QueryEngine.js')
+
+  resetCostState()
+  const paid = await runCase(QueryEngine, paidModel, 0.5)
+  assert.equal(paid.subtype, 'error_max_budget_usd')
+  assert.equal(paid.total_cost_usd, 1)
+  assert.equal(paid.modelUsage[paidModel]?.costUSD, 1)
+  assert.equal(
+    calculateCostFromTokens(paidModel, {
+      inputTokens: 10_000,
+      outputTokens: 0,
+      cacheReadInputTokens: 0,
+      cacheCreationInputTokens: 0,
+    }),
+    1,
+  )
+  assert.match(formatTotalCost(), /Total cost:\s+\$1\.00/)
+  assert.match(formatTotalCost(), /provider\/paid-model:.*\(\$1\.00\)/s)
+  assert.match(JSON.stringify(paid), /"total_cost_usd":1/)
+
+  resetCostState()
+  const free = await runCase(QueryEngine, freeModel, 0.01)
+  assert.equal(free.subtype, 'success')
+  assert.equal(free.total_cost_usd, 0)
+  assert.equal(free.modelUsage[freeModel]?.costUSD, 0)
+  assert.equal(free.modelUsage[freeModel]?.inputTokens, 10_000)
+} finally {
+  mock.restore()
+  resetCostState()
+  setSessionPersistenceDisabled(originalPersistence)
+  setAllowedSettingSources(originalSources)
+  setFlagSettingsPath(originalFlagPath)
+  setFlagSettingsInline(originalFlagInline)
+  resetSettingsCache()
+  rmSync(fixtureDir, { recursive: true, force: true })
+}
@@ -20,7 +20,7 @@ import {
 } from '../context.js'
 import { isEnvTruthy } from '../envUtils.js'
 import { getModelStrings, resolveOverriddenModel } from './modelStrings.js'
-import { formatModelPricing, getOpus46CostTier } from '../modelCost.js'
+import { getModelPricingString } from '../modelCost.js'
 import { getSettings_DEPRECATED } from '../settings/settings.js'
 import type { PermissionMode } from '../permissions/PermissionMode.js'
 import {
@@ -562,9 +562,15 @@ export function renderDefaultModelSetting(
  return renderModelName(parseUserSpecifiedModel(setting))
 }

-export function getOpus46PricingSuffix(fastMode: boolean): string {
+export function getOpus46PricingSuffix(
+  fastMode: boolean,
+  model: string = getModelStrings().opus48,
+): string {
  if (!isFirstPartyAnthropicProvider()) return ''
-  const pricing = formatModelPricing(getOpus46CostTier(fastMode))
+  const pricing = getModelPricingString(model, {
+    speed: fastMode ? 'fast' : 'standard',
+  })
+  if (!pricing) return ''
  const fastModeIndicator = fastMode ? ` (${LIGHTNING_BOLT})` : ''
  return ` ·${fastModeIndicator} ${pricing}`
 }
@@ -143,6 +143,19 @@ test('custom Anthropic endpoints use the third-party default description', async
  expect(defaultOption?.description).not.toContain('$')
 })

+test('custom Anthropic endpoints omit first-party pricing from every model option', async () => {
+  process.env.ANTHROPIC_BASE_URL = 'https://proxy.example/v1'
+  process.env.ANTHROPIC_API_KEY = 'proxy-key'
+
+  const { getModelOptions } = await importFreshModelOptionsModule('firstParty')
+  const options = getModelOptions()
+
+  expect(options.length).toBeGreaterThan(1)
+  for (const option of options) {
+    expect(option.description).not.toContain('$')
+  }
+})
+
 test('OpenRouter active profile cache merges with the static route catalog', async () => {
  process.env.CLAUDE_CODE_USE_OPENAI = '1'
  process.env.OPENAI_BASE_URL = 'https://openrouter.ai/api/v1'
@@ -16,12 +16,7 @@ import {
  isTeamPremiumSubscriber,
 } from '../auth.js'
 import { getModelStrings } from './modelStrings.js'
-import {
-  COST_TIER_3_15,
-  COST_HAIKU_35,
-  COST_HAIKU_45,
-  formatModelPricing,
-} from '../modelCost.js'
+import { getModelPricingString } from '../modelCost.js'
 import { getSettings_DEPRECATED } from '../settings/settings.js'
 import { checkOpus1mAccess, checkSonnet1mAccess } from './check1mAccess.js'
 import {
@@ -187,10 +182,16 @@ export function getDefaultOptionForUser(fastMode = false): ModelOption {
  return {
    value: null,
    label: 'Default (recommended)',
-    description: `Use the default model (currently ${renderDefaultModelSetting(currentDefaultModel)})${is3P ? '' : ` · ${formatModelPricing(COST_TIER_3_15)}`}`,
+    description: `Use the default model (currently ${renderDefaultModelSetting(currentDefaultModel)})${getPricingSuffix(getDefaultSonnetModel())}`,
  }
 }

