fix(ai): avoid double-counting reasoning tokens

closes #3581

packages/ai/CHANGELOG.md

@@ -8,6 +8,8 @@
 
 ### Fixed
 
+- Fixed OpenAI-compatible completion usage parsing to stop double-counting reasoning tokens already included in `completion_tokens` ([#3581](https://github.com/badlogic/pi-mono/issues/3581))
+- Fixed long cache retention compatibility by adding `compat.supportsLongCacheRetention`, allowing Anthropic Messages and OpenAI-compatible proxies to explicitly disable long-retention fields while enabling long retention by default when requested ([#3543](https://github.com/badlogic/pi-mono/issues/3543))
 - Fixed `openai-responses` compatibility by adding `compat.sendSessionIdHeader: false`, allowing strict OpenAI-compatible proxies to omit the underscore-containing `session_id` header while still sending other session-affinity headers ([#3579](https://github.com/badlogic/pi-mono/issues/3579))
 - Fixed `anthropic-messages` tool streaming compatibility by adding `compat.supportsEagerToolInputStreaming`, allowing Anthropic-compatible providers to omit per-tool `eager_input_streaming` and use the legacy fine-grained tool streaming beta header instead ([#3575](https://github.com/badlogic/pi-mono/issues/3575))
 - Fixed `supportsXhigh()` to recognize `openai-codex` `gpt-5.5`, preserving `xhigh` reasoning requests instead of clamping them to `high`.

packages/ai/src/providers/openai-completions.ts

@@ -465,15 +465,18 @@ function buildParams(
 	cacheRetention: CacheRetention = resolveCacheRetention(options?.cacheRetention),
 ) {
 	const messages = convertMessages(model, context, compat);
-	const cacheControl = getCompatCacheControl(model, compat, cacheRetention);
+	const cacheControl = getCompatCacheControl(compat, cacheRetention);
 
 	const params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {
 		model: model.id,
 		messages,
 		stream: true,
 		prompt_cache_key:
-			model.baseUrl.includes("api.openai.com") && cacheRetention !== "none" ? options?.sessionId : undefined,
-		prompt_cache_retention: model.baseUrl.includes("api.openai.com") && cacheRetention === "long" ? "24h" : undefined,
+			(model.baseUrl.includes("api.openai.com") && cacheRetention !== "none") ||
+			(cacheRetention === "long" && compat.supportsLongCacheRetention)
+				? options?.sessionId
+				: undefined,
+		prompt_cache_retention: cacheRetention === "long" && compat.supportsLongCacheRetention ? "24h" : undefined,
 	};
 
 	if (compat.supportsUsageInStreaming !== false) {
@@ -565,7 +568,6 @@ function mapReasoningEffort(
 }
 
 function getCompatCacheControl(
-	model: Model<"openai-completions">,
 	compat: ResolvedOpenAICompletionsCompat,
 	cacheRetention: CacheRetention,
 ): OpenAICompatCacheControl | undefined {
@@ -573,7 +575,7 @@ function getCompatCacheControl(
 		return undefined;
 	}
 
-	const ttl = cacheRetention === "long" && model.baseUrl.includes("api.anthropic.com") ? "1h" : undefined;
+	const ttl = cacheRetention === "long" && compat.supportsLongCacheRetention ? "1h" : undefined;
 	return { type: "ephemeral", ...(ttl ? { ttl } : {}) };
 }
 
@@ -937,14 +939,12 @@ function parseChunkUsage(
 		prompt_tokens?: number;
 		completion_tokens?: number;
 		prompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };
-		completion_tokens_details?: { reasoning_tokens?: number };
 	},
 	model: Model<"openai-completions">,
 ): AssistantMessage["usage"] {
 	const promptTokens = rawUsage.prompt_tokens || 0;
 	const reportedCachedTokens = rawUsage.prompt_tokens_details?.cached_tokens || 0;
 	const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;
-	const reasoningTokens = rawUsage.completion_tokens_details?.reasoning_tokens || 0;
 
 	// Normalize to pi-ai semantics:
 	// - cacheRead: hits from cache created by previous requests only
@@ -955,9 +955,8 @@ function parseChunkUsage(
 		cacheWriteTokens > 0 ? Math.max(0, reportedCachedTokens - cacheWriteTokens) : reportedCachedTokens;
 
 	const input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);
-	// Compute totalTokens ourselves since we add reasoning_tokens to output
-	// and some providers (e.g., Groq) don't include them in total_tokens
-	const outputTokens = (rawUsage.completion_tokens || 0) + reasoningTokens;
+	// OpenAI completion_tokens already includes reasoning_tokens.
+	const outputTokens = rawUsage.completion_tokens || 0;
 	const usage: AssistantMessage["usage"] = {
 		input,
 		output: outputTokens,
@@ -1055,6 +1054,7 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 		supportsStrictMode: true,
 		cacheControlFormat,
 		sendSessionAffinityHeaders: false,
+		supportsLongCacheRetention: true,
 	};
 }
 
@@ -1084,5 +1084,6 @@ function getCompat(model: Model<"openai-completions">): ResolvedOpenAICompletion
 		supportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,
 		cacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,
 		sendSessionAffinityHeaders: model.compat.sendSessionAffinityHeaders ?? detected.sendSessionAffinityHeaders,
+		supportsLongCacheRetention: model.compat.supportsLongCacheRetention ?? detected.supportsLongCacheRetention,
 	};
 }

packages/ai/test/openai-completions-tool-choice.test.ts

@@ -552,6 +552,41 @@ describe("openai-completions tool_choice", () => {
 		expect(toolCall).not.toHaveProperty("partialArgs");
 	});
 
+	it("does not double-count reasoning tokens in completion usage", async () => {
+		mockState.chunks = [
+			{
+				id: "chatcmpl-reasoning-usage",
+				choices: [{ delta: {}, finish_reason: "stop" }],
+				usage: {
+					prompt_tokens: 10,
+					completion_tokens: 33,
+					prompt_tokens_details: { cached_tokens: 0 },
+					completion_tokens_details: { reasoning_tokens: 21 },
+				},
+			},
+		];
+
+		const { compat: _compat, ...baseModel } = getModel("openai", "gpt-4o-mini")!;
+		const model = { ...baseModel, api: "openai-completions" } as const;
+		const response = await streamSimple(
+			model,
+			{
+				messages: [
+					{
+						role: "user",
+						content: "Use reasoning.",
+						timestamp: Date.now(),
+					},
+				],
+			},
+			{ apiKey: "test" },
+		).result();
+
+		expect(response.usage.input).toBe(10);
+		expect(response.usage.output).toBe(33);
+		expect(response.usage.totalTokens).toBe(43);
+	});
+
 	it("preserves prompt_tokens_details.cache_write_tokens from chunk usage", async () => {
 		mockState.chunks = [
 			{