From 565714a015e673514fc88cb48f4a9d8c8411fc00 Mon Sep 17 00:00:00 2001
From: "@taltas" <6816042+taltas@users.noreply.github.com>
Date: Fri, 16 Oct 2026 22:39:00 +0000
Subject: [PATCH] feat(providers): clamp Gemini 4 Argon limits on Vercel AI
 Gateway to Google's values

---
 .../__tests__/vercel-ai-gateway.spec.ts       | 27 +++++++++++++++++++
 .../providers/fetchers/vercel-ai-gateway.ts   | 13 +++++++++
 2 files changed, 40 insertions(+)

diff --git a/src/api/providers/fetchers/__tests__/vercel-ai-gateway.spec.ts b/src/api/providers/fetchers/__tests__/vercel-ai-gateway.spec.ts
index 1331ed4049..85ce9cf065 100644
--- a/src/api/providers/fetchers/__tests__/vercel-ai-gateway.spec.ts
+++ b/src/api/providers/fetchers/__tests__/vercel-ai-gateway.spec.ts
@@ -196,6 +196,33 @@ describe("Vercel AI Gateway Fetchers", () => {
 			},
 		}
 
+		it("applies Google's authoritative limits and thinking levels to google/gemini-4-argon", () => {
+			// The gateway catalog reports context_window and max_tokens of 1,000,000;
+			// Google's documented limits are a shared 1,048,576-token window and a
+			// 65,536-token single-request output cap.
+			const result = parseVercelAiGatewayModel({
+				id: "google/gemini-4-argon",
+				model: {
+					...baseModel,
+					id: "google/gemini-4-argon",
+					name: "Gemini 4 Argon",
+					owned_by: "google",
+					context_window: 1_000_000,
+					max_tokens: 1_000_000,
+					tags: ["reasoning", "vision", "tool-use", "implicit-caching"],
+				},
+			})
+
+			expect(result).toMatchObject({
+				contextWindow: 1_048_576,
+				maxTokens: 65_536,
+				supportsImages: true,
+				supportsReasoningEffort: ["low", "medium", "high"],
+				reasoningEffort: "medium",
+				supportsTemperature: false,
+			})
+		})
+
 		it.each(["openai/gpt-6-astra", "openai/gpt-6-astra-fast"])(
 			"applies required Astra request constraints to %s",
 			(id) => {
diff --git a/src/api/providers/fetchers/vercel-ai-gateway.ts b/src/api/providers/fetchers/vercel-ai-gateway.ts
index e9271f07af..04d785124a 100644
--- a/src/api/providers/fetchers/vercel-ai-gateway.ts
+++ b/src/api/providers/fetchers/vercel-ai-gateway.ts
@@ -125,6 +125,19 @@ export const parseVercelAiGatewayModel = ({ id, model }: { id: string; model: Ve
 		description: model.description ?? model.name,
 	}
 
+	if (id === "google/gemini-4-argon") {
+		// The gateway catalog reports context_window and max_tokens of 1,000,000,
+		// but Google's authoritative limits are a shared 1,048,576-token window
+		// (input + thinking + output) and a 65,536-token single-request output cap
+		// (longer decodes use continuation tokens).
+		modelInfo.contextWindow = 1_048_576
+		modelInfo.maxTokens = 65_536
+		modelInfo.supportsReasoningEffort = ["low", "medium", "high"]
+		modelInfo.reasoningEffort = "medium"
+		// The Google backend ignores temperature/top_p/top_k on this model.
+		modelInfo.supportsTemperature = false
+	}
+
 	if (id === "openai/gpt-6-astra" || id === "openai/gpt-6-astra-fast") {
 		modelInfo.supportsReasoningEffort = ["low", "medium", "high", "xhigh", "max"]
 		modelInfo.requiredReasoningEffort = true