+function getPricingSuffix(model: string): string {
+  if (!isFirstPartyAnthropicProvider()) return ''
+  const pricing = getModelPricingString(model)
+  return pricing ? ` · ${pricing}` : ''
+}
+
 function getCustomSonnetOption(): ModelOption | undefined {
  const is3P = getAPIProvider() !== 'firstParty'
  const customSonnetModel = process.env.ANTHROPIC_DEFAULT_SONNET_MODEL
@@ -216,7 +217,7 @@ function getSonnet46Option(): ModelOption {
  return {
    value: is3P ? getModelStrings().sonnet46 : 'sonnet',
    label: 'Sonnet',
-    description: `Sonnet 4.6 · Best for everyday tasks${is3P ? '' : ` · ${formatModelPricing(COST_TIER_3_15)}`}`,
+    description: `Sonnet 4.6 · Best for everyday tasks${getPricingSuffix(getModelStrings().sonnet46)}`,
    descriptionForModel:
      'Sonnet 4.6 - best for everyday tasks. Generally recommended for most coding tasks',
  }
@@ -253,7 +254,7 @@ function getOpus48Option(fastMode = false): ModelOption {
  return {
    value: is3P ? getModelStrings().opus48 : 'opus',
    label: 'Opus',
-    description: `Opus 4.8 · Most capable for complex work${getOpus46PricingSuffix(fastMode)}`,
+    description: `Opus 4.8 · Most capable for complex work${getOpus46PricingSuffix(fastMode, getModelStrings().opus48)}`,
    descriptionForModel: 'Opus 4.8 - most capable for complex work',
  }
 }
@@ -263,7 +264,7 @@ function getOpus47Option(fastMode = false): ModelOption {
  return {
    value: is3P ? getModelStrings().opus47 : 'opus',
    label: 'Opus',
-    description: `Opus 4.7 · Most capable for complex work${getOpus46PricingSuffix(fastMode)}`,
+    description: `Opus 4.7 · Most capable for complex work${getOpus46PricingSuffix(fastMode, getModelStrings().opus47)}`,
    descriptionForModel: 'Opus 4.7 - most capable for complex work',
  }
 }
@@ -273,7 +274,7 @@ function getOpus46Option(fastMode = false): ModelOption {
  return {
    value: is3P ? getModelStrings().opus46 : 'opus',
    label: 'Opus',
-    description: `Opus 4.6 · Most capable for complex work${getOpus46PricingSuffix(fastMode)}`,
+    description: `Opus 4.6 · Most capable for complex work${getOpus46PricingSuffix(fastMode, getModelStrings().opus46)}`,
    descriptionForModel: 'Opus 4.6 - most capable for complex work',
  }
 }
@@ -283,7 +284,7 @@ export function getSonnet46_1MOption(): ModelOption {
  return {
    value: is3P ? getModelStrings().sonnet46 + '[1m]' : 'sonnet[1m]',
    label: 'Sonnet (1M context)',
-    description: `Sonnet 4.6 for long sessions${is3P ? '' : ` · ${formatModelPricing(COST_TIER_3_15)}`}`,
+    description: `Sonnet 4.6 for long sessions${getPricingSuffix(getModelStrings().sonnet46)}`,
    descriptionForModel:
      'Sonnet 4.6 with 1M context window - for long sessions with large codebases',
  }
@@ -297,7 +298,7 @@ export function getOpus46_1MOption(fastMode = false): ModelOption {
  return {
    value: is3P ? getModelStrings().opus46 + '[1m]' : 'opus[1m]',
    label: 'Opus (1M context)',
-    description: `${opusName} for long sessions${getOpus46PricingSuffix(fastMode)}`,
+    description: `${opusName} for long sessions${getOpus46PricingSuffix(fastMode, is3P ? getModelStrings().opus46 : getModelStrings().opus48)}`,
    descriptionForModel: `${opusName} with 1M context window - for long sessions with large codebases`,
  }
 }
@@ -319,22 +320,20 @@ function getCustomHaikuOption(): ModelOption | undefined {
 }

 function getHaiku45Option(): ModelOption {
-  const is3P = getAPIProvider() !== 'firstParty'
  return {
    value: 'haiku',
    label: 'Haiku',
-    description: `Haiku 4.5 · Fastest for quick answers${is3P ? '' : ` · ${formatModelPricing(COST_HAIKU_45)}`}`,
+    description: `Haiku 4.5 · Fastest for quick answers${getPricingSuffix(getModelStrings().haiku45)}`,
    descriptionForModel:
      'Haiku 4.5 - fastest for quick answers. Lower cost but less capable than Sonnet 4.6.',
  }
 }

 function getHaiku35Option(): ModelOption {
-  const is3P = getAPIProvider() !== 'firstParty'
  return {
    value: 'haiku',
    label: 'Haiku',
-    description: `Haiku 3.5 for simple tasks${is3P ? '' : ` · ${formatModelPricing(COST_HAIKU_35)}`}`,
+    description: `Haiku 3.5 for simple tasks${getPricingSuffix(getModelStrings().haiku35)}`,
    descriptionForModel:
      'Haiku 3.5 - faster and lower cost, but less capable than Sonnet. Use for simple tasks.',
  }
@@ -352,17 +351,16 @@ function getMaxOpusOption(fastMode = false): ModelOption {
  return {
    value: 'opus',
    label: 'Opus',
-    description: `Opus 4.8 · Most capable for complex work${fastMode ? getOpus46PricingSuffix(true) : ''}`,
+    description: `Opus 4.8 · Most capable for complex work${fastMode ? getOpus46PricingSuffix(true, getModelStrings().opus48) : ''}`,
  }
 }

 export function getMaxSonnet46_1MOption(): ModelOption {
-  const is3P = getAPIProvider() !== 'firstParty'
  const billingInfo = isClaudeAISubscriber() ? ' · Billed as extra usage' : ''
  return {
    value: 'sonnet[1m]',
    label: 'Sonnet (1M context)',
-    description: `Sonnet 4.6 with 1M context${billingInfo}${is3P ? '' : ` · ${formatModelPricing(COST_TIER_3_15)}`}`,
+    description: `Sonnet 4.6 with 1M context${billingInfo}${getPricingSuffix(getModelStrings().sonnet46)}`,
  }
 }

@@ -371,7 +369,7 @@ export function getMaxOpus46_1MOption(fastMode = false): ModelOption {
  return {
    value: 'opus[1m]',
    label: 'Opus (1M context)',
-    description: `Opus 4.8 with 1M context${billingInfo}${getOpus46PricingSuffix(fastMode)}`,
+    description: `Opus 4.8 with 1M context${billingInfo}${getOpus46PricingSuffix(fastMode, getModelStrings().opus48)}`,
  }
 }

