diff --git a/site/src/pages/AgentsPage/AgentSettingsCompactionPage.tsx b/site/src/pages/AgentsPage/AgentSettingsCompactionPage.tsx
index 16a338e6b1e..f7e9cbbf92b 100644
--- a/site/src/pages/AgentsPage/AgentSettingsCompactionPage.tsx
+++ b/site/src/pages/AgentsPage/AgentSettingsCompactionPage.tsx
@@ -1,7 +1,8 @@
 import type { FC } from "react";
-import { useMutation, useQuery, useQueryClient } from "react-query";
+import { useMutation, useQueries, useQuery, useQueryClient } from "react-query";
 import {
 	deleteUserCompactionThreshold,
+	organizationChatModelOverrides,
 	updateUserCompactionThreshold,
 	userChatProviderConfigs,
 	userCompactionThresholds,
@@ -19,6 +20,25 @@ const AgentSettingsCompactionPage: FC = () => {
 	);
 	const providerConfigsQuery = useQuery(userChatProviderConfigs());
 	const thresholdsQuery = useQuery(userCompactionThresholds());
+	// Only refines the displayed trigger point; a failed request falls back
+	// to the chat model's own window.
+	const modelOverrideQueries = useQueries({
+		queries: organizations.map((organization) =>
+			organizationChatModelOverrides(organization.id),
+		),
+	});
+	const compactionModelIDByOrganization = new Map<string, string>();
+	for (const [index, query] of modelOverrideQueries.entries()) {
+		const compactionOverride = query.data?.overrides.find(
+			(override) => override.context === "compaction",
+		);
+		if (compactionOverride) {
+			compactionModelIDByOrganization.set(
+				organizations[index].id,
+				compactionOverride.model_config_id,
+			);
+		}
+	}
 	const saveThresholdMutation = useMutation(
 		updateUserCompactionThreshold(queryClient),
 	);
@@ -44,6 +64,7 @@ const AgentSettingsCompactionPage: FC = () => {
 			models={organizationModels.models}
 			providerTypeByID={providerTypeByID}
 			organizations={organizations}
+			compactionModelIDByOrganization={compactionModelIDByOrganization}
 			modelsError={organizationModels.error ?? organizationModels.partialError}
 			isLoadingModels={organizationModels.isLoading}
 			thresholds={thresholdsQuery.data?.thresholds}
diff --git a/site/src/pages/AgentsPage/AgentSettingsCompactionPageView.tsx b/site/src/pages/AgentsPage/AgentSettingsCompactionPageView.tsx
index 61136d266f2..1fa6aa89c99 100644
--- a/site/src/pages/AgentsPage/AgentSettingsCompactionPageView.tsx
+++ b/site/src/pages/AgentsPage/AgentSettingsCompactionPageView.tsx
@@ -7,6 +7,7 @@ export interface AgentSettingsCompactionPageViewProps {
 	models: readonly TypesGen.ChatModel[] | undefined;
 	providerTypeByID: ReadonlyMap<string, string>;
 	organizations: readonly TypesGen.Organization[];
+	compactionModelIDByOrganization?: ReadonlyMap<string, string>;
 	modelsError: unknown;
 	isLoadingModels: boolean;
 	thresholds: readonly TypesGen.UserChatCompactionThreshold[] | undefined;
@@ -25,6 +26,7 @@ export const AgentSettingsCompactionPageView: FC<
 	models,
 	providerTypeByID,
 	organizations,
+	compactionModelIDByOrganization,
 	modelsError,
 	isLoadingModels,
 	thresholds,
@@ -43,6 +45,7 @@ export const AgentSettingsCompactionPageView: FC<
 				models={models ?? []}
 				providerTypeByID={providerTypeByID}
 				organizations={organizations}
+				compactionModelIDByOrganization={compactionModelIDByOrganization}
 				modelsError={modelsError}
 				isLoadingModels={isLoadingModels}
 				thresholds={thresholds}
diff --git a/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.stories.tsx b/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.stories.tsx
index 8a3d57eb567..e9b9ccf369f 100644
--- a/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.stories.tsx
+++ b/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.stories.tsx
@@ -94,6 +94,36 @@ export const Default: Story = {
 	},
 };
 
+export const ContextWindowTracksDraft: Story = {
+	name: "Context Window Tracks Draft",
+	play: async ({ canvasElement }) => {
+		const canvas = within(canvasElement);
+		const gpt4oInput = await canvas.findByRole("textbox", {
+			name: /GPT-4o compaction threshold/i,
+		});
+
+		// 128K window: default 80% compacts at ~102K, the draft moves it to ~64K.
