diff --git a/changelog.d/fixes/12792-pricing-valuation.md b/changelog.d/fixes/12792-pricing-valuation.md
new file mode 100644
index 0000000000..7eea070726
--- /dev/null
+++ b/changelog.d/fixes/12792-pricing-valuation.md
@@ -0,0 +1 @@
+- **fix(routing):** off-table models check the free-model catalog before inheriting premium prices, latency bootstraps from the observed pool median, async tiers read live database pricing with a 90-day freshness gate, and the tier cache invalidates on every pricing write ([#12792](https://github.com/diegosouzapw/OmniRoute/pull/12792)) — thanks @maxmad64bis
diff --git a/config/quality/file-size-baseline.json b/config/quality/file-size-baseline.json
index 49c5f74724..41d80fbff6 100644
--- a/config/quality/file-size-baseline.json
+++ b/config/quality/file-size-baseline.json
@@ -467,7 +467,8 @@
     "src/sse/handlers/chat.ts": 2458,
     "src/sse/services/auth.ts": 3450,
     "tests/unit/account-fallback-service.test.ts": 2453,
-    "tests/unit/provider-validation-specialty.test.ts": 4656
+    "tests/unit/provider-validation-specialty.test.ts": 4656,
+    "open-sse/services/autoCombo/virtualFactory.ts": 1207
   },
   "_rebaseline_base_2026_08_10_proxyfetch": "Base-red fix (green-prs sweep, issue #9985): open-sse/utils/proxyFetch.ts 1207 > cap 1000 — new proxied-TLS fetch helper introduced by the Fal reference-image work. Owner-authorized quick rebaseline to green; structural slim tracked for v3.9.0.",
   "_rebaseline_2026_07_27_v3849_train2": "Merge-train 2 (7 PRs) — owner-approved 2026-07-27. Single entry: chatCore.ts 4955->5006 (#8595, Responses multi-turn image compaction before the context hard-reject). Genuine irreducible growth at the existing compaction chokepoint in handleChatCore — the PR adds a last-resort retry against the concrete budget plus the estimateFinalInputTokens helper, both wired at the pre-existing call site rather than a new branch. Covered by tests/unit/8560-responses-image-compaction.test.ts (4 tests).",
@@ -652,5 +653,6 @@
   "_rebaseline_2026_09_03_houminxi_combo_stacked": "Leva HouMinXi (#12624 #12626 #12632 #12637): open-sse/services/combo.ts 4075->4080 (+5), medido no tip com os quatro mergeados. Cada PR registrou o proprio crescimento contra o tip de onde forkou (o #12637 ja subira o cap para 4075); as 5 linhas restantes so aparecem quando eles empilham, porque mais de um toca o mesmo chokepoint de scoring reset-aware em combo.ts. Fiacao em ponto existente, sem extracao possivel sem partir a funcao de selecao de alvos. Coberto por combo-strategies e reset-aware-request-scope-12600 (119/119 focados na leva).",
   "_rebaseline_2026_09_04_12641_continuation_effective_input": "PR #12641 crescimento proprio: src/sse/handlers/chat.ts 2450->2454 (+4). A continuacao por previous_response_id encadeava a partir de clientRawRequest.body.input, que e capturado ANTES da reconstrucao do proprio chat.ts; quando o turno anterior ja era uma continuacao, esse campo guarda so o delta do cliente, e o erro se acumulava a cada salto ate a reconstrucao virar itens de tool sem prefixo. Persistir o input EFETIVO exige as linhas no ponto onde a reconstrucao termina, dentro do fluxo de despacho. Coberto por tests/unit/responses-continuation-store.test.ts (22/22 focados na leva).",
   "_rebaseline_2026_09_05_12671_combos_usage_guide_external_store": "combos/page.tsx 5018 -> 5066: #12671 replaces the effect-based localStorage read with useSyncExternalStore; the +48 lines are the store helpers (subscribe/getSnapshot/getServerSnapshot/emit) hoisted to module scope, which is the sanctioned shape and what let the react-hooks/set-state-in-effect suppression be dropped.",
-  "_rebaseline_2026_09_07_chatcore_nonstreaming_regression_fixes": "Own growth: open-sse/handlers/chatCore.ts 5984->6021 (+37). Two of my own PRs on top of #12867: #12963 pins the ok variant of the non-streaming leg result in its own binding (the discriminated-union narrowing was lost across the tool-loop reassignment, 13 TS2339 under tsconfig.typecheck-api.json), and #12990 restores four behaviours the same refactor dropped — abort classification through isLocalStreamLifecycleError, the omitted synthetic clientResponse, the claudePromptCacheLogMeta rebuild on the leg path, and the lazy fail-closed fence identity. Irreducible at the existing chokepoints: each edit sits where chatCore already owns the decision, and the helpers themselves (nonStreamingProviderLeg.ts, serverOwnedToolLoopWire.ts) are under cap. Covered by tests/unit/chatcore-translation-paths.test.ts (72/74; the 2 open are issue #13043)."
+  "_rebaseline_2026_09_07_chatcore_nonstreaming_regression_fixes": "Own growth: open-sse/handlers/chatCore.ts 5984->6021 (+37). Two of my own PRs on top of #12867: #12963 pins the ok variant of the non-streaming leg result in its own binding (the discriminated-union narrowing was lost across the tool-loop reassignment, 13 TS2339 under tsconfig.typecheck-api.json), and #12990 restores four behaviours the same refactor dropped — abort classification through isLocalStreamLifecycleError, the omitted synthetic clientResponse, the claudePromptCacheLogMeta rebuild on the leg path, and the lazy fail-closed fence identity. Irreducible at the existing chokepoints: each edit sits where chatCore already owns the decision, and the helpers themselves (nonStreamingProviderLeg.ts, serverOwnedToolLoopWire.ts) are under cap. Covered by tests/unit/chatcore-translation-paths.test.ts (72/74; the 2 open are issue #13043).",
+  "_rebaseline_2026_09_07_virtualfactory_crosses_the_new_file_cap": "open-sse/services/autoCombo/virtualFactory.ts crosses the 1200 new-file cap for the first time (1187 on the pre-wave tip, 1207 here). Growth is spread across the routing/free-tier wave, not one extractable block: #12794 feeds observed breaker state and model quality into snapshot scoring instead of neutral constants, #12792 adds the reliability factor the snapshot path was still ignoring and the pooled-latency bootstrap, and #12744 tightens the free-model predicate the factory consumes. FROZEN RATHER THAN SPLIT, deliberately, and this is debt: two cohesive extraction candidates are ready when someone owns the move — computeSnapshotWeights (~85 lines) and the credential-eligibility group hasUsableOAuthToken/hasProviderSpecificSessionData/isKeylessEligibleConnection/hasUsableConnectionCredential (~70 lines). Either alone clears 1200 from here. Splitting three contributors' just-merged work mid-batch was the larger risk."
 }
diff --git a/open-sse/services/__tests__/manifestAdapter.test.ts b/open-sse/services/__tests__/manifestAdapter.test.ts
index e9430851a4..5652888741 100644
--- a/open-sse/services/__tests__/manifestAdapter.test.ts
+++ b/open-sse/services/__tests__/manifestAdapter.test.ts
@@ -22,8 +22,8 @@ function makeTarget(provider: string, model: string): ResolvedComboTarget {
 
