diff --git a/src/lib/modelCapabilities.ts b/src/lib/modelCapabilities.ts
index 95330b45f8..3deb72832b 100644
--- a/src/lib/modelCapabilities.ts
+++ b/src/lib/modelCapabilities.ts
@@ -4,7 +4,7 @@ import {
 } from "@omniroute/open-sse/config/providerModels.ts";
 import { parseModel, resolveCanonicalProviderModel } from "@omniroute/open-sse/services/model.ts";
 import {
-  MODEL_SPECS,
+  findModelSpecIdByExactOrAlias,
   getAuthoritativeContextWindow,
   getAuthoritativeProviderContextWindow,
   getModelSpec,
@@ -285,17 +285,18 @@ function getAuthoritativeStaticContextWindow(
   return null;
 }
 
+// #8697-adjacent: this used to rescan Object.entries(MODEL_SPECS) per candidate per
+// call — the top hotspot in a full catalog-rebuild profile once the pricing-path and
+// getCanonicalModelSpecId() bottlenecks were fixed. Reuses the lazy index already built
+// for getCanonicalModelSpecId() (@/shared/constants/modelSpecs) instead of duplicating a
+// second cache over the same static table.
 function getStaticSpecCanonicalModelId(modelId: string | null, rawModel: string | null) {
   const candidates = [modelId, rawModel].filter(
     (candidate): candidate is string => typeof candidate === "string" && candidate.length > 0
   );
   for (const candidate of candidates) {
-    const lower = candidate.toLowerCase();
-    for (const [canonical, spec] of Object.entries(MODEL_SPECS)) {
-      if (canonical === "__default__") continue;
-      if (canonical.toLowerCase() === lower) return canonical;
-      if (spec.aliases?.some((alias) => alias.toLowerCase() === lower)) return canonical;
-    }
+    const hit = findModelSpecIdByExactOrAlias(candidate);
+    if (hit) return hit;
   }
   return null;
 }
@@ -311,7 +312,14 @@ function stripLatestAlias(modelId: string | null): string | null {
   return stripped && stripped !== modelId ? stripped : null;
 }
 
-function reverseModelsDevProviders(provider: string): string[] {
+// #8697-adjacent: MODELS_DEV_PROVIDER_MAP is a static module constant, so the result
+// of reverseModelsDevProviders() never changes for a given provider — memoized by
+// provider key instead of rescanning Object.entries(MODELS_DEV_PROVIDER_MAP) on every
+// call (called once per model in a catalog rebuild). Never evicted — bounded by the
+// number of distinct providers ever queried (~50-100 in practice), negligible memory.
+const reverseModelsDevProvidersCache = new Map<string, readonly string[]>();
+
+function reverseModelsDevProviders(provider: string): readonly string[] {
   // models.dev may store capabilities under a different OmniRoute provider id
   // that also maps from the same upstream models.dev provider. Build reverse
   // candidates from MODELS_DEV_PROVIDER_MAP (e.g. openai ↔ cx).
@@ -321,6 +329,9 @@ function reverseModelsDevProviders(provider: string): string[] {
   // list their alias (cx/cc), never the canonical id. Also probe the
   // provider's alias so a canonical id like "codex"/"claude" still matches
   // the map entries keyed only by "cx"/"cc" (#8429).
+  const cached = reverseModelsDevProvidersCache.get(provider);
+  if (cached) return cached;
+
   const out = new Set<string>();
   const providerAlias = PROVIDER_ID_TO_ALIAS[provider] || provider;
   for (const [modelsDevId, omniIds] of Object.entries(MODELS_DEV_PROVIDER_MAP)) {
@@ -334,7 +345,12 @@ function reverseModelsDevProviders(provider: string): string[] {
       for (const id of omniIds) out.add(id);
     }
   }
-  return [...out];
+  // Frozen: the result is now shared across every future call for this provider (via
+  // the cache above) instead of a fresh array per call — freeze prevents an accidental
+  // caller mutation (e.g. .push()) from corrupting the cache for everyone else.
+  const result = Object.freeze([...out]);
+  reverseModelsDevProvidersCache.set(provider, result);
+  return result;
 }
 
