diff --git a/docs/docs/reference/telemetry.md b/docs/docs/reference/telemetry.md
index c3b72cc6d..4535ae72b 100644
--- a/docs/docs/reference/telemetry.md
+++ b/docs/docs/reference/telemetry.md
@@ -11,7 +11,7 @@ We collect the following categories of events:
 | `session_start` | A new CLI session begins |
 | `session_end` | A CLI session ends (includes duration) |
 | `session_forked` | A session is forked from an existing one |
-| `generation` | An AI model generation completes (model ID, token counts, duration — no prompt content) |
+| `generation` | An AI model generation (step) completes — model ID, provider ID, agent, finish reason, cost, duration, and token breakdown: input, output, and when available: reasoning tokens (reasoning models only), cache-read tokens (prompt cache hit), cache-write tokens (new cache entry). No prompt content. |
 | `tool_call` | A tool is invoked (tool name and category — no arguments or output) |
 | `native_call` | A native engine call completes (method name and duration — no arguments) |
 | `command` | A CLI command is executed (command name only) |
diff --git a/packages/opencode/src/altimate/telemetry/index.ts b/packages/opencode/src/altimate/telemetry/index.ts
index e0f0130e4..ded93596c 100644
--- a/packages/opencode/src/altimate/telemetry/index.ts
+++ b/packages/opencode/src/altimate/telemetry/index.ts
@@ -14,13 +14,6 @@ export namespace Telemetry {
   const MAX_BUFFER_SIZE = 200
   const REQUEST_TIMEOUT_MS = 10_000
 
-  export type TokensPayload = {
-    input: number
-    output: number
-    reasoning: number
-    cache_read: number
-    cache_write: number
-  }
 
   export type Event =
     | {
@@ -50,9 +43,15 @@ export namespace Telemetry {
         provider_id: string
         agent: string
         finish_reason: string
-        tokens: TokensPayload
         cost: number
         duration_ms: number
+        // Flat token fields — only present when data is available from the provider.
+        // No nested objects: Azure App Insights custom measures must be top-level numbers.
+        tokens_input: number
+        tokens_output: number
+        tokens_reasoning?: number    // only for reasoning models
+        tokens_cache_read?: number   // only when a cached prompt was reused
+        tokens_cache_write?: number  // only when a new cache entry was written
       }
     | {
         type: "tool_call"
diff --git a/packages/opencode/src/session/processor.ts b/packages/opencode/src/session/processor.ts
index 70ea586b8..2fd644cfe 100644
--- a/packages/opencode/src/session/processor.ts
+++ b/packages/opencode/src/session/processor.ts
@@ -16,6 +16,9 @@ import { PermissionNext } from "@/permission/next"
 import { Question } from "@/question"
 import { PartID } from "./schema"
 import type { SessionID, MessageID } from "./schema"
+// altimate_change start — import Telemetry for per-generation token tracking
+import { Telemetry } from "@/altimate/telemetry"
+// altimate_change end
 
 export namespace SessionProcessor {
   const DOOM_LOOP_THRESHOLD = 3
@@ -47,6 +50,9 @@ export namespace SessionProcessor {
         log.info("process")
         needsCompaction = false
         const shouldBreak = (await Config.get()).experimental?.continue_loop_on_deny !== true
+        // altimate_change start — track step start time for generation telemetry duration_ms
+        let stepStartTime = Date.now()
+        // altimate_change end
         while (true) {
           try {
             let currentText: MessageV2.TextPart | undefined
@@ -240,6 +246,9 @@ export namespace SessionProcessor {
                     snapshot,
                     type: "step-start",
                   })
+                  // altimate_change start — record step start time for generation telemetry duration
+                  stepStartTime = Date.now()
+                  // altimate_change end
                   break
 