@@ -380,7 +378,7 @@ function getMergedOpus1MOption(fastMode = false): ModelOption {
  return {
    value: is3P ? getModelStrings().opus46 + '[1m]' : 'opus[1m]',
    label: 'Opus (1M context)',
-    description: `${is3P ? 'Opus 4.6' : 'Opus 4.8'} with 1M context · Most capable for complex work${!is3P && fastMode ? getOpus46PricingSuffix(fastMode) : ''}`,
+    description: `${is3P ? 'Opus 4.6' : 'Opus 4.8'} with 1M context · Most capable for complex work${!is3P && fastMode ? getOpus46PricingSuffix(fastMode, getModelStrings().opus48) : ''}`,
    descriptionForModel:
      `${is3P ? 'Opus 4.6' : 'Opus 4.8'} with 1M context - most capable for complex work`,
  }
@@ -1,10 +1,36 @@
 import { afterEach, beforeEach, expect, mock, test } from 'bun:test'
-import { hasUnknownModelCost, resetCostState } from '../bootstrap/state.js'
+import {
+  getAllowedSettingSources,
+  getFlagSettingsInline,
+  getFlagSettingsPath,
+  hasUnknownModelCost,
+  resetCostState,
+  setAllowedSettingSources,
+  setFlagSettingsInline,
+  setFlagSettingsPath,
+} from '../bootstrap/state.js'
 import {
  acquireSharedMutationLock,
  releaseSharedMutationLock,
 } from '../test/sharedMutationLock.js'
 import * as realFastMode from './fastMode.js'
+import * as realModel from './model/model.js'
+import { resetSettingsCache } from './settings/settingsCache.js'
+
+const realModelSnapshot = { ...realModel }
+
+type PricingOverride = {
+  inputTokens: number
+  outputTokens: number
+  promptCacheReadTokens: number
+  promptCacheWriteTokens: number
+  webSearchRequests: number
+}
+
+let pricingByModel: Record<string, PricingOverride>
+let originalSources: ReturnType<typeof getAllowedSettingSources>
+let originalFlagPath: string | undefined
+let originalFlagInline: Record<string, unknown> | null

 async function importFreshModelCost() {
  return import(`./modelCost.js?ts=${Date.now()}-${Math.random()}`)
@@ -12,12 +38,25 @@ async function importFreshModelCost() {

 beforeEach(async () => {
  await acquireSharedMutationLock('utils/modelCost.modelGate.test.ts')
+  originalSources = [...getAllowedSettingSources()]
+  originalFlagPath = getFlagSettingsPath()
+  originalFlagInline = getFlagSettingsInline()
+  pricingByModel = Object.create(null) as Record<string, PricingOverride>
+  setAllowedSettingSources(['flagSettings'])
+  setFlagSettingsPath(undefined)
+  setFlagSettingsInline({ modelPricing: pricingByModel })
+  resetSettingsCache()
 })

 afterEach(() => {
  try {
    mock.restore()
    mock.module('./fastMode.js', () => realFastMode)
+    mock.module('./model/model.js', () => realModelSnapshot)
+    setAllowedSettingSources(originalSources)
+    setFlagSettingsPath(originalFlagPath)
+    setFlagSettingsInline(originalFlagInline)
+    resetSettingsCache()
  } finally {
    releaseSharedMutationLock()
  }
@@ -111,3 +150,127 @@ test('proto-member model ids fall through the unknown-model path, not NaN', asyn
    resetCostState()
  }
 })
+
+test('exact unknown-model override prices every usage field and suppresses the unknown warning', async () => {
+  const model = 'provider/model:v1?profile=paid'
+  pricingByModel[model] = {
+    inputTokens: 1,
+    outputTokens: 2,
+    promptCacheReadTokens: 3,
+    promptCacheWriteTokens: 4,
+    webSearchRequests: 5,
+  }
+  const {
+    calculateUSDCost,
+    calculateCostFromTokens,
+    getModelCosts,
+    getModelPricingString,
+  } = await importFreshModelCost()
+  const usage = {
+    input_tokens: 1_000_000,
+    output_tokens: 2_000_000,
+    cache_read_input_tokens: 3_000_000,
+    cache_creation_input_tokens: 4_000_000,
+    server_tool_use: { web_search_requests: 2 },
+  } as Parameters<typeof calculateUSDCost>[1]
+
+  expect(getModelCosts(model, usage)).toEqual(pricingByModel[model])
+  expect(calculateUSDCost(model, usage)).toBe(40)
+  expect(
+    calculateCostFromTokens(model, {
+      inputTokens: 1_000_000,
+      outputTokens: 2_000_000,
+      cacheReadInputTokens: 3_000_000,
+      cacheCreationInputTokens: 4_000_000,
+    }),
+  ).toBe(30)
+  expect(getModelPricingString(model)).toBe('$1/$2 per Mtok')
+  expect(hasUnknownModelCost()).toBe(false)
+})
+
+test('all-zero exact override is authoritative for unknown and fast-mode known models', async () => {
+  const zero = {
+    inputTokens: 0,
+    outputTokens: 0,
+    promptCacheReadTokens: 0,
+    promptCacheWriteTokens: 0,
+    webSearchRequests: 0,
+  }
+  pricingByModel['nvidia/free-model'] = zero
+  pricingByModel['claude-opus-4-8'] = zero
+  mock.module('./fastMode.js', () => ({
+    ...realFastMode,
+    isFastModeEnabled: () => true,
+  }))
+  const { calculateUSDCost, getModelCosts, getModelPricingString } =
+    await importFreshModelCost()
+  const usage = {
+    input_tokens: 1_000_000,
+    output_tokens: 1_000_000,
+    cache_read_input_tokens: 1_000_000,
+    cache_creation_input_tokens: 1_000_000,
+    server_tool_use: { web_search_requests: 1 },
+    speed: 'fast',
+  } as Parameters<typeof calculateUSDCost>[1]
+
+  for (const model of ['nvidia/free-model', 'claude-opus-4-8']) {
+    expect(getModelCosts(model, usage)).toEqual(zero)
+    expect(calculateUSDCost(model, usage)).toBe(0)
+    expect(getModelPricingString(model)).toBe('$0/$0 per Mtok')
+  }
+  expect(hasUnknownModelCost()).toBe(false)
+})
+
+test('custom pricing matches exact resolved ids only, including unusual own keys', async () => {
+  mock.restore()
+  mock.module('./fastMode.js', () => realFastMode)
+  mock.module('./model/model.js', () => realModelSnapshot)
+  const configured = {
+    inputTokens: 7,
+    outputTokens: 11,
+    promptCacheReadTokens: 1,
+    promptCacheWriteTokens: 2,
+    webSearchRequests: 0.5,
+  }
+  const exact = 'Provider/model:v1?route=alpha'
+  pricingByModel[exact] = configured
+  for (const model of ['constructor', 'toString', '__proto__']) {
+    Object.defineProperty(pricingByModel, model, {
+      configurable: true,
+      enumerable: true,
+      value: configured,
+      writable: true,
+    })
+  }
+  const {
+    calculateUSDCost,
+    getModelCosts,
+    COST_TIER_5_25,
+    DEFAULT_UNKNOWN_MODEL_COST,
+  } = await importFreshModelCost()
+  const usage = {
+    input_tokens: 1_000_000,
+    output_tokens: 0,
+  } as Parameters<typeof calculateUSDCost>[1]
+
+  for (const model of [exact, 'constructor', 'toString', '__proto__']) {
+    expect(calculateUSDCost(model, usage)).toBe(7)
+  }
+  resetCostState()
+  for (const nearMatch of [
+    exact.toLowerCase(),
+    exact.slice(0, -1),
+    `${exact}/child`,
+  ]) {
+    expect(getModelCosts(nearMatch, usage)).toEqual(
+      DEFAULT_UNKNOWN_MODEL_COST,
+    )
+  }
+  expect(hasUnknownModelCost()).toBe(true)
+
+  resetCostState()
+  expect(getModelCosts('claude-opus-4-8-20260815', usage)).toEqual(
+    COST_TIER_5_25,
+  )
+  expect(hasUnknownModelCost()).toBe(false)
+})
@@ -3,6 +3,7 @@ import type { AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS } from
 import { logEvent } from 'src/services/analytics/index.js'
 import { setHasUnknownModelCost } from '../bootstrap/state.js'
 import { isFastModeEnabled } from './fastMode.js'
+import { getModelPricingOverride } from './settings/modelPricing.js'
 import {
  CLAUDE_3_5_HAIKU_CONFIG,
  CLAUDE_3_5_V2_SONNET_CONFIG,
@@ -87,7 +88,7 @@ export const COST_HAIKU_45 = {
  webSearchRequests: 0.01,
 } as const satisfies ModelCosts