 function getSyncedCapabilityForResolved(
@@ -694,8 +710,7 @@ export function capThinkingBudget(input: CapabilityInput, budget: number): numbe
   // default to "gemini". Without this a cap learned via the executor would be
   // invisible to bare-model callers. Provider-qualified inputs keep their own
   // provider, preserving per-provider independence.
-  const providerForLearned =
-    resolved.provider ?? (modelLower.includes("gemini") ? "gemini" : null);
+  const providerForLearned = resolved.provider ?? (modelLower.includes("gemini") ? "gemini" : null);
 
   const learned = getLearnedThinkingCap(providerForLearned, modelId);
   if (learned !== null) {
diff --git a/src/lib/modelMetadataRegistry.ts b/src/lib/modelMetadataRegistry.ts
index 2e03a981d3..9f4e93aa2a 100644
--- a/src/lib/modelMetadataRegistry.ts
+++ b/src/lib/modelMetadataRegistry.ts
@@ -258,25 +258,47 @@ export function getCanonicalModelMetadata(input: {
   };
 }
 
+// #8697 second bottleneck (after getModelsDevPricing memoization above): findInsensitive
+// rebuilt a full Object.entries() scan on every miss, twice per model (provider lookup +
+// model lookup) — ~6091 models × ~180-210 entries ≈ 1.2-1.3M allocations per catalog
+// rebuild. Replaced with a lowercase-key index built once per distinct object and cached
+// by identity (WeakMap) — getModelsDevPricing() returns the same object reference while
+// its cache is warm, so the index is reused across every resolveCatalogPricing() call in
+// a rebuild instead of rebuilt per lookup.
+const lowercaseIndexCache = new WeakMap<object, Map<string, unknown>>();
+
+function findInsensitive<T>(obj: Record<string, T> | null | undefined, key: string): T | undefined {
+  if (!obj || !key) return undefined;
+  if (key in obj) return obj[key];
+  let index = lowercaseIndexCache.get(obj);
+  if (!index) {
+    index = new Map();
+    for (const [k, v] of Object.entries(obj)) {
+      const lowerKey = k.toLowerCase();
+      // Warn once at index-build time (not per-lookup) if two keys collide
+      // case-insensitively — a real data-quality signal from an upstream sync (e.g.
+      // models.dev returning both "OpenAI" and "openai" as distinct provider keys).
+      // Matches the pre-fix scan's silent first-match-wins behavior, just surfaced
+      // instead of swallowed.
+      if (index.has(lowerKey)) {
+        console.warn(
+          `[modelMetadataRegistry] findInsensitive: case-insensitive key collision on "${lowerKey}" — keeping first-seen value, later one discarded`
+        );
+        continue;
+      }
+      index.set(lowerKey, v);
+    }
+    lowercaseIndexCache.set(obj, index);
+  }
+  return index.get(key.toLowerCase()) as T | undefined;
+}
+
 function resolveCatalogPricing(
   provider: string | null,
   model: string | null
 ): Record<string, number> | null {
   if (!provider || !model) return null;
 
-  const findInsensitive = <T>(
-    obj: Record<string, T> | null | undefined,
-    key: string
-  ): T | undefined => {
-    if (!obj || !key) return undefined;
-    if (key in obj) return obj[key];
-    const lower = key.toLowerCase();
-    for (const [k, v] of Object.entries(obj)) {
-      if (k.toLowerCase() === lower) return v;
-    }
-    return undefined;
-  };
-
   // Prefer models.dev synced pricing when present; fall back to hardcoded defaults.
   try {
     const modelsDev = getModelsDevPricing() as Record<
diff --git a/src/lib/modelsDevSync.ts b/src/lib/modelsDevSync.ts
index 8c34135feb..32a6e180f6 100644
--- a/src/lib/modelsDevSync.ts
+++ b/src/lib/modelsDevSync.ts
@@ -18,7 +18,7 @@
  */
 
 import { getDbInstance } from "./db/core";
-import { invalidateDbCache } from "./db/readCache";
+import { invalidateDbCache, getModelCatalogCacheVersion } from "./db/readCache";
 import { backupDbFile } from "./db/backup";
 
 import {
@@ -193,10 +193,25 @@ function mapCapabilityRecord(record: Record<string, unknown>): ModelCapabilityEn
   };
 }
 