+		await userEvent.type(gpt4oInput, "50");
+	},
+};
+
+export const CompactionOverrideShrinksWindow: Story = {
+	name: "Compaction Override Shrinks Window",
+	args: {
+		// The organization summarizes with the 16K model, so both enabled
+		// models show a 16K compaction window instead of their own.
+		compactionModelIDByOrganization: new Map([
+			[MockChatModel.organization_id, "model-3"],
+		]),
+	},
+};
+
+export const UnknownContextWindow: Story = {
+	args: {
+		models: [{ ...mockModels[0], context_limit: 0 }],
+	},
+};
+
 export const EmptyOrganizationDisplayNameFallsBackToName: Story = {
 	args: {
 		organizations: [organizationWithEmptyDisplayName],
diff --git a/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.tsx b/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.tsx
index 5c0c0a8e232..b7e7049de32 100644
--- a/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.tsx
+++ b/site/src/pages/AgentsPage/components/UserCompactionThresholdSettings.tsx
@@ -29,13 +29,23 @@ import {
 	TooltipContent,
 	TooltipTrigger,
 } from "#/components/Tooltip/Tooltip";
+import { formatContextLimit } from "#/modules/aiModels/ModelSelector";
 import { ProviderIcon } from "#/modules/aiModels/ProviderIcon";
 import { formatProviderLabel } from "#/utils/aiProviders";
+import {
+	compactionTriggerTokens,
+	resolveCompactionContextLimit,
+} from "../utils/modelOptions";
 
 interface UserCompactionThresholdSettingsProps {
 	models: readonly TypesGen.ChatModel[];
 	providerTypeByID: ReadonlyMap<string, string>;
 	organizations: readonly TypesGen.Organization[];
+	/**
+	 * Organization ID to the model config the organization routes compaction
+	 * through. Missing entries mean the chat model summarizes itself.
+	 */
+	compactionModelIDByOrganization?: ReadonlyMap<string, string>;
 	modelsError?: unknown;
 	isLoadingModels?: boolean;
 	thresholds: readonly TypesGen.UserChatCompactionThreshold[] | undefined;
@@ -48,6 +58,8 @@ interface UserCompactionThresholdSettingsProps {
 	onResetThreshold: (modelId: string) => Promise<unknown>;
 }
 
+const noCompactionOverrides: ReadonlyMap<string, string> = new Map();
+
 const parseThresholdDraft = (value: string): number | null => {
 	const trimmedValue = value.trim();
 	if (!/^\d+$/.test(trimmedValue)) {
@@ -80,6 +92,7 @@ export const UserCompactionThresholdSettings: FC<
 	models,
 	providerTypeByID,
 	organizations,
+	compactionModelIDByOrganization = noCompactionOverrides,
 	modelsError,
 	isLoadingModels,
 	thresholds,
@@ -330,6 +343,7 @@ export const UserCompactionThresholdSettings: FC<
 						<TableHeader>
 							<TableRow>
 								<TableHead className="text-content-secondary">Model</TableHead>
+								<TableHead className="w-0 whitespace-nowrap">Context</TableHead>
 								<TableHead className="w-0 whitespace-nowrap">Default</TableHead>
 								<TableHead className="w-0 whitespace-nowrap">
 									Threshold
@@ -361,6 +375,21 @@ export const UserCompactionThresholdSettings: FC<
 								const organizationName =
 									organizationNameByID.get(modelConfig.organization_id) ??
 									modelConfig.organization_id;
+								// Prefer the typed draft so the trigger point tracks
+								// what the user is about to save.