 describe("ManifestAdapter", () => {
   describe("generateRoutingHints - trivial query", () => {
-    it("returns prefer-free modifier for greeting", () => {
-      const hints = generateRoutingHints([], {
+    it("returns prefer-free modifier for greeting", async () => {
+      const hints = await generateRoutingHints([], {
         messages: [{ content: "Hello" }],
       });
       expect(hints.strategyModifier).toBe("prefer-free");
@@ -32,8 +32,8 @@ describe("ManifestAdapter", () => {
   });
 
   describe("generateRoutingHints - expert query", () => {
-    it("returns a valid modifier for complex input", () => {
-      const hints = generateRoutingHints([], {
+    it("returns a valid modifier for complex input", async () => {
+      const hints = await generateRoutingHints([], {
         messages: [
           {
             content:
@@ -47,25 +47,25 @@ describe("ManifestAdapter", () => {
   });
 
   describe("generateRoutingHints - target classification", () => {
-    it("marks free provider as eligible for trivial query", () => {
+    it("marks free provider as eligible for trivial query", async () => {
       const targets = [makeTarget("kiro", "claude-sonnet-4.5")];
-      const hints = generateRoutingHints(targets, {
+      const hints = await generateRoutingHints(targets, {
         messages: [{ content: "Hi" }],
       });
       expect(hints.eligibleTargets.length).toBeGreaterThanOrEqual(0);
     });
 
-    it("handles empty targets array gracefully", () => {
-      const hints = generateRoutingHints([], {
+    it("handles empty targets array gracefully", async () => {
+      const hints = await generateRoutingHints([], {
         messages: [{ content: "Hello" }],
       });
       expect(hints.eligibleTargets.length).toBe(0);
       expect(hints.underqualifiedTargets.length).toBe(0);
     });
 
-    it("classifies mixed targets for simple query", () => {
+    it("classifies mixed targets for simple query", async () => {
       const targets = [makeTarget("kiro", "claude-sonnet-4.5"), makeTarget("openai", "gpt-4o")];
-      const hints = generateRoutingHints(targets, {
+      const hints = await generateRoutingHints(targets, {
         messages: [{ content: "Hello" }],
       });
       expect(hints.eligibleTargets.length).toBeGreaterThanOrEqual(0);
@@ -73,20 +73,20 @@ describe("ManifestAdapter", () => {
   });
 
   describe("compareByCostEffectiveness", () => {
-    it("takes 3 arguments and returns a number", () => {
+    it("takes 3 arguments and returns a number", async () => {
       const a = makeTarget("deepseek", "deepseek-chat");
       const b = makeTarget("openai", "gpt-4o");
-      const hints = generateRoutingHints([a, b], {
+      const hints = await generateRoutingHints([a, b], {
         messages: [{ content: "Test" }],
       });
       const result = compareByCostEffectiveness(a, b, hints);
       expect(typeof result).toBe("number");
     });
 
-    it("returns negative when a is cheaper than b", () => {
+    it("returns negative when a is cheaper than b", async () => {
       const a = makeTarget("deepseek", "deepseek-chat");
       const b = makeTarget("openai", "gpt-4o");
-      const hints = generateRoutingHints([a, b], {
+      const hints = await generateRoutingHints([a, b], {
         messages: [{ content: "Test" }],
       });
       const result = compareByCostEffectiveness(a, b, hints);
@@ -115,16 +115,16 @@ describe("ManifestAdapter", () => {
   });
 
   describe("edge cases", () => {
-    it("handles empty targets array", () => {
-      const hints = generateRoutingHints([], {
+    it("handles empty targets array", async () => {
+      const hints = await generateRoutingHints([], {
         messages: [{ content: "Hello" }],
       });
       expect(hints.eligibleTargets.length).toBe(0);
       expect(hints.underqualifiedTargets.length).toBe(0);
     });
 