+// #8697: getModelsDevPricing() re-ran the SELECT + JSON.parse of ~180 blobs on
+// every call — called once per catalog model (up to ~6091x) instead of once per
+// request, freezing the whole server 41-54s on a cold /v1/models rebuild.
+// Memoized here, invalidated via the same modelCatalogCacheVersion signal
+// save/clearModelsDevPricing already bump through invalidateDbCache("pricing") —
+// reusing the existing pattern (getCachedRawProviderConnections et al. in
+// db/readCache.ts) instead of introducing a new invalidation mechanism.
+let pricingMemo: PricingByProvider | null = null;
+let pricingMemoVersion = -1; // -1: never equals a real cacheVersion (starts at 0), guarantees a miss on the first call
+
 /**
  * Read synced pricing from `models_dev_pricing` namespace.
  */
 export function getModelsDevPricing(): PricingByProvider {
+  const currentVersion = getModelCatalogCacheVersion();
+  if (pricingMemo !== null && pricingMemoVersion === currentVersion) {
+    return pricingMemo;
+  }
+
   const db = getDbInstance();
   const rows = db
     .prepare("SELECT key, value FROM key_value WHERE namespace = 'models_dev_pricing'")
@@ -213,6 +228,8 @@ export function getModelsDevPricing(): PricingByProvider {
       console.warn(`[MODELS_DEV] Corrupted pricing data for provider "${key}", skipping`);
     }
   }
+  pricingMemo = synced;
+  pricingMemoVersion = currentVersion;
   return synced;
 }
 
@@ -354,44 +371,26 @@ export function getSyncedCapability(
 ): ModelCapabilityEntry | null {
   if (!provider || !modelId) return null;
 
-  // Fast path: every provider is in the in-memory cache, skip SQLite entirely.
-  if (cachedCapabilitiesLoadedAll) {
-    const lookupCached = (p: string) => cachedCapabilities?.[p]?.[modelId] ?? null;
-    const directCached = lookupCached(provider);
-    if (directCached) return directCached;
-    const fallbacks = SYNCED_CAPABILITY_FALLBACK_ALIASES[provider];
-    if (fallbacks) {
-      for (const alt of fallbacks) {
-        const found = lookupCached(alt);
-        if (found) return found;
-      }
-    }
-    return null;
+  // #8697-adjacent: this used to hit SQLite with a per-model SELECT on every cold
+  // call, relying on some other caller (getSyncedCapabilities() with no args) to have
+  // already warmed the whole-table cache first — no such caller sits in the /v1/models
+  // catalog build path, so a cold rebuild ran one SQLite round-trip per model per call
+  // site instead of one bulk read for the whole rebuild. Self-warm here instead of
+  // depending on an external caller.
+  if (!cachedCapabilitiesLoadedAll) {
+    getSyncedCapabilities();
   }
 
-  // Cold path: hit SQLite. Prepare the statement once, reuse for every alias.
-  const db = getDbInstance();
-  ensureCapabilitiesTable();
-  const stmt = db.prepare(
-    "SELECT * FROM model_capabilities WHERE provider = ? AND model_id = ? LIMIT 1"
-  );
-  const lookupDb = (p: string): ModelCapabilityEntry | null => {
-    const row = stmt.get(p, modelId);
-    if (!row) return null;
-    return mapCapabilityRecord(toRecord(row));
-  };
-
-  const direct = lookupDb(provider);
-  if (direct) return direct;
-
+  const lookupCached = (p: string) => cachedCapabilities?.[p]?.[modelId] ?? null;
+  const directCached = lookupCached(provider);
+  if (directCached) return directCached;
   const fallbacks = SYNCED_CAPABILITY_FALLBACK_ALIASES[provider];
   if (fallbacks) {
     for (const alt of fallbacks) {
-      const found = lookupDb(alt);
+      const found = lookupCached(alt);
       if (found) return found;
     }
   }
-
   return null;
 }
 
diff --git a/src/shared/constants/modelSpecs.ts b/src/shared/constants/modelSpecs.ts
index 3128877453..c9f46e67d4 100644
--- a/src/shared/constants/modelSpecs.ts
+++ b/src/shared/constants/modelSpecs.ts
@@ -608,26 +608,83 @@ export const MODEL_SPECS: Record<string, ModelSpec> = {
   __default__: {},
 };
 
