Skip to content

🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.

chore(topics): switch to gpt-6 luna defaults (#18338)

Nimar0dfba8865d89
diff --git a/packages/shared/src/server/topics/text.test.ts b/packages/shared/src/server/topics/text.test.ts
index 3bc27b0679..ee116b58d3 100644
--- a/packages/shared/src/server/topics/text.test.ts
+++ b/packages/shared/src/server/topics/text.test.ts
@@ -18,7 +18,12 @@ afterEach(() => {
});
describe("generateTopicText", () => {
- it.each(["us.openai.gpt-5.6-luna", "us.openai.gpt-5.6-terra"])(
+ it.each([
+ "us.openai.gpt-5.6-luna",
+ "us.openai.gpt-5.6-terra",
+ "us.openai.gpt-6-luna",
+ "global.openai.gpt-6-luna",
+ ])(
"serializes %s structured output through Bedrock with the Topics profile",
async (model) => {
const schema = z.object({ summary: z.string() });
diff --git a/worker/src/features/topics/README.md b/worker/src/features/topics/README.md
index 48f371871d..696aab37b2 100644
--- a/worker/src/features/topics/README.md
+++ b/worker/src/features/topics/README.md
@@ -88,7 +88,14 @@ Summaries snapshot source environment and trace name; assignments copy that
snapshot, including empty environment strings. Reprocessing refreshes metadata
even when text is reused; session results have no trace name.
-The summary model is selected with `LANGFUSE_TOPICS_SUMMARY_MODEL`. Topic
+The summary model is selected with `LANGFUSE_TOPICS_SUMMARY_MODEL`. Use
+`us.openai.gpt-6-luna` for US routing or `global.openai.gpt-6-luna` for global
+routing; GPT-6 Luna has no EU geographic profile. Both support the existing
+Bedrock Converse tool-based structured output with reasoning disabled.
+Summary cost estimates include both GPT-6 Luna profiles and the existing
+`us.openai.gpt-5.6-luna` profile. See the
+[GPT-6 Luna model card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-luna.html)
+for supported source regions and pricing. Topic
naming uses `us.openai.gpt-5.6-terra` through Bedrock Converse with reasoning
disabled. The AWS model cards for [GPT-5.6 Luna](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-luna.html)
and [GPT-5.6 Terra](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-terra.html)
diff --git a/worker/src/features/topics/models.test.ts b/worker/src/features/topics/models.test.ts
index 375f3f3c2c..34ef3c26e3 100644
--- a/worker/src/features/topics/models.test.ts
+++ b/worker/src/features/topics/models.test.ts
@@ -5,6 +5,7 @@ const state = vi.hoisted(() => ({
embed: vi.fn(),
increment: vi.fn(),
region: vi.fn(),
+ summaryModel: vi.fn(),
}));
vi.mock("@langfuse/shared/src/server", () => ({
getLangfuseAIAwsProfile: () => "ai-test",
@@ -14,7 +15,7 @@ vi.mock("@langfuse/shared/src/server", () => ({
}));
vi.mock("@langfuse/shared/topics/server", () => ({
getTopicsModelConfig: () => ({
- summaryModel: "us.openai.gpt-5.6-luna",
+ summaryModel: state.summaryModel(),
embeddingModel: "eu.cohere.embed-v4:0",
}),
generateTopicText: (...args: unknown[]) => state.call(...args),
@@ -49,47 +50,56 @@ beforeEach(() => {
state.embed.mockReset();
state.increment.mockReset();
state.region.mockReset().mockReturnValue("eu-west-1");
+ state.summaryModel.mockReset().mockReturnValue("us.openai.gpt-5.6-luna");
});
describe("Topics naming boundary", () => {
- it("summarizes a trace for its facet without requiring citations", async () => {
- state.call.mockResolvedValue({
- output: { summary: "A billing request.", status: "applicable" },
- usage: { inputTokens: 100, outputTokens: 30 },
- });
- const result = await summarizeTopicTrace(
- facet,
- "RAW_TRANSCRIPT_SENTINEL",
- topicProcessingConfigSchema.parse({
- summaryModel: "us.openai.gpt-5.6-luna",
- }),
- );
- expect(result.output).toEqual({
- summary: "A billing request.",
- status: "applicable",
- });
- expect(result.providedUsageDetails).toEqual({