+								const effectiveThreshold =
+									parsedDraftValue ??
+									existingOverride ??
+									modelConfig.compression_threshold;
+								const contextLimit = resolveCompactionContextLimit(
+									modelConfig,
+									models,
+									compactionModelIDByOrganization,
+								);
+								const triggerTokens = compactionTriggerTokens(
+									contextLimit,
+									effectiveThreshold,
+								);
 
 								return (
 									<TableRow key={modelConfig.id}>
@@ -388,6 +417,20 @@ export const UserCompactionThresholdSettings: FC<
 												</p>
 											)}
 										</TableCell>
+										<TableCell className="w-0 whitespace-nowrap tabular-nums">
+											{contextLimit > 0 ? (
+												<div className="flex flex-col">
+													<span>{formatContextLimit(contextLimit)} tokens</span>
+													{triggerTokens !== undefined && (
+														<span className="text-2xs text-content-secondary">
+															Compacts at ~{formatContextLimit(triggerTokens)}
+														</span>
+													)}
+												</div>
+											) : (
+												<span className="text-content-secondary">Unknown</span>
+											)}
+										</TableCell>
 										<TableCell className="w-0 whitespace-nowrap tabular-nums">
 											{modelConfig.compression_threshold}%
 										</TableCell>
@@ -484,7 +527,7 @@ export const UserCompactionThresholdSettings: FC<
 						</TableBody>
 						<TableFooter className="bg-transparent">
 							<TableRow className="border-0">
-								<TableCell colSpan={3} className="border-0 p-0">
+								<TableCell colSpan={4} className="border-0 p-0">
 									<div className="mt-2 flex h-6 items-center justify-end gap-2 px-3">
 										{isSavedVisible ? (
 											<TemporarySavedState />
diff --git a/site/src/pages/AgentsPage/utils/modelOptions.test.ts b/site/src/pages/AgentsPage/utils/modelOptions.test.ts
index 188590d5527..6cc6e46b6e6 100644
--- a/site/src/pages/AgentsPage/utils/modelOptions.test.ts
+++ b/site/src/pages/AgentsPage/utils/modelOptions.test.ts
@@ -11,6 +11,7 @@ import {
 	MockChatProviderConfig,
 } from "#/testHelpers/chatModels";
 import {
+	compactionTriggerTokens,
 	countConfiguredProviderConfigs,
 	filterModelsWithEnabledProvider,
 	formatProviderLabel,
@@ -25,6 +26,7 @@ import {
 	NIL_UUID,
 	providerInfoByIDFromUserConfigs,
 	providerTypeByIDFromUserConfigs,
+	resolveCompactionContextLimit,
 	resolveCompactionThreshold,
 	resolveModelOptionId,
 	resolveModelSelector,
@@ -1210,3 +1212,89 @@ describe("resolveCompactionThreshold", () => {
 		expect(resolveCompactionThreshold("missing", [], models)).toBe(undefined);
 	});
 });
+
+describe("resolveCompactionContextLimit", () => {
+	const chatModel = createConfig({
+		id: "chat",
+		ai_provider_id: "prov-openai",
+		model: "gpt-4o",
+		context_limit: 1_000,
+	});
+	const smallSummarizer = createConfig({
+		id: "small",
+		ai_provider_id: "prov-openai",
+		model: "gpt-4o-mini",
+		context_limit: 100,
+	});
+	const largeSummarizer = createConfig({
+		id: "large",
+		ai_provider_id: "prov-anthropic",
+		model: "claude",
+		context_limit: 5_000,
+	});
+	const models = [chatModel, smallSummarizer, largeSummarizer];
+
+	it("uses the chat model window without an override", () => {
+		expect(resolveCompactionContextLimit(chatModel, models, new Map())).toBe(