-const DEFAULT_UNKNOWN_MODEL_COST = COST_TIER_5_25
+export const DEFAULT_UNKNOWN_MODEL_COST = COST_TIER_5_25

 /**
 * Get the cost tier for Opus 4.6 based on fast mode.
@@ -146,7 +147,18 @@ function tokensToUSDCost(modelCosts: ModelCosts, usage: Usage): number {
  )
 }

-export function getModelCosts(model: string, usage: Usage): ModelCosts {
+function getKnownModelCosts(
+  model: string,
+  usage: { speed?: Usage['speed'] },
+): ModelCosts | undefined {
+  // Custom prices match the exact resolved model id before any canonical
+  // first-party fallback. This intentionally does not normalize case, aliases,
+  // prefixes, route names, or provider-specific identifiers.
+  const override = getModelPricingOverride(model)
+  if (override) {
+    return override
+  }
+
  const shortName = getCanonicalName(model)

  // Check if this is a fast-mode-capable Opus model (4.8/4.7/4.6) with fast mode
@@ -168,11 +180,19 @@ export function getModelCosts(model: string, usage: Usage): ModelCosts {
  // getCanonicalName passes them through unchanged) would return a truthy
  // prototype value, skip the unknown-model path, and yield NaN costs downstream.
  // Match on own properties only.
-  if (!Object.hasOwn(MODEL_COSTS, shortName)) {
-    trackUnknownModelCost(model, shortName)
-    return DEFAULT_UNKNOWN_MODEL_COST
+  return Object.hasOwn(MODEL_COSTS, shortName)
+    ? MODEL_COSTS[shortName]
+    : undefined
+}
+
+export function getModelCosts(model: string, usage: Usage): ModelCosts {
+  const costs = getKnownModelCosts(model, usage)
+  if (costs) {
+    return costs
  }
-  return MODEL_COSTS[shortName]
+
+  trackUnknownModelCost(model, getCanonicalName(model))
+  return DEFAULT_UNKNOWN_MODEL_COST
 }

