diff --git a/src/services/api/claude.ts b/src/services/api/claude.ts
index c71d0103..93ae1502 100644
--- a/src/services/api/claude.ts
+++ b/src/services/api/claude.ts
@@ -243,6 +243,7 @@ import {
   logAPIError,
   logAPIQuery,
   logAPISuccessAndDuration,
+  normalizeUsage,
   type NonNullableUsage,
 } from "./logging.js";
 import {
@@ -2635,6 +2636,7 @@ async function* queryModel(
       const m: AssistantMessage = {
         message: {
           ...result,
+          usage: normalizeUsage(result.usage),
           content: normalizeContentFromAPI(
             result.content,
             tools,
@@ -2732,6 +2734,7 @@ async function* queryModel(
         const m: AssistantMessage = {
           message: {
             ...result,
+            usage: normalizeUsage(result.usage),
             content: normalizeContentFromAPI(
               result.content,
               tools,
@@ -2882,8 +2885,9 @@ async function* queryModel(
     // message_delta handler before any yield. Fallback pushes to newMessages
     // then yields, so tracking must be here to survive .return() at the yield.
     if (fallbackMessage) {
-      const fallbackUsage = fallbackMessage.message.usage;
-      usage = updateUsage(EMPTY_USAGE, fallbackUsage);
+      const fallbackUsage = normalizeUsage(fallbackMessage.message.usage);
+      fallbackMessage.message.usage = fallbackUsage;
+      usage = fallbackUsage;
       stopReason = fallbackMessage.message.stop_reason;
       const fallbackCost = calculateUSDCost(resolvedModel, fallbackUsage);
       costUSD += addToTotalSessionCost(
diff --git a/src/services/api/emptyUsage.test.ts b/src/services/api/emptyUsage.test.ts
new file mode 100644
index 00000000..144548ad
--- /dev/null
+++ b/src/services/api/emptyUsage.test.ts
@@ -0,0 +1,30 @@
+import { describe, expect, test } from 'bun:test'
+import { calculateUSDCost } from '../../utils/modelCost.js'
+import { EMPTY_USAGE, normalizeUsage } from './emptyUsage.js'
+
+describe('normalizeUsage', () => {
+  test('fills missing provider usage with zero-valued usage', () => {
+    expect(normalizeUsage(undefined)).toEqual(EMPTY_USAGE)
+  })
+
+  test('preserves reported token fields while defaulting missing nested fields', () => {
+    expect(
+      normalizeUsage({
+        input_tokens: 12,
+        output_tokens: 5,
+        cache_read_input_tokens: 3,
+      }),
+    ).toEqual({
+      ...EMPTY_USAGE,
+      input_tokens: 12,
+      output_tokens: 5,
+      cache_read_input_tokens: 3,
+    })
+  })
+
+  test('keeps cost calculation safe for provider responses without usage', () => {
+    expect(() =>
+      calculateUSDCost('anthropic/claude-opus-4.7', normalizeUsage(undefined)),
+    ).not.toThrow()
+  })
+})
diff --git a/src/services/api/emptyUsage.ts b/src/services/api/emptyUsage.ts
index ad8c25ff..a1b16a3b 100644
--- a/src/services/api/emptyUsage.ts
+++ b/src/services/api/emptyUsage.ts
@@ -1,3 +1,4 @@
+import type { BetaUsage } from '@anthropic-ai/sdk/resources/beta/messages/messages.mjs'
 import type { NonNullableUsage } from '../../entrypoints/sdk/sdkUtilityTypes.js'
 