+			1_000,
+		);
+	});
+
+	it("uses the override window when it is smaller", () => {
+		expect(
+			resolveCompactionContextLimit(
+				chatModel,
+				models,
+				new Map([[testOrganizationID, "small"]]),
+			),
+		).toBe(100);
+	});
+
+	it("keeps the chat model window when the override is larger", () => {
+		expect(
+			resolveCompactionContextLimit(
+				chatModel,
+				models,
+				new Map([[testOrganizationID, "large"]]),
+			),
+		).toBe(1_000);
+	});
+
+	it("falls back to the override window when the chat window is unknown", () => {
+		expect(
+			resolveCompactionContextLimit(
+				{ ...chatModel, context_limit: 0 },
+				models,
+				new Map([[testOrganizationID, "small"]]),
+			),
+		).toBe(100);
+	});
+
+	it("ignores overrides for other organizations and unknown models", () => {
+		expect(
+			resolveCompactionContextLimit(
+				chatModel,
+				models,
+				new Map([
+					["other-org", "small"],
+					[testOrganizationID, "missing"],
+				]),
+			),
+		).toBe(1_000);
+	});
+});
+
+describe("compactionTriggerTokens", () => {
+	it("scales the window by the threshold", () => {
+		expect(compactionTriggerTokens(128_000, 80)).toBe(102_400);
+		expect(compactionTriggerTokens(1_000, 70)).toBe(700);
+	});
+
+	it("returns undefined when the window is unknown", () => {
+		expect(compactionTriggerTokens(0, 80)).toBe(undefined);
+	});
+
+	it("returns undefined at 100% because compaction never triggers", () => {
+		expect(compactionTriggerTokens(128_000, 100)).toBe(undefined);
+	});
+});
diff --git a/site/src/pages/AgentsPage/utils/modelOptions.ts b/site/src/pages/AgentsPage/utils/modelOptions.ts
index b58f446842c..5f03bd1c886 100644
--- a/site/src/pages/AgentsPage/utils/modelOptions.ts
+++ b/site/src/pages/AgentsPage/utils/modelOptions.ts
@@ -342,6 +342,42 @@ export function resolveCompactionThreshold(
 	return model.compression_threshold;
 }
 
+/**
+ * Context window the compaction trigger is measured against. Mirrors the
+ * backend: when the organization routes compaction to an override model,
+ * the history must also fit that model's window, so the smaller of the
+ * two limits wins. Returns 0 when neither limit is known.
+ */
+export function resolveCompactionContextLimit(
+	model: TypesGen.ChatModel,
+	models: readonly TypesGen.ChatModel[],
+	compactionModelIDByOrganization: ReadonlyMap<string, string>,
+): number {
+	const chatLimit = model.context_limit > 0 ? model.context_limit : 0;
+	const overrideID = compactionModelIDByOrganization.get(model.organization_id);
+	const overrideLimit =
+		models.find((candidate) => candidate.id === overrideID)?.context_limit ?? 0;
+	if (overrideLimit > 0 && (chatLimit <= 0 || overrideLimit < chatLimit)) {
+		return overrideLimit;
+	}
+	return chatLimit;
+}
+
+/**
+ * Token count at which compaction triggers for the given context window and
+ * threshold, or undefined when the window is unknown or compaction is
+ * disabled (100%).
+ */
+export function compactionTriggerTokens(
+	contextLimit: number,
+	thresholdPercent: number,
+): number | undefined {
+	if (contextLimit <= 0 || thresholdPercent >= 100) {
+		return undefined;
+	}
+	return Math.round((contextLimit * thresholdPercent) / 100);
+}
+
 export const getModelSelectorPlaceholder = (
 	modelOptions: readonly ModelSelectorOption[],
 	isModelCatalogLoading: boolean,