 function trackUnknownModelCost(model: string, shortName: ModelShortName): void {
@@ -236,10 +256,10 @@ export function formatModelPricing(costs: ModelCosts): string {
 * Accepts either a short name or full model name
 * Returns undefined if model is not found
 */
-export function getModelPricingString(model: string): string | undefined {
-  const shortName = getCanonicalName(model)
-  // Own-property guard: a proto-member id (`constructor`, `__proto__`) would
-  // otherwise return an inherited value and render "$NaN/$NaN per Mtok".
-  if (!Object.hasOwn(MODEL_COSTS, shortName)) return undefined
-  return formatModelPricing(MODEL_COSTS[shortName])
+export function getModelPricingString(
+  model: string,
+  usage: { speed?: Usage['speed'] } = {},
+): string | undefined {
+  const costs = getKnownModelCosts(model, usage)
+  return costs ? formatModelPricing(costs) : undefined
 }
@@ -0,0 +1,185 @@
+import { afterEach, beforeEach, expect, mock, test } from 'bun:test'
+import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'fs'
+import { tmpdir } from 'os'
+import { join } from 'path'
+import {
+  getAllowedSettingSources,
+  getFlagSettingsInline,
+  getFlagSettingsPath,
+  getOriginalCwd,
+  setAllowedSettingSources,
+  setFlagSettingsInline,
+  setFlagSettingsPath,
+  setOriginalCwd,
+} from '../../bootstrap/state.js'
+import {
+  acquireSharedMutationLock,
+  releaseSharedMutationLock,
+} from '../../test/sharedMutationLock.js'
+import { resetSettingsCache } from './settingsCache.js'
+import type { SettingsJson } from './types.js'
+
+// Capture the actual module once so source-order tests can replace just the
+// per-source reader without leaking a fake settings implementation.
+// @ts-expect-error -- query suffix intentionally bypasses Bun's module cache.
+import * as realSettings from './settings.js?modelPricingRealSettings'
+
+const paid = (inputTokens: number) => ({
+  inputTokens,
+  outputTokens: 2,
+  promptCacheReadTokens: 0.1,
+  promptCacheWriteTokens: 2.5,
+})
+
+let tempDir: string
+let originalCwd: string
+let originalSources: ReturnType<typeof getAllowedSettingSources>
+let originalFlagPath: string | undefined
+let originalFlagInline: Record<string, unknown> | null
+let settingsOverrideActive = false
+let settingsBySource: Partial<Record<string, SettingsJson>> = {}
+
+beforeEach(async () => {
+  await acquireSharedMutationLock('utils/settings/modelPricing.test.ts')
+  mock.restore()
+  tempDir = mkdtempSync(join(tmpdir(), 'openclaude-model-pricing-'))
+  originalCwd = getOriginalCwd()
+  originalSources = [...getAllowedSettingSources()]
+  originalFlagPath = getFlagSettingsPath()
+  originalFlagInline = getFlagSettingsInline()
+  setOriginalCwd(tempDir)
+  setFlagSettingsPath(undefined)
+  setFlagSettingsInline(null)
+  resetSettingsCache()
+  settingsOverrideActive = false
+  settingsBySource = {}
+  mock.module('./settings.js', () => ({
+    ...realSettings,
+    getSettingsForSource: (source: Parameters<typeof realSettings.getSettingsForSource>[0]) =>
+      settingsOverrideActive
+        ? settingsBySource[source] ?? null
+        : realSettings.getSettingsForSource(source),
+  }))
+})
+
+afterEach(() => {
+  try {
+    mock.restore()
+    setOriginalCwd(originalCwd)
+    setAllowedSettingSources(originalSources)
+    setFlagSettingsPath(originalFlagPath)
+    setFlagSettingsInline(originalFlagInline)
+    settingsOverrideActive = false
+    settingsBySource = {}
+    resetSettingsCache()
+    rmSync(tempDir, { recursive: true, force: true })
+  } finally {
+    releaseSharedMutationLock()
+  }
+})
+
+async function importFreshModelPricing() {
+  return import(`./modelPricing.js?ts=${Date.now()}-${Math.random()}`)
+}
+
+function writeJson(path: string, data: unknown): void {
+  writeFileSync(path, `${JSON.stringify(data)}\n`, 'utf8')
+}
+
+test('shared project settings cannot influence personal pricing', async () => {
+  const settingsDir = join(tempDir, '.openclaude')
+  mkdirSync(settingsDir, { recursive: true })
+  writeJson(join(settingsDir, 'settings.json'), {
+    modelPricing: { 'repo-controlled-model': paid(999) },
+  })
+  setAllowedSettingSources(['projectSettings'])
+  resetSettingsCache()
+
+  const { getModelPricingOverride } = await importFreshModelPricing()
+  expect(getModelPricingOverride('repo-controlled-model')).toBeUndefined()
+})
+
+test('canonical source order keeps managed pricing above CLI pricing', async () => {
+  settingsBySource = {
+    userSettings: { modelPricing: { model: paid(1) } },
+    flagSettings: { modelPricing: { model: paid(2) } },
+    policySettings: { modelPricing: { model: paid(3) } },
+  }
+  settingsOverrideActive = true
+  // policySettings and flagSettings remain enabled by policy even when the
+  // user restricts ordinary setting sources.
+  setAllowedSettingSources(['userSettings'])
+
+  const { getModelPricingOverride } = await importFreshModelPricing()
+  expect(getModelPricingOverride('model')?.inputTokens).toBe(3)
+})
+
+test('settings cache reset reloads exact pricing without a bespoke watcher', async () => {
+  const settingsPath = join(tempDir, 'flag-settings.json')
+  setFlagSettingsPath(settingsPath)
+  setAllowedSettingSources(['flagSettings'])
+  writeJson(settingsPath, {
+    modelPricing: { 'provider/model:v1?route=paid': paid(4) },
+  })
+  resetSettingsCache()
+
+  const { getModelPricingOverride } = await importFreshModelPricing()
+  const first = getModelPricingOverride('provider/model:v1?route=paid')
+  const firstClone = getModelPricingOverride('provider/model:v1?route=paid')
+  expect(first?.inputTokens).toBe(4)
+  expect(first?.webSearchRequests).toBe(0.01)
+  expect(Object.isFrozen(first)).toBe(true)
+  expect(firstClone).not.toBe(first)
+  expect(firstClone).toEqual(first)
+
+  writeJson(settingsPath, {
+    modelPricing: { 'provider/model:v1?route=paid': paid(8) },
+  })
+  expect(getModelPricingOverride('provider/model:v1?route=paid')?.inputTokens).toBe(4)
+
+  resetSettingsCache()
+  expect(getModelPricingOverride('provider/model:v1?route=paid')?.inputTokens).toBe(8)
+})
+
+test('prototype-like ids survive JSON parsing and require exact own-key lookup', async () => {
+  const settingsPath = join(tempDir, 'flag-settings.json')
+  const modelPricing = Object.create(null) as Record<string, unknown>
+  for (const model of ['constructor', 'toString', '__proto__']) {
+    modelPricing[model] = paid(6)
+  }
+  setFlagSettingsPath(settingsPath)
+  setAllowedSettingSources(['flagSettings'])
+  writeJson(settingsPath, { modelPricing })
+  resetSettingsCache()
+
+  const { getModelPricingOverride } = await importFreshModelPricing()
+  for (const model of ['constructor', 'toString', '__proto__']) {
+    expect(getModelPricingOverride(model)?.inputTokens).toBe(6)
+  }
+  expect(getModelPricingOverride('tostring')).toBeUndefined()
+})
+
+test('SDK flag settings replace complete entries without inheriting lower rates', async () => {
+  const settingsPath = join(tempDir, 'flag-settings.json')
+  const inlinePricing = Object.create(null) as Record<string, unknown>
+  inlinePricing.model = paid(8)
+  inlinePricing.__proto__ = paid(9)
+  setFlagSettingsPath(settingsPath)
+  setFlagSettingsInline({ modelPricing: inlinePricing })
+  setAllowedSettingSources(['flagSettings'])