+// #8697-adjacent: getCanonicalModelSpecId() re-scanned Object.keys/entries(MODEL_SPECS)
+// up to 3 times per call (exact ci, alias ci, prefix) — the top hotspot in a full
+// catalog-rebuild profile once the pricing-path bottlenecks were fixed. MODEL_SPECS is
+// a static module constant (never mutated at runtime), so the lowercase index below is
+// built once, lazily, on first use and never invalidated. Iteration order for the
+// prefix-match candidates is preserved exactly (same Object.keys() insertion order) so
+// resolution outcomes for ambiguous prefixes are unchanged.
+let modelSpecIndex: {
+  exactCi: Map<string, string>;
+  aliasCi: Map<string, string>;
+  aliasExact: Map<string, string>;
+  prefixCandidates: Array<[lowerKey: string, canonical: string]>;
+} | null = null;
+
+function getModelSpecIndex() {
+  if (modelSpecIndex) return modelSpecIndex;
+  const exactCi = new Map<string, string>();
+  const aliasCi = new Map<string, string>();
+  const aliasExact = new Map<string, string>();
+  const prefixCandidates: Array<[string, string]> = [];
+  for (const [canonical, spec] of Object.entries(MODEL_SPECS)) {
+    const lowerCanonical = canonical.toLowerCase();
+    if (!exactCi.has(lowerCanonical)) exactCi.set(lowerCanonical, canonical);
+    for (const alias of spec.aliases || []) {
+      const lowerAlias = alias.toLowerCase();
+      if (!aliasCi.has(lowerAlias)) aliasCi.set(lowerAlias, canonical);
+      if (!aliasExact.has(alias)) aliasExact.set(alias, canonical);
+    }
+    if (canonical !== "__default__") prefixCandidates.push([lowerCanonical, canonical]);
+  }
+  modelSpecIndex = { exactCi, aliasCi, aliasExact, prefixCandidates };
+  return modelSpecIndex;
+}
+
+/**
+ * Exact + alias case-insensitive lookup only (no prefix phase) — shared by
+ * modelCapabilities.ts's getStaticSpecCanonicalModelId(), which tries multiple id
+ * candidates and never wanted prefix matching. Reuses the same lazy index as
+ * getCanonicalModelSpecId() below instead of each caller maintaining its own cache
+ * over the same static MODEL_SPECS table.
+ *
+ * Contract: returns `null` for `__default__` (never a real canonical id), for an
+ * unrecognized `modelId`, or for an empty string. Matching is case-insensitive on
+ * both the canonical id and its aliases; there is no prefix-matching phase (unlike
+ * getCanonicalModelSpecId() below) — callers that need prefix matching should use
+ * that function instead.
+ */
+export function findModelSpecIdByExactOrAlias(modelId: string): string | null {
+  const lower = modelId.toLowerCase();
+  const index = getModelSpecIndex();
+  const exactHit = index.exactCi.get(lower);
+  if (exactHit && exactHit !== "__default__") return exactHit;
+  const aliasHit = index.aliasCi.get(lower);
+  if (aliasHit && aliasHit !== "__default__") return aliasHit;
+  return null;
+}
+
 export function getCanonicalModelSpecId(modelId: string): string | null {
   if (MODEL_SPECS[modelId]) return modelId;
 
   // Case-insensitive lookups: upstream model ids are often capitalized
   // (e.g. "MiniMax-M2.7") while specs/aliases use lowercase ids (#3141).
   const lower = modelId.toLowerCase();
+  const index = getModelSpecIndex();
 
   // Exact match (case-insensitive)
-  for (const canonical of Object.keys(MODEL_SPECS)) {
-    if (canonical.toLowerCase() === lower) return canonical;
-  }
+  const exactHit = index.exactCi.get(lower);
+  if (exactHit) return exactHit;
 
   // Buscas por alias (case-insensitive)
-  for (const [canonical, spec] of Object.entries(MODEL_SPECS)) {
-    if (spec.aliases?.some((alias) => alias.toLowerCase() === lower)) return canonical;
-  }
+  const aliasHit = index.aliasCi.get(lower);
+  if (aliasHit) return aliasHit;
 
-  // Prefix matching (case-insensitive)
-  for (const key of Object.keys(MODEL_SPECS)) {
-    if (key !== "__default__" && lower.startsWith(key.toLowerCase())) return key;
+  // Prefix matching (case-insensitive) — same insertion-order iteration as before,
+  // first match wins.
+  for (const [lowerKey, canonical] of index.prefixCandidates) {
+    if (lower.startsWith(lowerKey)) return canonical;
   }
 
   return null;
@@ -721,9 +778,12 @@ export function capThinkingBudget(modelId: string, budget: number): number {
   return Math.min(budget, cap);
 }
 