-    it("returns valid hints structure with no targets", () => {
-      const hints = generateRoutingHints([], {
+    it("returns valid hints structure with no targets", async () => {
+      const hints = await generateRoutingHints([], {
         messages: [{ content: "Test" }],
       });
       expect("specificityLevel" in hints).toBe(true);
diff --git a/open-sse/services/autoCombo/complexityRouter.ts b/open-sse/services/autoCombo/complexityRouter.ts
index 602fb1450b..5b49adf483 100644
--- a/open-sse/services/autoCombo/complexityRouter.ts
+++ b/open-sse/services/autoCombo/complexityRouter.ts
@@ -75,11 +75,11 @@ export function classifyRequestComplexity(input: RuleInput): ComplexityClassific
  * failure — fail-open, so scoring stays tier-neutral. Extracted from combo.ts to
  * keep the complexity-routing logic in one module.
  */
-export function buildComplexityRoutingHint(
+export async function buildComplexityRoutingHint(
   modelTargets: Parameters<typeof generateRoutingHints>[0],
   body: { messages?: unknown; tools?: unknown; model?: unknown } | null | undefined,
   log: { info: (tag: string, message: string) => void }
-): RoutingHint | null {
+): Promise<RoutingHint | null> {
   try {
     const ruleInput = {
       messages: Array.isArray(body?.messages)
@@ -92,7 +92,7 @@ export function buildComplexityRoutingHint(
         : undefined,
       model: typeof body?.model === "string" ? body.model : undefined,
     };
-    const hint = generateRoutingHints(modelTargets, ruleInput);
+    const hint = await generateRoutingHints(modelTargets, ruleInput);
     // Tool-use escalation: floor the recommended tier at "cheap" so scoring
     // favors function-calling-reliable models for agentic requests.
     const classification = classifyRequestComplexity(ruleInput);
diff --git a/open-sse/services/autoCombo/engine.ts b/open-sse/services/autoCombo/engine.ts
index 9e91552f36..07a94e2dc4 100644
--- a/open-sse/services/autoCombo/engine.ts
+++ b/open-sse/services/autoCombo/engine.ts
@@ -39,6 +39,7 @@ export interface AutoComboConfig {
    *     silently overspending.
    */
   budgetFallback?: "cheapest" | "strict";
+  estimatedInputTokens?: number; // tokens the budget is computed against (default 1000)
   explorationRate: number; // 0.05 = 5% exploratory
   /** If set, RouterStrategy name to use for selection ('rules' | 'cost' | 'latency') */
   routerStrategy?: string;
@@ -311,9 +312,13 @@ export function selectProvider(
     for (const c of candidates) {
       costMap.set(`${c.provider}\0${c.model}`, c.costPer1MTokens);
     }
+    const estimatedTokens =
+      Number.isFinite(config.estimatedInputTokens) && config.estimatedInputTokens! > 0
+        ? config.estimatedInputTokens!
+        : 1000;
     const estimatedCostFor = (s: ScoredProvider) => {
       const cost = costMap.get(`${s.provider}\0${s.model}`) ?? 0;
-      return (cost / 1_000_000) * 1000;
+      return (cost / 1_000_000) * estimatedTokens;
     };
     if (estimatedCostFor(selected) > config.budgetCap) {
       const budgetOk = candidates_.filter((s) => estimatedCostFor(s) <= config.budgetCap!);
diff --git a/open-sse/services/autoCombo/scoring.ts b/open-sse/services/autoCombo/scoring.ts
index 295d5379b3..a9977ccba5 100644
--- a/open-sse/services/autoCombo/scoring.ts
+++ b/open-sse/services/autoCombo/scoring.ts
@@ -293,6 +293,14 @@ export function computePoolMaxima(pool: ProviderCandidate[]): PoolMaxima {
  * `speedRanking.ts` so both consumers of the same signal agree, including on
  * garbage input.
  */
+/** Reliability factor from an observed failure (else error) rate; absent reads fully reliable. */
+export function reliabilityFactor(candidate: {
+  failureRate?: number | null;
+  errorRate?: number | null;
+}): number {
+  return clamp01(1 - boundedRate(candidate.failureRate ?? candidate.errorRate));
+}
+
 function boundedRate(value: number | null | undefined): number {
   if (typeof value !== "number" || !Number.isFinite(value) || value < 0) return 0;
   return Math.min(1, value);
@@ -341,7 +349,7 @@ export function calculateFactors(
     // bounded BEFORE the subtraction, exactly as `toBoundedRate` does there --
     // `clamp01(1 - NaN)` would be 0, i.e. "fails every call", which is the
     // opposite of what corrupt telemetry should mean.
-    reliability: clamp01(1 - boundedRate(candidate.failureRate ?? candidate.errorRate)),
+    reliability: reliabilityFactor(candidate),
   };
 }
 
diff --git a/open-sse/services/autoCombo/virtualFactory.ts b/open-sse/services/autoCombo/virtualFactory.ts
index f14cec5cd4..51f0be6606 100644
--- a/open-sse/services/autoCombo/virtualFactory.ts
+++ b/open-sse/services/autoCombo/virtualFactory.ts
@@ -1,6 +1,6 @@
 import { AutoComboConfig } from "./engine";
 import { MODE_PACKS } from "./modePacks";
-import { DEFAULT_WEIGHTS, ScoringWeights } from "./scoring";
+import { DEFAULT_WEIGHTS, reliabilityFactor, ScoringWeights } from "./scoring";
 import { getCachedProviderConnections } from "@/lib/db/readCache";
 import { getSettings } from "@/lib/db/settings";
 import { getProviderRegistry } from "./providerRegistryAccessor";
@@ -28,6 +28,7 @@ import {
 import { classifyTier } from "../tierResolver";
 import { getQualityScore } from "../routing/quality.ts";
 import { readBreakerStates, snapshotHealthFactor, type BreakerState } from "./snapshotBreaker.ts";
+import { resolveVirtualCost } from "../providerCostData";
 import type { AutoVariant } from "./autoPrefix";
 import { buildFamilyCandidateFilter, type ModelFamily } from "./modelFamily";
 import { getHiddenModelsByProvider } from "@/models";
@@ -107,6 +108,8 @@ export interface VirtualAutoComboCandidate {
   model: string;
   modelStr: string; // e.g., 'openai/gpt-4o'
   costPer1MTokens: number; // from providerRegistry
+  /** Observed failure rate 0..1 when known; null/absent reads fully reliable. */
+  failureRate?: number | null;
   /** Build-local capability snapshot. Runtime calls rebuild it; catalog entries reuse it. */
   resolvedContextLength?: number | null;
   resolvedMaxOutputTokens?: number | null;
@@ -444,7 +447,7 @@ function getNoAuthCandidates(
         connectionId: SYNTHETIC_NOAUTH_CONNECTION_ID,
         model: modelId,
         modelStr: `${routingPrefix}/${modelId}`,
-        costPer1MTokens: 0,
+        costPer1MTokens: resolveVirtualCost(providerId, modelId),
       });
     }
   }
@@ -726,7 +729,7 @@ export async function prepareVirtualAutoComboInputs(
         allowedConnectionIds,
         model: modelId,
         modelStr: `${providerId}/${modelId}`,
-        costPer1MTokens: 0, // Not used in virtual auto-combo (LKGP uses session stickiness)
+        costPer1MTokens: resolveVirtualCost(providerId, modelId),
       });
     }
   }
@@ -893,6 +896,10 @@ export function computeSnapshotWeights(
     // (no runtime data at snapshot time, so equal baseline)
     if (weights.latencyInv > 0) score += weights.latencyInv * 0.5;
 
+    // reliability (#12792): the same failure-rate factor as scoring.ts; absent reads as
+    // fully reliable. The snapshot path was the only one still ignoring it.
+    if (weights.reliability > 0) score += weights.reliability * reliabilityFactor(c);
+
     // health: build-time breaker state (OPEN 0, CLOSED 1, else neutral 0.5); quota stays neutral
     score += weights.health * snapshotHealthFactor(breakerByProvider, c.provider);
     score += weights.quota * 0.5;
diff --git a/open-sse/services/combo.ts b/open-sse/services/combo.ts
index 61236649fd..acc72de4ae 100644
--- a/open-sse/services/combo.ts
+++ b/open-sse/services/combo.ts
@@ -251,6 +251,63 @@ function getBootstrapLatencyMs(modelId: string): number {
   return DEFAULT_MODEL_P95_MS[normalized] ?? 1500;
 }
 