+  writeJson(settingsPath, {
+    modelPricing: {
+      model: { ...paid(4), webSearchRequests: 7 },
+      'file-only-model': paid(5),
+    },
+  })
+  resetSettingsCache()
+
+  const { getModelPricingOverride } = await importFreshModelPricing()
+  expect(getModelPricingOverride('model')).toMatchObject({
+    inputTokens: 8,
+    webSearchRequests: 0.01,
+  })
+  expect(getModelPricingOverride('file-only-model')?.inputTokens).toBe(5)
+  expect(getModelPricingOverride('__proto__')?.inputTokens).toBe(9)
+})
@@ -0,0 +1,64 @@
+import {
+  isSettingSourceEnabled,
+  SETTING_SOURCES,
+  type SettingSource,
+} from './constants.js'
+import { getSettingsForSource } from './settings.js'
+
+export const DEFAULT_MODEL_WEB_SEARCH_RATE_USD_PER_REQUEST = 0.01
+
+export type ResolvedModelPricing = Readonly<{
+  inputTokens: number
+  outputTokens: number
+  promptCacheReadTokens: number
+  promptCacheWriteTokens: number
+  webSearchRequests: number
+}>
+
+const TRUSTED_MODEL_PRICING_SOURCES = new Set<SettingSource>([
+  'userSettings',
+  'localSettings',
+  'flagSettings',
+  'policySettings',
+])
+
+/**
+ * Resolve a price for the exact model id used by the API request.
+ *
+ * Source precedence follows SETTING_SOURCES (later wins), but deliberately
+ * excludes shared project settings so repository content cannot change a
+ * user's personal USD accounting. The settings loader owns caching and reload
+ * invalidation; this function keeps no additional state.
+ */
+export function getModelPricingOverride(
+  model: string,
+): ResolvedModelPricing | undefined {
+  let resolved: ResolvedModelPricing | undefined
+
+  for (const source of SETTING_SOURCES) {
+    if (
+      !TRUSTED_MODEL_PRICING_SOURCES.has(source) ||
+      !isSettingSourceEnabled(source)
+    ) {
+      continue
+    }
+
+    const pricing = getSettingsForSource(source)?.modelPricing
+    if (!pricing || !Object.hasOwn(pricing, model)) {
+      continue
+    }
+
+    const entry = pricing[model]
+    resolved = Object.freeze({
+      inputTokens: entry.inputTokens,
+      outputTokens: entry.outputTokens,
+      promptCacheReadTokens: entry.promptCacheReadTokens,
+      promptCacheWriteTokens: entry.promptCacheWriteTokens,
+      webSearchRequests:
+        entry.webSearchRequests ??
+        DEFAULT_MODEL_WEB_SEARCH_RATE_USD_PER_REQUEST,
+    })
+  }
+
+  return resolved
+}
@@ -0,0 +1,140 @@
+import { expect, test } from 'bun:test'
+import { mkdtempSync, rmSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { generateSettingsJSONSchema } from './schemaOutput.js'
+import { parseSettingsFile } from './settings.js'
+import { SettingsSchema } from './types.js'
+
+const completePrice = {
+  inputTokens: 1,
+  outputTokens: 2,
+  promptCacheReadTokens: 0.1,
+  promptCacheWriteTokens: 2.5,
+}
+
+test('modelPricing accepts explicit zero and defaults web-search pricing later', () => {
+  const result = SettingsSchema().safeParse({
+    modelPricing: {
+      'nvidia/free-model': {
+        inputTokens: 0,
+        outputTokens: 0,
+        promptCacheReadTokens: 0,
+        promptCacheWriteTokens: 0,
+        webSearchRequests: 0,
+      },
+      'provider/paid:model?variant=1': completePrice,
+    },
+  })
+
+  expect(result.success).toBe(true)
+  if (result.success) {
+    expect(result.data.modelPricing).toEqual({
+      'nvidia/free-model': {
+        inputTokens: 0,
+        outputTokens: 0,
+        promptCacheReadTokens: 0,
+        promptCacheWriteTokens: 0,
+        webSearchRequests: 0,
+      },
+      'provider/paid:model?variant=1': completePrice,
+    })
+  }
+})
+
+test('invalid modelPricing is dropped without invalidating unrelated settings', () => {
+  const tooManyEntries = Object.fromEntries(
+    Array.from({ length: 257 }, (_, i) => [`model-${i}`, completePrice]),
+  )
+  const invalidValues: unknown[] = [
+    'not-an-object',
+    { model: 'not-an-entry-object' },
+    { model: { ...completePrice, unknownField: 1 } },
+    { model: { ...completePrice, inputTokens: -1 } },
+    { model: { ...completePrice, outputTokens: Number.POSITIVE_INFINITY } },
+    { model: { ...completePrice, outputTokens: Number.NaN } },
+    { model: { ...completePrice, inputTokens: 100_001 } },
+    { model: { ...completePrice, webSearchRequests: 1_001 } },
+    { model: { inputTokens: 1, outputTokens: 2 } },
+    { '': completePrice },
+    { ['x'.repeat(513)]: completePrice },
+    tooManyEntries,
+  ]
+
+  for (const modelPricing of invalidValues) {
+    const result = SettingsSchema().safeParse({
+      model: 'sonnet',
+      modelPricing,
+    })
+    expect(result.success).toBe(true)
+    if (result.success) {
+      expect(result.data.model).toBe('sonnet')
+      expect(result.data.modelPricing).toBeUndefined()
+    }
+  }
+})
+
+test('parseSettingsFile reports the raw modelPricing value that it drops', () => {
+  const fixtureDir = mkdtempSync(join(tmpdir(), 'model-pricing-diagnostic-'))
+  const settingsPath = join(fixtureDir, 'settings.json')
+  const modelPricing = {
+    model: { ...completePrice, inputTokens: -1 },
+  }
+  writeFileSync(
+    settingsPath,
+    JSON.stringify({ model: 'sonnet', modelPricing }),
+    'utf8',
+  )
+
+  try {
+    const result = parseSettingsFile(settingsPath)
+
+    expect(result.settings?.model).toBe('sonnet')
+    expect(result.settings?.modelPricing).toBeUndefined()
+    expect(result.errors).toEqual([
+      {
+        file: settingsPath,
+        path: 'modelPricing',
+        message: 'Invalid modelPricing value was ignored',
+        expected: undefined,
+        invalidValue: modelPricing,
+        suggestion: undefined,
+        docLink: undefined,
+      },
+    ])
+  } finally {
+    rmSync(fixtureDir, { recursive: true, force: true })
+  }
+})
+
+test('generated settings schema preserves modelPricing constraints', () => {
+  const schema = JSON.parse(generateSettingsJSONSchema()) as {
+    properties: {
+      modelPricing: {
+        additionalProperties: {
+          properties: Record<string, unknown>
+          required: string[]
+        }
+        maxProperties: number
+        propertyNames: { type: string; maxLength: number; minLength: number }
+      }
+    }
+  }
+  const pricing = schema.properties.modelPricing
+
+  expect(pricing.maxProperties).toBe(256)
+  expect(pricing.propertyNames).toEqual({
+    type: 'string',
+    minLength: 1,
+    maxLength: 512,
+  })
+  expect(pricing.additionalProperties.required).toEqual([
+    'inputTokens',
+    'outputTokens',
+    'promptCacheReadTokens',
+    'promptCacheWriteTokens',
+  ])
+  expect(pricing.additionalProperties.properties).toHaveProperty(
+    'webSearchRequests',
+  )
+})
@@ -46,6 +46,7 @@ import {
 } from './settingsCache.js'
 import { type SettingsJson, SettingsSchema } from './types.js'
 import {
+  filterInvalidModelPricing,
  filterInvalidPermissionRules,
  formatZodError,
  type SettingsWithErrors,
@@ -215,15 +216,22 @@ function parseSettingsFileUncached(path: string): {
    // Filter invalid permission rules before schema validation so one bad
    // rule doesn't cause the entire settings file to be rejected.
    const ruleWarnings = filterInvalidPermissionRules(data, path)
+    const modelPricingWarnings = filterInvalidModelPricing(data, path)