- summary_input: 100,
- summary_output: 30,
- });
- expect(result.usageDetails).toEqual({
- summary_input: 100,
- summary_output: 30,
- total: 130,
- });
- expect(result.providedCostDetails).toEqual({});
- expect(result.costDetails.summary_input).toBeCloseTo(0.00002, 10);
- expect(result.costDetails.summary_output).toBeCloseTo(0.000036, 10);
- expect(result.costDetails.total).toBeCloseTo(0.000056, 10);
- const request = state.call.mock.calls[0][0];
- expect(request).toMatchObject({
- model: "us.openai.gpt-5.6-luna",
- region: "eu-west-1",
- profile: "ai-test",
- });
- expect(request.messages[0].content).toContain(facet.prompt);
- expect(request.messages[1].content).toBe("RAW_TRANSCRIPT_SENTINEL");
- });
+ it.each([
+ ["us.openai.gpt-5.6-luna", 0.00002, 0.000036],
+ ["us.openai.gpt-6-luna", 0.000011, 0.0000165],
+ ["global.openai.gpt-6-luna", 0.00001, 0.000015],
+ ])(
+ "summarizes a trace and prices %s usage",
+ async (model, inputCost, outputCost) => {
+ state.summaryModel.mockReturnValue(model);
+ state.call.mockResolvedValue({
+ output: { summary: "A billing request.", status: "applicable" },
+ usage: { inputTokens: 100, outputTokens: 30 },
+ });
+ const result = await summarizeTopicTrace(
+ facet,
+ "RAW_TRANSCRIPT_SENTINEL",
+ topicProcessingConfigSchema.parse({
+ summaryModel: model,
+ }),
+ );
+ expect(result.output).toEqual({
+ summary: "A billing request.",
+ status: "applicable",
+ });
+ expect(result.providedUsageDetails).toEqual({
+ summary_input: 100,
+ summary_output: 30,
+ });
+ expect(result.usageDetails).toEqual({
+ summary_input: 100,
+ summary_output: 30,
+ total: 130,
+ });
+ expect(result.providedCostDetails).toEqual({});
+ expect(result.costDetails.summary_input).toBeCloseTo(inputCost, 10);
+ expect(result.costDetails.summary_output).toBeCloseTo(outputCost, 10);
+ expect(result.costDetails.total).toBeCloseTo(inputCost + outputCost, 10);
+ const request = state.call.mock.calls[0][0];
+ expect(request).toMatchObject({
+ model,
+ region: "eu-west-1",
+ profile: "ai-test",
+ });
+ expect(request.messages[0].content).toContain(facet.prompt);
+ expect(request.messages[1].content).toBe("RAW_TRANSCRIPT_SENTINEL");
+ },
+ );
it("rejects missing Bedrock configuration before calling the provider", async () => {
state.region.mockReturnValue(undefined);
diff --git a/worker/src/features/topics/models.ts b/worker/src/features/topics/models.ts
index 5e7e11d8a9..dbee9c650b 100644
--- a/worker/src/features/topics/models.ts
+++ b/worker/src/features/topics/models.ts
@@ -23,7 +23,13 @@ import {
} from "./provider-error";
export const TOPICS_NAMING_MODEL = "us.openai.gpt-5.6-terra";
-const TOPICS_SUMMARY_COST_MODEL = "us.openai.gpt-5.6-luna";
+// Summary inputs are capped below the 272k-token long-context pricing threshold.
+const TOPICS_SUMMARY_RATES: Record<string, { input: number; output: number }> =
+ {
+ "us.openai.gpt-5.6-luna": { input: 0.2, output: 1.2 },
+ "us.openai.gpt-6-luna": { input: 0.11, output: 0.55 },
+ "global.openai.gpt-6-luna": { input: 0.1, output: 0.5 },
+ };
const TOPICS_EMBEDDING_COST_MODELS = new Set([
"us.cohere.embed-v4:0",
"eu.cohere.embed-v4:0",
@@ -96,7 +102,7 @@ async function structuredCall<T>(
input: tokens > 272_000 ? 4 : 2,
output: tokens > 272_000 ? 18 : 12,
}
- : { input: 0.2, output: 1.2 };
+ : TOPICS_SUMMARY_RATES[model];
const result = await generateTopicText({
...connection,
model,
@@ -110,10 +116,7 @@ async function structuredCall<T>(
});
throw topicProviderError(error);
});
- const actualRates =
- stage === "naming" || model === TOPICS_SUMMARY_COST_MODEL
- ? rates(result.usage.inputTokens ?? inputLimit)
- : null;
+ const actualRates = rates(result.usage.inputTokens ?? inputLimit);
recordTopicTokenUsage(stage, {
input: result.usage.inputTokens,
output: result.usage.outputTokens,