+export function poolMedianP95Ms(
+  stats: Record<string, { p95LatencyMs?: unknown }>
+): number | undefined {
+  const vals = Object.values(stats)
+    .map((st) => Number(st?.p95LatencyMs))
+    .filter((v) => Number.isFinite(v) && v > 0)
+    .sort((a, b) => a - b);
+  return vals.length ? vals[(vals.length - 1) >> 1] : undefined;
+}
+
+const BOOTSTRAP_WARN_WINDOW_MS = 3600_000;
+export let bootstrapLatencyHits = 0; // exported for testability (reset in tests)
+export let bootstrapLatencyTotal = 0;
+let bootstrapWarnedAt = 0;
+export function resetBootstrapCounters(): void {
+  bootstrapLatencyHits = 0;
+  bootstrapLatencyTotal = 0;
+  bootstrapWarnedAt = 0;
+}
+export function bootstrapMs(model: string, poolMedian: number | undefined): number {
+  bootstrapLatencyTotal++;
+  const table = DEFAULT_MODEL_P95_MS[String(model || "").toLowerCase()];
+  if (table !== undefined) return table;
+  bootstrapLatencyHits++;
+  return poolMedian ?? 1500;
+}
+
+// Pure and testable without timers: the throttled 1h warn + cold-start exemption live here.
+export function shouldWarnBootstrap(
+  hits: number,
+  total: number,
+  hasStats: boolean,
+  now: number,
+  lastWarn: number
+): boolean {
+  if (!hasStats || total === 0) return false;
+  if (hits / total <= 0.3) return false;
+  return now - lastWarn >= BOOTSTRAP_WARN_WINDOW_MS;
+}
+
+function maybeWarnBootstrapDominant(hasStats: boolean): void {
+  if (
+    !shouldWarnBootstrap(
+      bootstrapLatencyHits,
+      bootstrapLatencyTotal,
+      hasStats,
+      Date.now(),
+      bootstrapWarnedAt
+    )
+  )
+    return;
+  bootstrapWarnedAt = Date.now();
+  console.warn(
+    `[combo] bootstrap latency dominant (${bootstrapLatencyHits}/${bootstrapLatencyTotal}) — scoring runs on guesses`
+  );
+}
+
 export async function buildAutoCandidates(
   targets: ResolvedComboTarget[],
   comboName: string,
@@ -278,6 +335,8 @@ export async function buildAutoCandidates(
   } catch {
     // keep empty stats — auto-combo will use runtime + bootstrap signals
   }
+  const poolMedian = poolMedianP95Ms(historicalLatencyStats);
+  const hasStats = Object.keys(historicalLatencyStats).length > 0;
 
   const uniqueProviders = Array.from(
     new Set(
@@ -364,10 +423,10 @@ export async function buildAutoCandidates(
       const p95LatencyMs = hasHistoricalSignal
         ? Number.isFinite(historicalP95Latency) && historicalP95Latency > 0
           ? historicalP95Latency
-          : getBootstrapLatencyMs(model)
+          : bootstrapMs(model, poolMedian)
         : Number.isFinite(avgLatency) && avgLatency > 0
           ? avgLatency
-          : getBootstrapLatencyMs(model);
+          : bootstrapMs(model, poolMedian);
 
       const errorRate = hasHistoricalSignal
         ? Number.isFinite(historicalSuccessRate) &&
@@ -516,6 +575,7 @@ export async function buildAutoCandidates(
 
   // Filter out candidates whose model is hidden by the user in the dashboard,
   // then drop vendor-retired ids so auto-combo cannot pick them (#11625).
+  maybeWarnBootstrapDominant(hasStats);
   return rejectRetiredAutoComboCandidates(
     candidates.filter((c) => {
       const hiddenModels = hiddenModelsMap.get(c.provider);
diff --git a/open-sse/services/combo/applyStrategyOrdering.ts b/open-sse/services/combo/applyStrategyOrdering.ts
index 405b042863..a0dc11ebc0 100644
--- a/open-sse/services/combo/applyStrategyOrdering.ts
+++ b/open-sse/services/combo/applyStrategyOrdering.ts
@@ -157,7 +157,7 @@ export async function applyStrategyOrdering(
     orderedTargets = await sortTargetsByCost(orderedTargets);
     if (config.manifestRouting === true) {
       try {
-        const manifestHint = generateRoutingHints(
+        const manifestHint = await generateRoutingHints(
           orderedTargets.filter((t) => t.kind === "model"),
           {
             messages: Array.isArray(body?.messages)
diff --git a/open-sse/services/combo/resolveAutoStrategy.ts b/open-sse/services/combo/resolveAutoStrategy.ts
index 9882f2da4e..134f8100b6 100644
--- a/open-sse/services/combo/resolveAutoStrategy.ts
+++ b/open-sse/services/combo/resolveAutoStrategy.ts
@@ -315,7 +315,7 @@ export async function resolveAutoStrategyOrder(
   // specificityMatch favor candidates whose tier matches the request.
   const autoManifestHint: RoutingHint | null =
     config.complexityAwareRouting === true
-      ? buildComplexityRoutingHint(
+      ? await buildComplexityRoutingHint(
           eligibleTargets.filter((t) => t.kind === "model"),
           body,
           log
@@ -408,6 +408,7 @@ export async function resolveAutoStrategyOrder(
             modePack,
             budgetCap,
             budgetFallback,
+            estimatedInputTokens,
             explorationRate,
           },
           routableCandidates,
diff --git a/open-sse/services/manifestAdapter.ts b/open-sse/services/manifestAdapter.ts
index a0528e48be..5ca43aa78a 100644
--- a/open-sse/services/manifestAdapter.ts
+++ b/open-sse/services/manifestAdapter.ts
@@ -1,7 +1,7 @@
 import type { TierAssignment, ProviderTier } from "./tierTypes";
 import { PROVIDER_TIER } from "./tierTypes";
 import type { SpecificityResult, SpecificityLevel } from "./specificityTypes";
-import { classifyTier } from "./tierResolver";
+import { classifyTier, classifyTierAsync } from "./tierResolver";
 import {
   analyzeSpecificity,
   getSpecificityLevel,
@@ -29,17 +29,24 @@ export interface RoutingHint {
   strategyModifier: StrategyModifier;
 }
 
-export function generateRoutingHints(
+export async function generateRoutingHints(
   targets: ResolvedComboTarget[],
   input: RuleInput
-): RoutingHint {
+): Promise<RoutingHint> {
   const tierAssignments = new Map<string, TierAssignment>();
-  for (const target of targets) {
-    if (target.kind !== "model") continue;
-    const key = `${target.provider}::${target.modelStr}`;
-    if (!tierAssignments.has(key)) {
-      tierAssignments.set(key, classifyTier(target.provider, target.modelStr));
-    }
+  const modelTargets = targets.filter((t) => t.kind === "model");
+  const settled = await Promise.all(
+    modelTargets.map(async (target) => {
+      const key = `${target.provider}::${target.modelStr}`;
+      try {
+        return [key, await classifyTierAsync(target.provider, target.modelStr)] as const;
+      } catch {
+        return [key, classifyTier(target.provider, target.modelStr)] as const;
+      }
+    })
+  );
+  for (const [key, assignment] of settled) {
+    if (!tierAssignments.has(key)) tierAssignments.set(key, assignment);
   }
 