    const result = SettingsSchema().safeParse(data)

    if (!result.success) {
      const errors = formatZodError(result.error, path)
-      return { settings: null, errors: [...ruleWarnings, ...errors] }
+      return {
+        settings: null,
+        errors: [...ruleWarnings, ...modelPricingWarnings, ...errors],
+      }
    }

-    return { settings: result.data, errors: ruleWarnings }
+    return {
+      settings: result.data,
+      errors: [...ruleWarnings, ...modelPricingWarnings],
+    }
  } catch (error) {
    handleFileSystemError(error, path)
    return { settings: null, errors: [] }
@@ -524,7 +532,7 @@ export function updateSettingsForSource(
 }

 /**
- * Custom merge function for arrays - concatenate and deduplicate
+ * Custom merge function for structured settings values.
 */
 function mergeArrays<T>(targetArray: T[], sourceArray: T[]): T[] {
  return uniq([...targetArray, ...sourceArray])
@@ -532,16 +540,45 @@ function mergeArrays<T>(targetArray: T[], sourceArray: T[]): T[] {

 /**
 * Custom merge function for lodash mergeWith when merging settings.
- * Arrays are concatenated and deduplicated; other values use default lodash merge behavior.
+ * Arrays are concatenated and deduplicated. Model-pricing maps preserve
+ * arbitrary exact keys and replace complete entries. Other values use default
+ * lodash merge behavior.
 * Exported for testing.
 */
 export function settingsMergeCustomizer(
  objValue: unknown,
  srcValue: unknown,
+  key?: PropertyKey,
 ): unknown {
  if (Array.isArray(objValue) && Array.isArray(srcValue)) {
    return mergeArrays(objValue, srcValue)
  }
+  if (
+    key === 'modelPricing' &&
+    typeof srcValue === 'object' &&
+    srcValue !== null &&
+    !Array.isArray(srcValue)
+  ) {
+    // Model ids are arbitrary strings, including Object.prototype names.
+    // Copy into a null-prototype map so mergeWith cannot interpret them as
+    // structure. Replace each complete higher-priority entry atomically: a
+    // missing optional webSearchRequests field must use its documented
+    // default, not inherit a lower-priority source's value.
+    const merged = Object.create(null) as Record<string, unknown>
+    if (
+      typeof objValue === 'object' &&
+      objValue !== null &&
+      !Array.isArray(objValue)
+    ) {
+      for (const [model, entry] of Object.entries(objValue)) {
+        merged[model] = entry
+      }
+    }
+    for (const [model, entry] of Object.entries(srcValue)) {
+      merged[model] = entry
+    }
+    return merged
+  }
  // Return undefined to let lodash handle default merge behavior
  return undefined
 }
@@ -1,5 +1,5 @@
 import { feature } from 'bun:bundle'
-import { z } from 'zod/v4'
+import { toJSONSchema, z } from 'zod/v4'
 import { SandboxSettingsSchema } from '../../entrypoints/sandboxTypes.js'
 import { isEnvTruthy } from '../envUtils.js'
 import { lazySchema } from '../lazySchema.js'
@@ -30,6 +30,103 @@ import { type HookCommand, HooksSchema } from '../../schemas/hooks.js'
 import { AutoFixConfigSchema } from '../../services/autoFix/autoFixConfig.js'
 import { count } from '../array.js'

