fix(model): default NVIDIA NIM main loop model

src/utils/model/model.openai-shim-providers.test.ts

@@ -306,6 +306,29 @@ test('getDefaultMainLoopModelSetting defaults MiniMax to M3', async () => {
   expect(getDefaultMainLoopModel()).toBe('MiniMax-M3')
 })
 
+test('getDefaultMainLoopModelSetting uses the NVIDIA NIM route model', async () => {
+  process.env.NVIDIA_NIM = '1'
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  process.env.OPENAI_MODEL = 'meta/llama-3.3-70b-instruct'
+
+  const {
+    getDefaultMainLoopModel,
+    getDefaultMainLoopModelSetting,
+  } = await importFreshModelModule()
+  expect(getDefaultMainLoopModelSetting()).toBe('meta/llama-3.3-70b-instruct')
+  expect(getDefaultMainLoopModel()).toBe('meta/llama-3.3-70b-instruct')
+})
+
+test('getDefaultMainLoopModelSetting falls back to the NVIDIA NIM descriptor default', async () => {
+  process.env.NVIDIA_NIM = '1'
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+
+  const { getDefaultMainLoopModelSetting } = await importFreshModelModule()
+  expect(getDefaultMainLoopModelSetting()).toBe(
+    'nvidia/llama-3.1-nemotron-70b-instruct',
+  )
+})
+
 test('getDefaultMainLoopModelSetting defaults Xiaomi MiMo to mimo-v2.5-pro', async () => {
   process.env.MIMO_API_KEY = 'mimo-test'
   process.env.CLAUDE_CODE_USE_OPENAI = '1'

src/utils/model/model.ts

@@ -30,6 +30,7 @@ import { type ModelAlias, isModelAlias } from './aliases.js'
 import { capitalize } from '../stringUtils.js'
 import { DEFAULT_GEMINI_MODEL } from '../providerProfile.js'
 import { getAntModelOverrideConfig, resolveAntModel } from './antModels.js'
+import { getRouteDefaultModel } from '../../integrations/routeMetadata.js'
 
 export type ModelShortName = string
 export type ModelName = string
@@ -388,6 +389,15 @@ export function getDefaultMainLoopModelSetting(): ModelName | ModelAlias {
   if (getAPIProvider() === 'codex') {
     return process.env.OPENAI_MODEL || 'gpt-5.5'
   }
+  // NVIDIA NIM uses OpenAI-compatible model ids. Keep this fallback aligned
+  // with the route descriptor so headless sessions never send a Claude model.
+  if (getAPIProvider() === 'nvidia-nim') {
+    return (
+      process.env.OPENAI_MODEL ||
+      getRouteDefaultModel('nvidia-nim') ||
+      'nvidia/llama-3.1-nemotron-70b-instruct'
+    )
+  }
   // xAI provider: always use the configured Grok model (default grok-4.3)
   if (getAPIProvider() === 'xai') {
     return process.env.OPENAI_MODEL || 'grok-4.3'

src/utils/modelCost.modelGate.test.ts

@@ -22,6 +22,29 @@ afterEach(() => {
   }
 })
 
+test('unknown models do not inherit the configured default model price', async () => {
+  mock.module('./model/model.js', () => ({
+    firstPartyNameToCanonical: (model: string) => {
+      if (model.includes('claude-haiku-4-5')) return 'claude-haiku-4-5'
+      return model
+    },
+    getCanonicalName: (model: string) => {
+      if (model.includes('claude-haiku-4-5')) return 'claude-haiku-4-5'
+      return model
+    },
+    getDefaultMainLoopModelSetting: () => 'claude-haiku-4-5',
+  }))
+  const { getModelCosts, COST_HAIKU_45, COST_TIER_5_25 } =
+    await importFreshModelCost()
+
+  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+  const usage = {} as any
+  const costs = getModelCosts('meta/llama-3.3-70b-instruct', usage)
+
+  expect(costs).toEqual(COST_TIER_5_25)
+  expect(costs).not.toEqual(COST_HAIKU_45)
+})
+
 // Regression for #1769: fast mode is now enabled for Opus 4.8, but getModelCosts
 // only applied the elevated fast-mode tier to opus-4-6, so fast-mode 4.8 was
 // billed at the normal rate while the picker advertised the fast-mode price.

src/utils/modelCost.ts

@@ -21,7 +21,6 @@ import {
 import {
   firstPartyNameToCanonical,
   getCanonicalName,
-  getDefaultMainLoopModelSetting,
   type ModelShortName,
 } from './model/model.js'
 
@@ -166,10 +165,7 @@ export function getModelCosts(model: string, usage: Usage): ModelCosts {
   const costs = MODEL_COSTS[shortName]
   if (!costs) {
     trackUnknownModelCost(model, shortName)
-    return (
-      MODEL_COSTS[getCanonicalName(getDefaultMainLoopModelSetting())] ??
-      DEFAULT_UNKNOWN_MODEL_COST
-    )
+    return DEFAULT_UNKNOWN_MODEL_COST
   }
   return costs
 }
@@ -184,7 +180,8 @@ function trackUnknownModelCost(model: string, shortName: ModelShortName): void {
 }
 
 // Calculate the cost of a query in US dollars.
-// If the model's costs are not found, use the default model's costs.
+// Unknown models use the explicit unknown-model estimate and are marked in
+// session state; they must never inherit an unrelated configured default.
 export function calculateUSDCost(resolvedModel: string, usage: Usage): number {
   const modelCosts = getModelCosts(resolvedModel, usage)
   return tokensToUSDCost(modelCosts, usage)