   const specificity = analyzeSpecificity(input);
diff --git a/open-sse/services/providerCostData.ts b/open-sse/services/providerCostData.ts
index 8464635c92..675514d816 100644
--- a/open-sse/services/providerCostData.ts
+++ b/open-sse/services/providerCostData.ts
@@ -1,4 +1,5 @@
 import { getPricingForModel as getDefaultPricingForModel } from "@/shared/constants/pricing";
+import { isFreeModel } from "@/shared/utils/freeModels";
 import type { TierConfig } from "./tierTypes";
 
 export interface ModelPricing {
@@ -38,14 +39,16 @@ export const KNOWN_MODEL_PRICING: Record<string, ModelPricing> = {
 };
 
 export function getModelPricing(provider: string, model: string): ModelPricing {
-  const providerKey = `${provider}/${model}`.toLowerCase();
-  if (KNOWN_MODEL_PRICING[providerKey]) {
-    return KNOWN_MODEL_PRICING[providerKey];
-  }
-  const providerPricing = getDefaultPricingForModel(provider, model);
-  if (providerPricing) {
-    const inputCostPer1M = Number(providerPricing.input);
-    const outputCostPer1M = Number(providerPricing.output);
+  const normalized = String(model || "")
+    .split("/")
+    .pop()!
+    .toLowerCase();
+  const providerHit = KNOWN_MODEL_PRICING[`${provider}/${normalized}`.toLowerCase()];
+  if (providerHit) return providerHit;
+  const defaultPricing = getDefaultPricingForModel(provider, model);
+  if (defaultPricing) {
+    const inputCostPer1M = Number(defaultPricing.input);
+    const outputCostPer1M = Number(defaultPricing.output);
     if (Number.isFinite(inputCostPer1M) && Number.isFinite(outputCostPer1M)) {
       return {
         inputCostPer1M,
@@ -54,13 +57,18 @@ export function getModelPricing(provider: string, model: string): ModelPricing {
       };
     }
   }
-  const directKey = model.toLowerCase();
-  if (KNOWN_MODEL_PRICING[directKey]) {
-    return KNOWN_MODEL_PRICING[directKey];
-  }
+  const genericHit = KNOWN_MODEL_PRICING[normalized];
+  if (genericHit) return genericHit;
+  if (isFreeModel(provider, { id: normalized }))
+    return { inputCostPer1M: 0, outputCostPer1M: 0, isFree: true };
   return { inputCostPer1M: 5.0, outputCostPer1M: 15.0, isFree: false };
 }
 
+/** Input cost per 1M tokens a virtual auto-combo candidate is scored at. */
+export function resolveVirtualCost(providerId: string, modelId: string): number {
+  return getModelPricing(providerId, modelId).inputCostPer1M;
+}
+
 export function isExplicitlyFree(provider: string, config: TierConfig): boolean {
   return config.freeProviders.includes(provider.toLowerCase());
 }
diff --git a/open-sse/services/tierResolver.ts b/open-sse/services/tierResolver.ts
index b1702c5d9c..d83e43bef0 100644
--- a/open-sse/services/tierResolver.ts
+++ b/open-sse/services/tierResolver.ts
@@ -142,3 +142,40 @@ export function getTierStats(): Record<ProviderTier, number> {
   }
   return stats;
 }
+
+export let tierAsyncFallbackTotal = 0; // exported for testability
+
+export async function classifyTierAsync(provider: string, model: string): Promise<TierAssignment> {
+  try {
+    const { getPricingForModel } = await import("@/lib/db/settings");
+    const db = await getPricingForModel(provider, model);
+    const input = Number((db as { input?: unknown } | null)?.input);
+    const output = Number((db as { output?: unknown } | null)?.output);
+    if (Number.isFinite(input) && input >= 0) {
+      const out = Number.isFinite(output) && output >= 0 ? output : input;
+      // Same thresholds as the sync path below; a DB $0 lands FREE by threshold
+      // (the DB carries no isFree flag of its own).
+      let tier: ProviderTier;
+      if (input <= currentConfig.defaults.freeThreshold) tier = PROVIDER_TIER.FREE;
+      else if (input <= currentConfig.defaults.cheapThreshold) tier = PROVIDER_TIER.CHEAP;
+      else tier = PROVIDER_TIER.PREMIUM;
+      const sync = getModelPricing(provider, model); // keeps freeQuotaLimit when DB is mute
+      const assignment: TierAssignment = {
+        provider,
+        model,
+        tier,
+        reason: `DB cost-based: $${input}/M input`,
+        costPer1MInput: input,
+        costPer1MOutput: out,
+        hasFreeTier: input === 0,
+        freeQuotaLimit: sync.freeQuotaLimit,
+      };
+      tierCache.set(cacheKey(provider, model), assignment);
+      return assignment;
+    }
+  } catch {
+    // fall through to sync
+  }
+  tierAsyncFallbackTotal++;
+  return classifyTier(provider, model);
+}
diff --git a/package.json b/package.json
index bd3367796d..e8f3389354 100644
--- a/package.json
+++ b/package.json
@@ -265,6 +265,7 @@
     "coverage:report": "cross-env NODE_OPTIONS=--max-old-space-size=8192 c8 report --merge-async --output-dir coverage --exclude=tests/** --exclude=**/*.test.* --reporter=text --reporter=text-summary --reporter=html --reporter=json-summary --reporter=lcov",
     "coverage:summary": "node scripts/check/test-report-summary.mjs --input coverage/coverage-summary.json --output coverage/coverage-report.md",
     "check:pr-test-policy": "node scripts/check/check-pr-test-policy.mjs",
+    "check:pricing-freshness": "node scripts/check/check-pricing-freshness.mjs",
     "coverage:report:legacy": "c8 report --output-dir coverage --exclude=open-sse --reporter=text --reporter=text-summary",
     "test:all": "npm run test:unit && npm run test:vitest && npm run test:vitest:ui && npm run test:ecosystem && npm run test:e2e",
     "check": "npm run lint && npm run test",
diff --git a/scripts/check/check-pricing-freshness.mjs b/scripts/check/check-pricing-freshness.mjs
new file mode 100644
index 0000000000..075077d13d
--- /dev/null
+++ b/scripts/check/check-pricing-freshness.mjs
@@ -0,0 +1,26 @@
+#!/usr/bin/env node
+// STRICT gate: KNOWN_MODEL_PRICING must have been touched within PRICING_STALE_AFTER_DAYS,
+// or tiers silently drift from real prices. Exits 1 on stale, 0 on fresh.
+// Run: node scripts/check/check-pricing-freshness.mjs
+import { execFileSync } from "node:child_process";
+
+const PRICING_STALE_AFTER_DAYS = 90;
+const TARGET = "open-sse/services/providerCostData.ts";
+
+function lastTouchDays() {
+  const out = execFileSync("git", ["log", "--follow", "-1", "--format=%ct", "--", TARGET], {
+    encoding: "utf8",
+  }).trim();
+  const touched = Number(out);
+  if (!Number.isFinite(touched) || touched <= 0) return Infinity;
+  return (Date.now() / 1000 - touched) / 86400;
+}
+
+const days = lastTouchDays();
+if (days > PRICING_STALE_AFTER_DAYS) {
+  console.error(
+    `STALE: ${TARGET} untouched for ${Math.floor(days)}d (> ${PRICING_STALE_AFTER_DAYS}d) — refresh prices or bump the gate with justification`
+  );
+  process.exit(1);
+}
+console.log(`pricing fresh: ${TARGET} touched ${Math.floor(days)}d ago`);
diff --git a/scripts/quality/run-all-gates.mjs b/scripts/quality/run-all-gates.mjs
index 54102b6abc..aa6b68ea90 100644
--- a/scripts/quality/run-all-gates.mjs
+++ b/scripts/quality/run-all-gates.mjs
@@ -29,6 +29,7 @@ const GATES = [
   // Group B — fast (<5s)
   { name: "check:provider-consistency", cmd: ["node", "--import", "tsx", "scripts/check/check-provider-consistency.ts"] },
   { name: "check:provider-assets", cmd: ["node", "scripts/check/check-provider-assets.mjs"] },
+  { name: "check:pricing-freshness", cmd: ["node", "scripts/check/check-pricing-freshness.mjs"] },
   { name: "check:provider-order-sync", cmd: ["node", "scripts/check/check-provider-order-sync.mjs"] },
   { name: "check:public-creds", cmd: ["node", "scripts/check/check-public-creds.mjs"] },
   { name: "check:error-helper", cmd: ["node", "scripts/check/check-error-helper.mjs"] },
diff --git a/src/lib/db/settings/pricing.ts b/src/lib/db/settings/pricing.ts
index b0f7dd1b74..e2496550a2 100644
--- a/src/lib/db/settings/pricing.ts
+++ b/src/lib/db/settings/pricing.ts
@@ -13,6 +13,16 @@ type PricingByProvider = Record<string, PricingModels>;
 export type PricingSource = "default" | "litellm" | "modelsDev" | "user";
 export type PricingSourceMap = Record<string, Record<string, PricingSource>>;
 