 /**
@@ -20,3 +21,42 @@ export const EMPTY_USAGE: Readonly<NonNullableUsage> = {
   iterations: [],
   speed: 'standard',
 }
+
+/**
+ * Some Anthropic-compatible providers omit usage in successful fallback
+ * responses. Normalize at the provider boundary so transcript writes and cost
+ * tracking can stay total-order safe instead of crashing on missing token fields.
+ */
+export function normalizeUsage(
+  usage: Partial<BetaUsage> | null | undefined,
+): NonNullableUsage {
+  return {
+    input_tokens: usage?.input_tokens ?? EMPTY_USAGE.input_tokens,
+    cache_creation_input_tokens:
+      usage?.cache_creation_input_tokens ??
+      EMPTY_USAGE.cache_creation_input_tokens,
+    cache_read_input_tokens:
+      usage?.cache_read_input_tokens ?? EMPTY_USAGE.cache_read_input_tokens,
+    output_tokens: usage?.output_tokens ?? EMPTY_USAGE.output_tokens,
+    server_tool_use: {
+      web_search_requests:
+        usage?.server_tool_use?.web_search_requests ??
+        EMPTY_USAGE.server_tool_use.web_search_requests,
+      web_fetch_requests:
+        usage?.server_tool_use?.web_fetch_requests ??
+        EMPTY_USAGE.server_tool_use.web_fetch_requests,
+    },
+    service_tier: usage?.service_tier ?? EMPTY_USAGE.service_tier,
+    cache_creation: {
+      ephemeral_1h_input_tokens:
+        usage?.cache_creation?.ephemeral_1h_input_tokens ??
+        EMPTY_USAGE.cache_creation.ephemeral_1h_input_tokens,
+      ephemeral_5m_input_tokens:
+        usage?.cache_creation?.ephemeral_5m_input_tokens ??
+        EMPTY_USAGE.cache_creation.ephemeral_5m_input_tokens,
+    },
+    inference_geo: usage?.inference_geo ?? EMPTY_USAGE.inference_geo,
+    iterations: usage?.iterations ?? EMPTY_USAGE.iterations,
+    speed: usage?.speed ?? EMPTY_USAGE.speed,
+  }
+}
diff --git a/src/services/api/logging.ts b/src/services/api/logging.ts
index a411c12c..769d4b9d 100644
--- a/src/services/api/logging.ts
+++ b/src/services/api/logging.ts
@@ -35,12 +35,12 @@ import {
   logEvent,
 } from '../analytics/index.js'
 import { sanitizeToolNameForAnalytics } from '../analytics/metadata.js'
-import { EMPTY_USAGE } from './emptyUsage.js'
+import { EMPTY_USAGE, normalizeUsage } from './emptyUsage.js'
 import { classifyAPIError } from './errors.js'
 import { extractConnectionErrorDetails } from './errorUtils.js'
 
 export type { NonNullableUsage }
-export { EMPTY_USAGE }
+export { EMPTY_USAGE, normalizeUsage }
 
 // Strategy used for global prompt caching
 export type GlobalCacheStrategy = 'tool_based' | 'system_prompt' | 'none'
diff --git a/src/utils/sideQuery.ts b/src/utils/sideQuery.ts
index e01e5b26..4c560286 100644
--- a/src/utils/sideQuery.ts
+++ b/src/utils/sideQuery.ts
@@ -14,6 +14,7 @@ import { logEvent } from '../services/analytics/index.js'
 import type { AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS } from '../services/analytics/metadata.js'
 import { getAPIMetadata } from '../services/api/claude.js'
 import { getAnthropicClient } from '../services/api/client.js'
+import { normalizeUsage } from '../services/api/emptyUsage.js'
 import { getModelBetas, modelSupportsStructuredOutputs } from './betas.js'
 import { computeFingerprint } from './fingerprint.js'
 import { normalizeModelStringForAPI } from './model/model.js'
@@ -192,6 +193,7 @@ export async function sideQuery(opts: SideQueryOptions): Promise<BetaMessage> {
 
   const requestId =
     (response as { _request_id?: string | null })._request_id ?? undefined
+  response.usage = normalizeUsage(response.usage)
   const now = Date.now()
   const lastCompletion = getLastApiCompletionTimestamp()
   logEvent('tengu_api_success', {