                 case "finish-step":
@@ -251,6 +260,26 @@ export namespace SessionProcessor {
                   input.assistantMessage.finish = value.finishReason
                   input.assistantMessage.cost += usage.cost
                   input.assistantMessage.tokens = usage.tokens
+                  // altimate_change start — emit per-generation telemetry with token breakdown
+                  // Only include token fields that are actually provided by the API (never default to 0).
+                  Telemetry.track({
+                    type: "generation",
+                    timestamp: Date.now(),
+                    session_id: input.sessionID,
+                    message_id: input.assistantMessage.id,
+                    model_id: streamInput.model.id,
+                    provider_id: streamInput.model.providerID,
+                    agent: streamInput.agent.name,
+                    finish_reason: value.finishReason,
+                    cost: usage.cost,
+                    duration_ms: Date.now() - stepStartTime,
+                    tokens_input: usage.tokens.input,
+                    tokens_output: usage.tokens.output,
+                    ...(value.usage.reasoningTokens !== undefined && { tokens_reasoning: usage.tokens.reasoning }),
+                    ...(value.usage.cachedInputTokens !== undefined && { tokens_cache_read: usage.tokens.cache.read }),
+                    ...(usage.tokens.cache.write > 0 && { tokens_cache_write: usage.tokens.cache.write }),
+                  })
+                  // altimate_change end
                   await Session.updatePart({
                     id: PartID.ascending(),
                     reason: value.finishReason,
diff --git a/packages/opencode/test/session/processor.test.ts b/packages/opencode/test/session/processor.test.ts
index 857966a5a..fd12e483d 100644
--- a/packages/opencode/test/session/processor.test.ts
+++ b/packages/opencode/test/session/processor.test.ts
@@ -468,20 +468,18 @@ describe("generation telemetry", () => {
       provider_id: "anthropic",
       agent: "builder",
       finish_reason: "end_turn",
-      tokens: {
-        input: 1000,
-        output: 500,
-        reasoning: 200,
-        cache_read: 800,
-        cache_write: 100,
-      },
+      tokens_input: 1000,
+      tokens_output: 500,
+      tokens_reasoning: 200,
+      tokens_cache_read: 800,
+      tokens_cache_write: 100,
       cost: 0.05,
       duration_ms: 3000,
     }
 
     expect(event.model_id).toBe("claude-opus-4-6")
-    expect(event.tokens.input).toBe(1000)
-    expect(event.tokens.cache_read).toBe(800)
+    expect(event.tokens_input).toBe(1000)
+    expect(event.tokens_cache_read).toBe(800)
     expect(event.cost).toBe(0.05)
     expect(event.finish_reason).toBe("end_turn")
   })
diff --git a/packages/opencode/test/telemetry/telemetry.test.ts b/packages/opencode/test/telemetry/telemetry.test.ts
index c7c575736..b8f4b1fb1 100644
--- a/packages/opencode/test/telemetry/telemetry.test.ts
+++ b/packages/opencode/test/telemetry/telemetry.test.ts
@@ -624,7 +624,7 @@ describe("telemetry.toAppInsightsEnvelopes (indirect)", () => {
     }
   })
 
-  test("nested tokens object is flattened with tokens_ prefix", async () => {
+  test("flat token fields appear in measurements", async () => {
     const { fetchCalls, cleanup } = await initWithMockedFetch()
     try {
       Telemetry.track({
@@ -636,13 +636,11 @@ describe("telemetry.toAppInsightsEnvelopes (indirect)", () => {
         provider_id: "anthropic",
         agent: "builder",
         finish_reason: "end_turn",
-        tokens: {
-          input: 100,
-          output: 200,
-          reasoning: 50,
-          cache_read: 10,
-          cache_write: 5,
-        },
+        tokens_input: 100,
+        tokens_output: 200,
+        tokens_reasoning: 50,
+        tokens_cache_read: 10,
+        tokens_cache_write: 5,
         cost: 0.01,
         duration_ms: 2000,
       })
@@ -656,8 +654,6 @@ describe("telemetry.toAppInsightsEnvelopes (indirect)", () => {
       expect(measurements.tokens_reasoning).toBe(50)
       expect(measurements.tokens_cache_read).toBe(10)
       expect(measurements.tokens_cache_write).toBe(5)
-      // Raw "tokens" key should not appear in properties
-      expect(envelopes[0].data.baseData.properties.tokens).toBeUndefined()
     } finally {
       cleanup()
     }