+export const MAX_MODEL_PRICING_ENTRIES = 256
+export const MAX_MODEL_PRICING_ID_LENGTH = 512
+export const MAX_MODEL_TOKEN_RATE_USD_PER_MILLION = 100_000
+export const MAX_MODEL_WEB_SEARCH_RATE_USD_PER_REQUEST = 1_000
+
+const ModelTokenRateSchema = z
+  .number()
+  .finite()
+  .nonnegative()
+  .max(MAX_MODEL_TOKEN_RATE_USD_PER_MILLION)
+
+export const ModelPricingEntrySchema = z
+  .object({
+    inputTokens: ModelTokenRateSchema.describe(
+      'Input-token price in USD per 1,000,000 tokens.',
+    ),
+    outputTokens: ModelTokenRateSchema.describe(
+      'Output-token price in USD per 1,000,000 tokens.',
+    ),
+    promptCacheReadTokens: ModelTokenRateSchema.describe(
+      'Prompt-cache read price in USD per 1,000,000 tokens.',
+    ),
+    promptCacheWriteTokens: ModelTokenRateSchema.describe(
+      'Prompt-cache write price in USD per 1,000,000 tokens.',
+    ),
+    webSearchRequests: z
+      .number()
+      .finite()
+      .nonnegative()
+      .max(MAX_MODEL_WEB_SEARCH_RATE_USD_PER_REQUEST)
+      .optional()
+      .describe(
+        'Web-search price in USD per request. Defaults to $0.01 when omitted.',
+      ),
+  })
+  .strict()
+
+const MODEL_PRICING_KEY_PREFIX = '\0'
+
+const ModelPricingJSONSchema = z.record(
+  z.string().min(1).max(MAX_MODEL_PRICING_ID_LENGTH),
+  ModelPricingEntrySchema,
+)
+
+const ModelPricingSchema = z.preprocess(
+  value => {
+    if (typeof value !== 'object' || value === null || Array.isArray(value)) {
+      return value
+    }
+    // Zod's record parser treats Object.prototype names as structural record
+    // members. Prefix every raw own key before validation so arbitrary exact
+    // model ids such as `constructor`, `toString`, and `__proto__` stay data.
+    const entries = Object.entries(value)
+    if (
+      entries.length > MAX_MODEL_PRICING_ENTRIES ||
+      entries.some(
+        ([model]) =>
+          model.length < 1 || model.length > MAX_MODEL_PRICING_ID_LENGTH,
+      )
+    ) {
+      return null
+    }
+    const encoded: Record<string, unknown> = Object.create(null) as Record<
+      string,
+      unknown
+    >
+    for (const [model, entry] of entries) {
+      encoded[MODEL_PRICING_KEY_PREFIX + model] = entry
+    }
+    return encoded
+  },
+  z
+    .record(z.string(), ModelPricingEntrySchema)
+    .transform(value => {
+      const decoded = Object.create(null) as Record<
+        string,
+        z.infer<typeof ModelPricingEntrySchema>
+      >
+      for (const [encodedModel, entry] of Object.entries(value)) {
+        decoded[encodedModel.slice(MODEL_PRICING_KEY_PREFIX.length)] = entry
+      }
+      return decoded
+    }),
+)
+
+export const ModelPricingDiagnosticSchema = z.unknown().superRefine(
+  (value, context) => {
+    if (!ModelPricingSchema.safeParse(value).success) {
+      context.addIssue({
+        code: 'custom',
+        message: 'Invalid modelPricing value was ignored',
+        params: { received: value },
+      })
+    }
+  },
+)
+
 /**
 * Schema for environment variables
 */
@@ -115,6 +212,19 @@ export const ExtraKnownMarketplaceSchema = lazySchema(() =>
  }),
 )

+// The prefix transform above is deliberately invisible to settings authors,
+// so expose the equivalent public input contract to JSON Schema consumers.
+// Zod documents this hook for schemas whose runtime representation needs a
+// custom JSON Schema (the preprocess itself is otherwise unrepresentable).
+ModelPricingSchema._zod.toJSONSchema = () => {
+  const schema = toJSONSchema(ModelPricingJSONSchema, {
+    unrepresentable: 'any',
+  }) as Record<string, unknown>
+  delete schema.$schema
+  schema.maxProperties = MAX_MODEL_PRICING_ENTRIES
+  return schema
+}
+
 /**
 * Schema for allowed MCP server entry in enterprise allowlist.
 * Supports matching by serverName, serverCommand, or serverUrl (mutually exclusive).
@@ -867,6 +977,17 @@ export const SettingsSchema = lazySchema(() =>
            'CLAUDE_CODE_OPENAI_CONTEXT_WINDOWS / CLAUDE_CODE_OPENAI_MAX_OUTPUT_TOKENS env vars take precedence. ' +
            'Example: { "qwen3.6-plus": { "contextWindow": 1048576, "maxOutputTokens": 32768 } }',
        ),
+      modelPricing: ModelPricingSchema.optional()
+        // One malformed pricing map must not invalidate unrelated settings in
+        // the same source. Drop the whole field so callers never combine a
+        // partially-trusted price with built-in cache rates.
+        .catch(undefined)
+        .describe(
+          'Exact-model pricing overrides. Token fields are USD per 1,000,000 tokens; ' +
+            'webSearchRequests is USD per request and defaults to $0.01 when omitted. ' +
+            'All four token fields are required. Keys are exact, case-sensitive resolved model ids. ' +
+            'Shared project settings are ignored for personal USD accounting.',
+        ),
      fastMode: z
        .boolean()
        .optional()
@@ -1306,6 +1427,7 @@ export type DeniedMcpServerEntry = z.infer<
  ReturnType<typeof DeniedMcpServerEntrySchema>
 >
 export type SettingsJson = z.infer<ReturnType<typeof SettingsSchema>>
+export type ModelPricingEntry = z.infer<typeof ModelPricingEntrySchema>

 /**
 * Type guard for MCP server entry with serverName
@@ -1,11 +1,14 @@
 import type { ConfigScope } from 'src/services/mcp/types.js'
-import type { ZodError, ZodIssue } from 'zod/v4'
+import { type ZodError, type ZodIssue, z } from 'zod/v4'
 import { jsonParse } from '../slowOperations.js'
 import { plural } from '../stringUtils.js'
 import { validatePermissionRule } from './permissionValidation.js'
 import { generateSettingsJSONSchema } from './schemaOutput.js'
 import type { SettingsJson } from './types.js'
-import { SettingsSchema } from './types.js'
+import {
+  ModelPricingDiagnosticSchema,
+  SettingsSchema,
+} from './types.js'
 import { getValidationTip } from './validationTips.js'

 /**
@@ -263,3 +266,23 @@ export function filterInvalidPermissionRules(
  }
  return warnings
 }
+
+export function filterInvalidModelPricing(
+  data: unknown,
+  filePath: string,
+): ValidationError[] {
+  if (
+    !data ||
+    typeof data !== 'object' ||
+    !Object.hasOwn(data, 'modelPricing')
+  ) {
+    return []
+  }
+
+  const result = z
+    .object({ modelPricing: ModelPricingDiagnosticSchema })
+    .safeParse({
+      modelPricing: (data as Record<string, unknown>).modelPricing,
+    })
+  return result.success ? [] : formatZodError(result.error, filePath)
+}