+async function touchPricing(): Promise<void> {
+  invalidateDbCache("pricing");
+  try {
+    const { clearTierCache } = await import("@omniroute/open-sse/services/tierResolver");
+    clearTierCache();
+  } catch {
+    // fail-open: a missed tier invalidation must never break a price write
+  }
+}
+
 function readPricingNamespace(
   db: ReturnType<typeof getDbInstance>,
   namespace: string
@@ -195,7 +205,7 @@ export async function updatePricing(pricingData: PricingByProvider) {
   });
   tx();
   backupDbFile("pre-write");
-  invalidateDbCache("pricing"); // Bust the pricing read cache
+  await touchPricing();
   const updated: PricingByProvider = {};
   const allRows = db.prepare("SELECT key, value FROM key_value WHERE namespace = 'pricing'").all();
   for (const row of allRows) {
@@ -234,6 +244,7 @@ export async function resetPricing(provider: string, model?: string) {
   }
 
   backupDbFile("pre-write");
+  await touchPricing();
   const allRows = db.prepare("SELECT key, value FROM key_value WHERE namespace = 'pricing'").all();
   const result: Record<string, unknown> = {};
   for (const row of allRows) {
@@ -250,5 +261,6 @@ export async function resetAllPricing() {
   const db = getDbInstance();
   db.prepare("DELETE FROM key_value WHERE namespace = 'pricing'").run();
   backupDbFile("pre-write");
+  await touchPricing();
   return {};
 }
diff --git a/tests/integration/manifest-routing.test.ts b/tests/integration/manifest-routing.test.ts
index 458b9af4ad..878eda7efb 100644
--- a/tests/integration/manifest-routing.test.ts
+++ b/tests/integration/manifest-routing.test.ts
@@ -3,7 +3,7 @@ import assert from "node:assert/strict";
 import { generateRoutingHints } from "../../open-sse/services/manifestAdapter.ts";
 
 test("manifest routing generates hints without error", async () => {
-  const hints = generateRoutingHints([], {
+  const hints = await generateRoutingHints([], {
     messages: [{ content: "Test" }],
   });
   assert.equal(hints.specificityLevel, "trivial");
@@ -11,7 +11,7 @@ test("manifest routing generates hints without error", async () => {
 });
 
 test("manifest routing failure gracefully falls back - empty targets handled", async () => {
-  const hints = generateRoutingHints([], {
+  const hints = await generateRoutingHints([], {
     messages: [{ content: "Hello world" }],
   });
   assert.equal(hints.eligibleTargets.length, 0);
@@ -20,7 +20,7 @@ test("manifest routing failure gracefully falls back - empty targets handled", a
 });
 
 test("specificity score is non-negative and bounded", async () => {
-  const hints = generateRoutingHints([], {
+  const hints = await generateRoutingHints([], {
     messages: [{ content: "Hello" }],
   });
   assert.ok(hints.specificity.score >= 0);
@@ -28,7 +28,7 @@ test("specificity score is non-negative and bounded", async () => {
 });
 
 test("routing hints contain all required fields", async () => {
-  const hints = generateRoutingHints([], {
+  const hints = await generateRoutingHints([], {
     messages: [{ content: "Test message" }],
   });
   assert.ok("specificityLevel" in hints);
@@ -40,7 +40,7 @@ test("routing hints contain all required fields", async () => {
 });
 