+// #8697-adjacent: rescanned Object.entries(MODEL_SPECS) on every call, unconditionally
+// once per model in a catalog rebuild — verified 1:1 call ratio (no early
+// short-circuit). Case-sensitive exact match (Array.includes(), no .toLowerCase()) —
+// deliberately NOT reusing the case-insensitive aliasCi index above, which would
+// silently broaden matches and change behavior.
 export function resolveModelAlias(modelId: string): string {
-  for (const [canonical, spec] of Object.entries(MODEL_SPECS)) {
-    if (spec.aliases?.includes(modelId)) return canonical;
-  }
-  return modelId;
+  const hit = getModelSpecIndex().aliasExact.get(modelId);
+  return hit ?? modelId;
 }
diff --git a/tests/unit/catalog-pricing-lookup-index-8697.test.ts b/tests/unit/catalog-pricing-lookup-index-8697.test.ts
new file mode 100644
index 0000000000..4afab30799
--- /dev/null
+++ b/tests/unit/catalog-pricing-lookup-index-8697.test.ts
@@ -0,0 +1,105 @@
+import assert from "node:assert/strict";
+import { describe, it, before, after } from "node:test";
+import { enrichCatalogModelEntry } from "../../src/lib/modelMetadataRegistry.ts";
+import {
+  saveModelsDevPricing,
+  clearModelsDevPricing,
+  type PricingByProvider,
+} from "../../src/lib/modelsDevSync.ts";
+
+const PROVIDER_COUNT = 180;
+const MODELS_PER_PROVIDER = 34;
+const ITERATIONS = 500;
+
+describe("catalog pricing lookup index (#8697 second bottleneck — findInsensitive)", () => {
+  before(() => {
+    // Mixed-case keys force the case-insensitive fallback scan in
+    // findInsensitive() — mirrors real models.dev data where provider/model
+    // casing does not always match the catalog's, and a large provider count
+    // mirrors the ~180 synced providers from the #8697 profiling run.
+    const pricing: PricingByProvider = {};
+    for (let p = 0; p < PROVIDER_COUNT; p++) {
+      const providerKey = `Provider${p}`;
+      pricing[providerKey] = {};
+      for (let m = 0; m < MODELS_PER_PROVIDER; m++) {
+        pricing[providerKey][`Model${m}`] = { input: p + m * 0.01, output: p + m * 0.02 };
+      }
+    }
+    pricing.Openai = { "Gpt-4o": { input: 2.5, output: 10 } };
+    saveModelsDevPricing(pricing);
+  });
+
+  after(() => {
+    try {
+      clearModelsDevPricing();
+    } catch {
+      // ignore
+    }
+  });
+
+  it("resolves case-insensitive pricing correctly for every provider/model pair", () => {
+    const entry = enrichCatalogModelEntry({
+      id: "provider42/model7",
+      owned_by: "provider42",
+      root: "model7",
+    });
+    assert.ok(entry.pricing, "pricing should resolve via case-insensitive lookup");
+    assert.equal((entry.pricing as { input: number }).input, 42.07);
+  });
+
+  it("does not rescan the pricing tables per lookup (regression guard for O(providers*models) scans)", () => {
+    // `provider`/`gpt-4o` always resolve through the same fast metadata path
+    // (real registered provider) so both scenarios below pay an identical
+    // getCanonicalModelMetadata cost — isolating the delta to pricing
+    // resolution alone, independent of unrelated catalog-metadata overhead.
+    const entryWithPricingPreset = () =>
+      enrichCatalogModelEntry({
+        id: "openai/gpt-4o",
+        owned_by: "openai",
+        root: "gpt-4o",
+        pricing: { input: 1, output: 1 }, // nextEntry.pricing != null → resolveCatalogPricing() never runs
+      });
+    const entryNeedingPricingResolution = () =>
+      enrichCatalogModelEntry({
+        id: "openai/gpt-4o",
+        owned_by: "openai",
+        root: "gpt-4o",
+      });
+
+    // Warm up (index build, module init) outside the measured window.
+    entryWithPricingPreset();
+    entryNeedingPricingResolution();
+
+    const originalEntries = Object.entries;
+    let calls = 0;
+    Object.entries = function patchedEntries(...args: Parameters<typeof Object.entries>) {
+      calls++;
+      return originalEntries.apply(this, args as never);
+    } as typeof Object.entries;
+
+    let baselineCalls: number;
