diff --git a/src/handlers/message.ts b/src/handlers/message.ts
index 248228a..ee3df11 100644
--- a/src/handlers/message.ts
+++ b/src/handlers/message.ts
@@ -27,6 +27,12 @@ import {
   TOOL_NAME,
   TOOL_PARAMETERS,
 } from "@arizeai/openinference-semantic-conventions"
+import {
+  ATTR_GEN_AI_REQUEST_MODEL,
+  ATTR_GEN_AI_RESPONSE_MODEL,
+  ATTR_GEN_AI_USAGE_INPUT_TOKENS,
+  ATTR_GEN_AI_USAGE_OUTPUT_TOKENS,
+} from "@opentelemetry/semantic-conventions/incubating"
 import {
   agentAttrs,
   errorSummary,
@@ -42,6 +48,11 @@ import type { HandlerContext } from "../types.ts"
 
 const OPENINFERENCE_SPAN_KIND = SemanticConventions.OPENINFERENCE_SPAN_KIND
 const LLM_FINISH_REASON = "llm.finish_reason"
+// Spelled out rather than taken from semconv: the incubating constants emit
+// `gen_ai.usage.cache_read.input_tokens`, while Anthropic's API and every GenAI
+// consumer we target read the underscored form below.
+const ATTR_GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS = "gen_ai.usage.cache_read_input_tokens"
+const ATTR_GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS = "gen_ai.usage.cache_creation_input_tokens"
 
 type SubtaskPart = {
   type: "subtask"
@@ -130,6 +141,11 @@ export function handleMessageUpdated(e: EventMessageUpdated, ctx: HandlerContext
       [LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_READ]: assistant.tokens.cache.read,
       [LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_WRITE]: assistant.tokens.cache.write,
       [LLM_TOKEN_COUNT_TOTAL]: totalTokens,
+      [ATTR_GEN_AI_RESPONSE_MODEL]: modelID,
+      [ATTR_GEN_AI_USAGE_INPUT_TOKENS]: assistant.tokens.input,
+      [ATTR_GEN_AI_USAGE_OUTPUT_TOKENS]: assistant.tokens.output,
+      [ATTR_GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS]: assistant.tokens.cache.read,
+      [ATTR_GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS]: assistant.tokens.cache.write,
       [LLM_FINISH_REASON]: assistant.error ? "error" : (assistant.finish ?? "stop"),
       [LLM_COST_TOTAL]: assistant.cost,
       ...(outputText
@@ -454,6 +470,7 @@ export function startMessageSpan(
         [LLM_PROVIDER]: providerID,
         "gen_ai.provider.name": genAiProviderName(providerID),
         [LLM_MODEL_NAME]: modelID,
+        [ATTR_GEN_AI_REQUEST_MODEL]: modelID,
         ...(inputText
           ? {
               [INPUT_VALUE]: inputText,
diff --git a/tests/handlers/spans.test.ts b/tests/handlers/spans.test.ts
index 18851d0..ef2bdb3 100644
--- a/tests/handlers/spans.test.ts
+++ b/tests/handlers/spans.test.ts
@@ -370,6 +370,16 @@ describe("message (LLM) spans", () => {
     expect(tracer.spans[0]!.attributes[LLM_MODEL_NAME]).toBe("claude-sonnet-4")
   })
 
+  // Asserted as literal keys, not via the semconv constants the source imports:
+  // these strings are the wire format Gen AI consumers match on, so the test has
+  // to fail if a constant is renamed upstream.
+  test("startMessageSpan sets OTel GenAI model attribute", () => {
+    const { ctx, tracer } = makeCtx()
+    startMessageSpan("ses_1", "msg_1", "user_1", "claude-sonnet-4", "amazon-bedrock", 1000, ctx)
+    expect(tracer.spans[0]!.attributes["gen_ai.request.model"]).toBe("claude-sonnet-4")
+    expect(tracer.spans[0]!.attributes["gen_ai.provider.name"]).toBe("aws.bedrock")
+  })
+
   test("startMessageSpan is a no-op when span already exists for sessionID:messageID", () => {
     const { ctx, tracer } = makeCtx()
     startMessageSpan("ses_1", "msg_1", "user_1", "claude", "anthropic", 1000, ctx)
@@ -434,6 +444,44 @@ describe("message (LLM) spans", () => {
     expect(span.attributes["agent.type"]).toBe("subagent")
   })
 
+  test("handleMessageUpdated sets OTel GenAI token attributes on span", () => {
+    const { ctx, tracer } = makeCtx()
+    startMessageSpan("ses_1", "msg_1", "user_1", "claude-3-5-sonnet", "anthropic", 1000, ctx)
+    handleMessageUpdated(
+      makeAssistantMessageUpdated({
+        id: "msg_1",
+        modelID: "claude-3-5-sonnet",
+        tokens: { input: 200, output: 80, reasoning: 10, cache: { read: 30, write: 5 } },
+      }),
+      ctx,
+    )
+    const span = tracer.spans[0]!
+    expect(span.attributes["gen_ai.response.model"]).toBe("claude-3-5-sonnet")
+    expect(span.attributes["gen_ai.usage.input_tokens"]).toBe(200)
+    expect(span.attributes["gen_ai.usage.output_tokens"]).toBe(80)
+    // Underscored, matching Anthropic's API and Gen AI consumers. The semconv
+    // constants spell these with a dot (gen_ai.usage.cache_read.input_tokens).
+    expect(span.attributes["gen_ai.usage.cache_read_input_tokens"]).toBe(30)
+    expect(span.attributes["gen_ai.usage.cache_creation_input_tokens"]).toBe(5)
+  })
+
+  test("GenAI token attributes do not displace the OpenInference ones", () => {
+    const { ctx, tracer } = makeCtx()
+    startMessageSpan("ses_1", "msg_1", "user_1", "claude-3-5-sonnet", "anthropic", 1000, ctx)
+    handleMessageUpdated(
+      makeAssistantMessageUpdated({
+        id: "msg_1",
+        tokens: { input: 200, output: 80, reasoning: 10, cache: { read: 30, write: 5 } },
+      }),
+      ctx,
+    )
+    const span = tracer.spans[0]!
+    expect(span.attributes[LLM_TOKEN_COUNT_PROMPT]).toBe(200)
+    expect(span.attributes["gen_ai.usage.input_tokens"]).toBe(200)
+    expect(span.attributes[LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_WRITE]).toBe(5)
+    expect(span.attributes["gen_ai.usage.cache_creation_input_tokens"]).toBe(5)
+  })
+
   test("handleMessageUpdated no-ops span handling when no span exists for messageID", () => {
     const { ctx, tracer } = makeCtx()
     const spansBefore = tracer.spans.length