 test("trivial query recommends free tier", async () => {
-  const hints = generateRoutingHints([], {
+  const hints = await generateRoutingHints([], {
     messages: [{ content: "Hello" }],
   });
   assert.equal(hints.recommendedMinTier, "free");
diff --git a/tests/unit/auto-combo-budget.test.ts b/tests/unit/auto-combo-budget.test.ts
new file mode 100644
index 0000000000..1e9313ba58
--- /dev/null
+++ b/tests/unit/auto-combo-budget.test.ts
@@ -0,0 +1,57 @@
+// tests/unit/auto-combo-budget.test.ts
+import test from "node:test";
+import assert from "node:assert/strict";
+import { selectProvider } from "../../open-sse/services/autoCombo/engine.ts";
+import { getSelfHealingManager } from "../../open-sse/services/autoCombo/selfHealing.ts";
+import { DEFAULT_WEIGHTS } from "../../open-sse/services/autoCombo/scoring.ts";
+
+const healer = getSelfHealingManager();
+test.beforeEach(() => {
+  healer.exclusions.clear();
+  healer.incidentMode = false;
+});
+
+function config(over: Record<string, unknown> = {}) {
+  return {
+    id: "budget-probe",
+    name: "Budget Probe",
+    type: "auto",
+    candidatePool: [],
+    weights: DEFAULT_WEIGHTS,
+    explorationRate: 0,
+    ...over,
+  } as never;
+}
+
+function cand(provider: string, model: string, cost: number) {
+  return {
+    provider,
+    model,
+    quotaRemaining: 95,
+    quotaTotal: 100,
+    circuitBreakerState: "CLOSED",
+    costPer1MTokens: cost,
+    p95LatencyMs: 120,
+    latencyStdDev: 8,
+  } as never;
+}
+
+test("default budget math is unchanged without the new field", () => {
+  const c = [cand("openai", "gpt-4o", 5)];
+  const r = selectProvider(config({ budgetCap: 1 }), c);
+  assert.equal(r.provider, "openai"); // (5/1M)*1000 = $0.005 < $1
+});
+
+test("32k estimated tokens enforce a 32x tighter budget", () => {
+  const c = [cand("openai", "gpt-4o", 5)];
+  // (5/1M)*32000 = $0.16 — exceeds a $0.01 cap only when tokens are honored.
+  // Message verified: "No candidate fits within the configured budget cap…" (engine.ts:59-62).
+  assert.throws(
+    () =>
+      selectProvider(
+        config({ budgetCap: 0.01, budgetFallback: "strict", estimatedInputTokens: 32000 }),
+        c
+      ),
+    /budget cap/i
+  );
+});
diff --git a/tests/unit/combo-bootstrap-latency.test.ts b/tests/unit/combo-bootstrap-latency.test.ts
new file mode 100644
index 0000000000..0499d7c36a
--- /dev/null
+++ b/tests/unit/combo-bootstrap-latency.test.ts
@@ -0,0 +1,53 @@
+// tests/unit/combo-bootstrap-latency.test.ts
+import { test } from "node:test";
+import assert from "node:assert/strict";
+
+// Median helper lives in combo.ts; import it directly (export it non-default).
+import {
+  poolMedianP95Ms,
+  bootstrapMs,
+  shouldWarnBootstrap,
+  resetBootstrapCounters,
+  bootstrapLatencyHits,
+  bootstrapLatencyTotal,
+} from "../../open-sse/services/combo.ts";
+
+test("median ignores corrupt entries", () => {
+  const m = poolMedianP95Ms({
+    "a/x": { p95LatencyMs: 1000 },
+    "b/y": { p95LatencyMs: NaN },
+    "c/z": { p95LatencyMs: -5 },
+    "d/w": { p95LatencyMs: 3000 },
+  } as never);
+  assert.equal(m, 1000);
+});
+
+test("empty stats yield undefined (caller falls back to 1500, no cold-start alarm)", () => {
+  assert.equal(poolMedianP95Ms({}), undefined);
+});
+
+test("counters start at zero", () => {
+  resetBootstrapCounters();
+  assert.equal(bootstrapLatencyHits, 0);
+  assert.equal(bootstrapLatencyTotal, 0);
+});
+
+test("table hit bumps Total but not Hits; miss bumps both", () => {
+  resetBootstrapCounters();
+  assert.equal(bootstrapMs("gpt-4o-mini", 9999), 2764); // table hit
+  assert.equal(bootstrapLatencyTotal, 1);
+  assert.equal(bootstrapLatencyHits, 0);
+  assert.equal(bootstrapMs("some-new-model-xyz", 1234), 1234); // miss -> median
+  assert.equal(bootstrapLatencyTotal, 2);
+  assert.equal(bootstrapLatencyHits, 1);
+  assert.equal(bootstrapMs("another-new-model", undefined), 1500); // miss, no median
+  resetBootstrapCounters();
+});
+
+test("warn gate is pure: cold start and low ratio stay silent, dominant warns once", () => {
+  assert.equal(shouldWarnBootstrap(10, 10, false, 9999, 0), false); // cold start
+  assert.equal(shouldWarnBootstrap(2, 10, true, 9999, 0), false); // ratio <= 0.3
+  assert.equal(shouldWarnBootstrap(5, 10, true, 9999, 0), false); // throttled 1h
+  assert.equal(shouldWarnBootstrap(5, 10, true, 3600_001, 0), true); // window elapsed
+  assert.equal(shouldWarnBootstrap(5, 10, true, 9999, 9000), false); // throttled 1h
+});
diff --git a/tests/unit/complexity-router.test.ts b/tests/unit/complexity-router.test.ts
index 7a6e6745c7..de7b565d75 100644
--- a/tests/unit/complexity-router.test.ts
+++ b/tests/unit/complexity-router.test.ts
@@ -78,8 +78,8 @@ test("classifyRequestComplexity — tool schemas escalate the tier above free",
   assert.notEqual(c.recommendedTier, "free", "tool-using requests must not route to the free tier");
 });
 
-test("buildComplexityRoutingHint — a tool-using request floors the hint tier above free", () => {
-  const hint = buildComplexityRoutingHint(
+test("buildComplexityRoutingHint — a tool-using request floors the hint tier above free", async () => {
+  const hint = await buildComplexityRoutingHint(
     modelTargets(),
     {
       messages: [{ role: "user", content: "weather?" }],
@@ -96,8 +96,8 @@ test("buildComplexityRoutingHint — a tool-using request floors the hint tier a
   );
 });
 