+    let withPricingCalls: number;
+    try {
+      calls = 0;
+      for (let i = 0; i < ITERATIONS; i++) entryWithPricingPreset();
+      baselineCalls = calls;
+
+      calls = 0;
+      for (let i = 0; i < ITERATIONS; i++) entryNeedingPricingResolution();
+      withPricingCalls = calls;
+    } finally {
+      Object.entries = originalEntries;
+    }
+
+    const delta = withPricingCalls - baselineCalls;
+    // Pre-fix: findInsensitive() called Object.entries() on every miss, twice per
+    // lookup (provider scan + model scan) → delta ≈ 2 * ITERATIONS. Indexed O(1)
+    // lookup: the index is built once per distinct object and reused, so delta
+    // stays a small constant regardless of ITERATIONS.
+    assert.ok(
+      delta < ITERATIONS,
+      `expected Object.entries() call delta to stay constant (not scale with ${ITERATIONS} ` +
+        `iterations), got delta=${delta} — findInsensitive() may have regressed to a linear scan per lookup`
+    );
+  });
+});
diff --git a/tests/unit/model-spec-lookup-index-8697.test.ts b/tests/unit/model-spec-lookup-index-8697.test.ts
new file mode 100644
index 0000000000..24618149f2
--- /dev/null
+++ b/tests/unit/model-spec-lookup-index-8697.test.ts
@@ -0,0 +1,53 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import { getCanonicalModelSpecId, getModelSpec } from "../../src/shared/constants/modelSpecs.ts";
+
+describe("model spec lookup index (#8697-adjacent — getCanonicalModelSpecId)", () => {
+  it("still resolves case-insensitive exact matches", () => {
+    // Real MODEL_SPECS entries — exercised via a mixed-case id, forcing the
+    // case-insensitive fallback the index covers.
+    const canonical = getCanonicalModelSpecId("GPT-5.6");
+    assert.ok(
+      canonical,
+      "expected a canonical id to resolve for a known model, case-insensitively"
+    );
+    assert.equal(getModelSpec("GPT-5.6"), getModelSpec(canonical!));
+  });
+
+  it("returns null for a genuinely unknown model id", () => {
+    assert.equal(getCanonicalModelSpecId("definitely-not-a-real-model-xyz-123"), null);
+  });
+
+  it("does not rescan MODEL_SPECS per lookup (regression guard for O(n) scans)", () => {
+    // Warm the lazy index outside the measured window.
+    getCanonicalModelSpecId("gpt-5.6");
+
+    const originalEntries = Object.entries;
+    const originalKeys = Object.keys;
+    let entriesCalls = 0;
+    let keysCalls = 0;
+    Object.entries = function patchedEntries(...args: Parameters<typeof Object.entries>) {
+      entriesCalls++;
+      return originalEntries.apply(this, args as never);
+    } as typeof Object.entries;
+    Object.keys = function patchedKeys(...args: Parameters<typeof Object.keys>) {
+      keysCalls++;
+      return originalKeys.apply(this, args as never);
+    } as typeof Object.keys;
+
+    try {
+      for (let i = 0; i < 500; i++) {
+        getCanonicalModelSpecId("gpt-5.6");
+      }
+    } finally {
+      Object.entries = originalEntries;
+      Object.keys = originalKeys;
+    }
+
+    // Pre-fix: every miss re-ran Object.keys()/Object.entries() up to 3x per call.
+    // Indexed: the lazy index is built once and reused, so no further
+    // Object.keys/entries calls should happen at all across 500 repeated lookups.
+    assert.equal(entriesCalls, 0, `expected 0 Object.entries() calls, got ${entriesCalls}`);
+    assert.equal(keysCalls, 0, `expected 0 Object.keys() calls, got ${keysCalls}`);
+  });
+});
diff --git a/tests/unit/models-dev-pricing-memoization-8697.test.ts b/tests/unit/models-dev-pricing-memoization-8697.test.ts
new file mode 100644
index 0000000000..9ab1c12277
--- /dev/null
+++ b/tests/unit/models-dev-pricing-memoization-8697.test.ts
@@ -0,0 +1,59 @@
+import assert from "node:assert/strict";
+import { describe, it, before, after, mock } from "node:test";
+import { getDbInstance } from "../../src/lib/db/core.ts";
+import {
+  getModelsDevPricing,
+  saveModelsDevPricing,
+  clearModelsDevPricing,
+  type PricingByProvider,