-test("buildComplexityRoutingHint — a null body is safe and still builds a tier-neutral hint", () => {
-  const hint = buildComplexityRoutingHint(modelTargets(), null, NOOP_LOG);
+test("buildComplexityRoutingHint — a null body is safe and still builds a tier-neutral hint", async () => {
+  const hint = await buildComplexityRoutingHint(modelTargets(), null, NOOP_LOG);
   assert.ok(hint, "a null body must not throw — messages default to [] and a hint is built");
   if (!hint) return;
   assert.ok(
diff --git a/tests/unit/pricing-fallback.test.ts b/tests/unit/pricing-fallback.test.ts
new file mode 100644
index 0000000000..13e2f4fffc
--- /dev/null
+++ b/tests/unit/pricing-fallback.test.ts
@@ -0,0 +1,33 @@
+// tests/unit/pricing-fallback.test.ts
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { getModelPricing, KNOWN_MODEL_PRICING } from "../../open-sse/services/providerCostData.ts";
+
+// `ai21/jamba-large-1.7` is a documented free model (FREE_MODEL_BUDGETS,
+// one-time-initial) that is NOT in the table (grep jamba -> 0 hit).
+// `openai/gpt-4o` is paid. Fixtures are verbatim catalog ids — never a suffixed
+// invention (Set.has is exact: a mutilated id is no longer in the catalog).
+test("off-table free model is free, never premium-priced", () => {
+  const p = getModelPricing("ai21", "jamba-large-1.7");
+  assert.equal(p.isFree, true);
+  assert.equal(p.inputCostPer1M, 0);
+  assert.equal(p.outputCostPer1M, 0);
+});
+
+test("prefixed model id still resolves through the catalog", () => {
+  const p = getModelPricing("ai21", "ai21/jamba-large-1.7");
+  assert.equal(p.isFree, true);
+});
+
+test("off-table paid model keeps the conservative fallback", () => {
+  const p = getModelPricing("openai", "gpt-9-never-existed");
+  assert.equal(p.isFree, false);
+  assert.equal(p.inputCostPer1M, 5.0);
+  assert.equal(p.outputCostPer1M, 15.0);
+});
+
+test("table hits are untouched", () => {
+  assert.equal(Object.keys(KNOWN_MODEL_PRICING).length, 21);
+  assert.deepEqual(getModelPricing("openai", "gpt-4o"), KNOWN_MODEL_PRICING["gpt-4o"]);
+  assert.deepEqual(getModelPricing("longcat", "LongCat-2.0"), KNOWN_MODEL_PRICING["longcat-2.0"]);
+});
diff --git a/tests/unit/tier-pricing-cache.test.ts b/tests/unit/tier-pricing-cache.test.ts
new file mode 100644
index 0000000000..844cfaafa6
--- /dev/null
+++ b/tests/unit/tier-pricing-cache.test.ts
@@ -0,0 +1,50 @@
+// tests/unit/tier-pricing-cache.test.ts
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import {
+  classifyTier,
+  classifyTierAsync,
+  clearTierCache,
+} from "../../open-sse/services/tierResolver.ts";
+import { updatePricing, resetPricing, resetAllPricing } from "../../src/lib/db/settings/pricing.ts";
+
+test.beforeEach(() => clearTierCache());
+
+test("async tier reflects a DB price write without restart", async () => {
+  try {
+    await updatePricing({ testprov: { "testmodel-cache-1": { input: 0, output: 0 } } });
+    const a = await classifyTierAsync("testprov", "testmodel-cache-1");
+    assert.equal(a.tier, "free");
+  } finally {
+    await resetPricing("testprov", "testmodel-cache-1");
+  }
+});
+
+test("resetAllPricing (full wipe, no read-cache bust) also falls back cleanly", async () => {
+  try {
+    await updatePricing({ testprov: { "testmodel-cache-2": { input: 0, output: 0 } } });
+    assert.equal((await classifyTierAsync("testprov", "testmodel-cache-2")).tier, "free");
+    await resetAllPricing();
+    const back = await classifyTierAsync("testprov", "testmodel-cache-2");
+    assert.equal(back.tier, classifyTier("testprov", "testmodel-cache-2").tier);
+  } finally {
+    await resetPricing("testprov");
+  }
+});
+
+test("async tier falls back to sync when DB is down", async () => {
+  const a = await classifyTierAsync("openai", "gpt-4o");
+  assert.deepEqual(a, classifyTier("openai", "gpt-4o"));
+});
+
+test("sync tier keeps serving the table while DB changes", async () => {
+  // Fictitious model on a real provider: never pollutes shared state if cleanup fails.
+  const before = classifyTier("openai", "gpt-9-never-existed");
+  try {
+    await updatePricing({ openai: { "gpt-9-never-existed": { input: 0, output: 0 } } });
+    const during = classifyTier("openai", "gpt-9-never-existed");
+    assert.equal(during.tier, before.tier); // sync path untouched by DB
+  } finally {
+    await resetPricing("openai", "gpt-9-never-existed");
+  }
+});
diff --git a/tests/unit/virtualFactory-valuation.test.ts b/tests/unit/virtualFactory-valuation.test.ts
new file mode 100644
index 0000000000..79f9665c1f
--- /dev/null
+++ b/tests/unit/virtualFactory-valuation.test.ts
@@ -0,0 +1,47 @@
+// tests/unit/virtualFactory-valuation.test.ts (part 1 — reliability)
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { computeSnapshotWeights } from "../../open-sse/services/autoCombo/virtualFactory.ts";
+import { resolveVirtualCost } from "../../open-sse/services/providerCostData.ts";
+import { DEFAULT_WEIGHTS } from "../../open-sse/services/autoCombo/scoring.ts";
+
+function candidate(over: Record<string, unknown> = {}) {
+  return {
+    provider: "openai",
+    connectionId: null,
+    model: "gpt-4o",
+    modelStr: "openai/gpt-4o",
+    costPer1MTokens: 5,
+    ...over,
+  } as never;
+}
+
+test("absent failure rate reads fully reliable (neutral 1)", () => {
+  const weights = { ...DEFAULT_WEIGHTS, reliability: 0.2 };
+  const scores = computeSnapshotWeights([candidate()], weights);
+  const base = computeSnapshotWeights([candidate()], { ...DEFAULT_WEIGHTS, reliability: 0 });
+  const delta0 = scores.get("openai/gpt-4o")! - base.get("openai/gpt-4o")!;
+  assert.ok(Math.abs(delta0 - 0.2) < 1e-9, `expected 0.2, got ${delta0}`);
+});
+
+test("failureRate 0.2 scales the reliability weight to 0.8", () => {
+  const weights = { ...DEFAULT_WEIGHTS, reliability: 0.2 };
+  const scores = computeSnapshotWeights([candidate({ failureRate: 0.2 })], weights);
+  const base = computeSnapshotWeights([candidate()], { ...DEFAULT_WEIGHTS, reliability: 0 });
+  const delta = scores.get("openai/gpt-4o")! - base.get("openai/gpt-4o")!;
+  assert.ok(Math.abs(delta - 0.16) < 1e-9, `expected 0.16, got ${delta}`);
+});
+
+test("errorRate backs up a missing failureRate", () => {
+  const weights = { ...DEFAULT_WEIGHTS, reliability: 0.2 };
+  const scores = computeSnapshotWeights([candidate({ errorRate: 0.5 })], weights);
+  const base = computeSnapshotWeights([candidate()], { ...DEFAULT_WEIGHTS, reliability: 0 });
+  const delta = scores.get("openai/gpt-4o")! - base.get("openai/gpt-4o")!;
+  assert.ok(Math.abs(delta - 0.1) < 1e-9, `expected 0.1, got ${delta}`);
+});
+
+test("virtual cost helper returns real table costs, not zero", () => {
+  assert.equal(resolveVirtualCost("openai", "gpt-4o"), 2.5);
+  assert.equal(resolveVirtualCost("ai21", "jamba-large-1.7"), 0); // catalog free
+  assert.equal(resolveVirtualCost("openai", "gpt-9-never-existed"), 5.0); // conservative fallback
+});