+} from "../../src/lib/modelsDevSync.ts";
+
+describe("getModelsDevPricing memoization (#8697)", () => {
+  before(() => {
+    const pricing: PricingByProvider = {
+      openai: {
+        "gpt-4o": { input: 2.5, output: 10 },
+      },
+    };
+    saveModelsDevPricing(pricing);
+  });
+
+  after(() => {
+    try {
+      clearModelsDevPricing();
+    } catch {
+      // ignore
+    }
+  });
+
+  it("hits the DB once for repeated reads within the same cache version", () => {
+    const db = getDbInstance();
+    const prepareSpy = mock.method(db, "prepare");
+    const callsBefore = prepareSpy.mock.calls.length;
+
+    getModelsDevPricing();
+    getModelsDevPricing();
+    getModelsDevPricing();
+
+    const callsAfter = prepareSpy.mock.calls.length;
+    prepareSpy.mock.restore();
+
+    // The N+1 bug re-runs the SELECT + JSON.parse on every call — memoized,
+    // 3 calls should cost at most 1 real DB round-trip (0 if a prior test
+    // already warmed the cache at the same version).
+    assert.ok(
+      callsAfter - callsBefore <= 1,
+      `expected at most 1 db.prepare() call across 3 reads, got ${callsAfter - callsBefore}`
+    );
+  });
+
+  it("returns fresh data after a write invalidates the cache", () => {
+    getModelsDevPricing(); // warm the cache
+    saveModelsDevPricing({
+      anthropic: { "claude-x": { input: 1, output: 2 } },
+    });
+    const pricing = getModelsDevPricing();
+    assert.ok(pricing.anthropic, "cache should reflect the write, not a stale snapshot");
+    assert.equal(pricing.anthropic["claude-x"].input, 1);
+  });
+});
diff --git a/tests/unit/resolve-model-alias-index-8697.test.ts b/tests/unit/resolve-model-alias-index-8697.test.ts
new file mode 100644
index 0000000000..f521014b89
--- /dev/null
+++ b/tests/unit/resolve-model-alias-index-8697.test.ts
@@ -0,0 +1,51 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import { resolveModelAlias } from "../../src/shared/constants/modelSpecs.ts";
+
+describe("resolveModelAlias lookup index (#8697-adjacent)", () => {
+  it("still resolves a known exact alias", () => {
+    // Real MODEL_SPECS alias, case-sensitive exact match.
+    assert.equal(resolveModelAlias("openai/gpt-5.6"), "gpt-5.6");
+  });
+
+  it("does not match a case-varied alias (case-sensitive semantics preserved)", () => {
+    // resolveModelAlias uses Array.includes(), never .toLowerCase() — a case-varied
+    // input must NOT resolve, unlike the case-insensitive getCanonicalModelSpecId().
+    assert.equal(resolveModelAlias("OpenAI/GPT-5.6"), "OpenAI/GPT-5.6");
+  });
+
+  it("returns the input unchanged for an unknown alias", () => {
+    assert.equal(
+      resolveModelAlias("definitely-not-a-real-alias-xyz"),
+      "definitely-not-a-real-alias-xyz"
+    );
+  });
+
+  it("does not rescan MODEL_SPECS per call (regression guard for O(n) scans)", () => {
+    // Warm up outside the measured window.
+    resolveModelAlias("openai/gpt-5.6");
+
+    const originalEntries = Object.entries;
+    let calls = 0;
+    Object.entries = function patchedEntries(...args: Parameters<typeof Object.entries>) {
+      calls++;
+      return originalEntries.apply(this, args as never);
+    } as typeof Object.entries;
+
+    try {
+      for (let i = 0; i < 500; i++) {
+        resolveModelAlias("openai/gpt-5.6");
+      }
+    } finally {
+      Object.entries = originalEntries;
+    }
+
+    // Pre-fix: every call re-ran Object.entries(MODEL_SPECS). Indexed: the lazy
+    // index is built once and reused, so no further Object.entries calls happen.
+    assert.equal(
+      calls,
+      0,
+      `expected 0 Object.entries() calls across 500 repeated lookups, got ${calls}`
+    );
+  });
+});
diff --git a/tests/unit/reverse-models-dev-providers-8697.test.ts b/tests/unit/reverse-models-dev-providers-8697.test.ts
new file mode 100644
index 0000000000..1e3c762e1d
--- /dev/null
+++ b/tests/unit/reverse-models-dev-providers-8697.test.ts
@@ -0,0 +1,47 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import { getResolvedModelCapabilities } from "../../src/lib/modelCapabilities.ts";
+import { MODELS_DEV_PROVIDER_MAP } from "../../src/lib/modelsDevSync/transform.ts";
+
+describe("reverseModelsDevProviders memoization (#8697-adjacent)", () => {
+  it("stays correct across repeated calls for the same provider", () => {
+    // codex/claude only list their alias (cx/cc) in MODELS_DEV_PROVIDER_MAP — exercises
+    // the reverse-lookup fallback this function builds (#8429).
+    const first = getResolvedModelCapabilities({ provider: "codex", model: "gpt-5.6" });
+    const second = getResolvedModelCapabilities({ provider: "codex", model: "gpt-5.6" });
+    assert.deepEqual(first, second, "memoized reverse-provider lookup must not change results");
+  });
+
+  it("does not rescan MODELS_DEV_PROVIDER_MAP per call (regression guard for O(n) scans)", () => {
+    // Warm up outside the measured window.
+    getResolvedModelCapabilities({ provider: "codex", model: "gpt-5.6" });
+
+    // getResolvedModelCapabilities' wider call chain legitimately calls Object.entries()
+    // on unrelated objects (e.g. once per call, elsewhere in the chain) — count only calls
+    // targeting MODELS_DEV_PROVIDER_MAP specifically, the object reverseModelsDevProviders()
+    // scans, to isolate this fix's contribution precisely.
+    const originalEntries = Object.entries;
+    let mapScans = 0;
+    Object.entries = function patchedEntries(...args: Parameters<typeof Object.entries>) {
+      if (args[0] === MODELS_DEV_PROVIDER_MAP) mapScans++;
+      return originalEntries.apply(this, args as never);
+    } as typeof Object.entries;
+
+    try {
+      for (let i = 0; i < 300; i++) {
+        getResolvedModelCapabilities({ provider: "codex", model: "gpt-5.6" });
+      }
+    } finally {
+      Object.entries = originalEntries;
+    }
+
+    // Pre-fix: reverseModelsDevProviders() rescanned Object.entries(MODELS_DEV_PROVIDER_MAP)
+    // on every call → mapScans would be ~300. Memoized by provider key: 0 scans once the
+    // "codex" entry is cached (the warm-up call above already populated it).
+    assert.equal(
+      mapScans,
+      0,
+      `expected 0 Object.entries(MODELS_DEV_PROVIDER_MAP) scans across 300 repeated calls, got ${mapScans}`
+    );
+  });
+});
diff --git a/tests/unit/synced-capability-warmup-8697.test.ts b/tests/unit/synced-capability-warmup-8697.test.ts
new file mode 100644
index 0000000000..90045d8cdd
--- /dev/null
+++ b/tests/unit/synced-capability-warmup-8697.test.ts
@@ -0,0 +1,32 @@
+import assert from "node:assert/strict";
+import { describe, it, mock } from "node:test";
+import { getDbInstance } from "../../src/lib/db/core.ts";
+import { getSyncedCapability } from "../../src/lib/modelsDevSync.ts";
+
+describe("getSyncedCapability warm-up (#8697-adjacent)", () => {
+  it("does not run a DB round-trip per distinct model lookup (regression guard for the missing bulk warm-up)", () => {
+    const db = getDbInstance();
+    const prepareSpy = mock.method(db, "prepare");
+    const callsBefore = prepareSpy.mock.calls.length;
+
+    // A catalog rebuild calls getSyncedCapability() once per distinct model — this
+    // used to run one SQLite SELECT per call on a cold cache (no warm-up caller sits
+    // in the /v1/models build path). Self-warmed, only the one-time bulk load (plus
+    // its CREATE TABLE IF NOT EXISTS guard) should touch the DB, regardless of how
+    // many distinct models are looked up afterward.
+    const N = 200;
+    for (let i = 0; i < N; i++) {
+      getSyncedCapability("openai", `synthetic-model-${i}`);
+    }
+
+    const callsAfter = prepareSpy.mock.calls.length;
+    prepareSpy.mock.restore();
+
+    assert.ok(
+      callsAfter - callsBefore <= 2,
+      `expected at most 2 db.prepare() calls (bulk load + table guard) across ${N} distinct ` +
+        `model lookups, got ${callsAfter - callsBefore} — getSyncedCapability() may have regressed ` +
+        `to a per-model SQLite round-trip`
+    );
+  });
+});