fix: use realtime text and Azure HD for Azure Realtime Fast

tools/cost_calculator/README.md

@@ -84,10 +84,13 @@ The GPT Live meter is hourly and is divided by 60 for USD-per-minute pricing.
   meter; `gpt-realtime` uses the original output meter.
 - **Azure Realtime** uses the documented Voice Live **Pro** tier.
   **Azure Realtime Fast** uses a **user-specified Lite-tier reference**,
-  not a verified published Fast tariff. Both use the existing Central India
-  tier text and native-audio meters, with no separate STT, Azure HD, or GPT Live
-  session charge. They appear in managed native speech-to-speech selection,
-  comparison, and CSV export only; no BYOM or alternate-pipeline prices are
+  not a verified published Fast tariff. Azure Realtime uses managed native
+  speech-to-speech with no separate STT or Azure HD charge. Azure Realtime Fast
+  uses **Realtime text response + Azure HD**: Lite-tier native audio input,
+  text input/output, and standard-speech output meters, without native audio
+  output charges. Both use the existing Central India tier meters and appear
+  in their respective managed pipeline selection, comparison, and CSV export.
+  Neither adds a GPT Live session charge; no BYOM or alternate-pipeline prices are
   inferred from GPT models. Fast's assumption is labeled in the UI and CSV.
   The editable audio tokens/second remain planning assumptions, not verified
   Azure Realtime-family tokenization. No price-snapshot refresh is required

tools/cost_calculator/app.mjs

@@ -111,7 +111,12 @@ function rateEditor(model, pipeline, book) {
   }
   if (pipeline === "live") $("price-note").textContent += " GPT Live uses Azure OpenAI's South India Global voice rate: $3/hour = $0.05/minute in this snapshot. Backend model usage is charged separately. GPT Live generates its own audio, so no Azure HD charge is added.";
   if (model.oss && (hosting === "byom" || book.usesGlobalReference)) $("price-note").textContent += " No OSS cached-input meter was found; reused text is conservatively charged at full input price.";
-  if (model.managedNativeOnly) $("price-note").textContent += " Azure Realtime models use native audio only in this calculator; no separate Azure STT, HD voice, or GPT Live session charge is added. Audio token rates remain editable assumptions, not verified Azure Realtime tokenization.";
+  if (model.managedPipeline) {
+    $("price-note").textContent += model.managedPipeline === "hybrid" ?
+      " Azure Realtime Fast uses native audio input, text response, and Azure HD speech output. No native audio output, separate Azure STT, or GPT Live session charge is added." :
+      " Azure Realtime uses native audio only in this calculator; no separate Azure STT, HD voice, or GPT Live session charge is added.";
+    $("price-note").textContent += " Audio token rates remain editable assumptions, not verified Azure Realtime tokenization.";
+  }
   if (model.tierAssumption) $("price-note").textContent += " Azure Realtime Fast uses the user-specified Voice Live Lite tier reference, not a verified published Fast tariff.";
 }
 
@@ -134,8 +139,11 @@ function renderOverview(a, model, pipeline, result, book) {
     book.usesGlobalReference ?
     "Managed tariff unavailable: showing an AOAI Global model-price + Central India BYO speech reference instead. This does not establish managed model support." :
     `Economic scenario only. ${$("hosting").value === "managed" ? `Central India managed-tier reference: ${model.tier || "unverified"}.` : "Model rates: South India Global. STT: Central India BYO reference."} Azure HD: Central India. Confirm model and transport availability.`;
-  if (model.managedNativeOnly) {
-    $("model-note").textContent = `${model.tierAssumption ? "User-specified Lite-tier reference, not a verified published Azure Realtime Fast tariff." : "Central India Voice Live Pro tier reference."} Native audio only; no separate STT/HD charge. Confirm regional availability and actual audio token rates.`;
+  if (model.managedPipeline) {
+    const audioNote = model.managedPipeline === "hybrid" ?
+      "Native audio input + text response + Central India Azure HD output; no native audio output charge." :
+      "Native audio only; no separate STT/HD charge.";
+    $("model-note").textContent = `${model.tierAssumption ? "User-specified Lite-tier reference, not a verified published Azure Realtime Fast tariff." : "Central India Voice Live Pro tier reference."} ${audioNote} Confirm regional availability and actual audio token rates.`;
   }
 }
 

tools/cost_calculator/calculator.mjs

@@ -14,9 +14,9 @@ export const MODELS = [
   { id: "gpt-realtime-2.1-mini", name: "GPT Realtime 2.1 mini", tier: "Std", realtime: true },
   { id: "gpt-realtime", name: "GPT Realtime 1.0 (gpt-realtime)", tier: "Pro", realtime: true },
   { id: "gpt-realtime-1.5", name: "GPT Realtime 1.5", tier: "Pro", realtime: true },
-  { id: "azure-realtime", name: "Azure Realtime", tier: "Pro", realtime: true, managedNativeOnly: true },
+  { id: "azure-realtime", name: "Azure Realtime", tier: "Pro", realtime: true, managedPipeline: "native" },
   { id: "azure-realtime-fast", name: "Azure Realtime Fast", tier: "Lite", realtime: true,
-    managedNativeOnly: true, tierAssumption: true },
+    managedPipeline: "hybrid", tierAssumption: true },
   { id: "gpt-4.1", name: "GPT-4.1", tier: "Pro" },
   { id: "gpt-4.1-mini", name: "GPT-4.1 mini", tier: "Std" },
   { id: "gpt-4.1-nano", name: "GPT-4.1 nano", tier: "Lite" },
@@ -31,7 +31,7 @@ export const MODELS = [
 ];
 
 export function supportsScenario(model, hosting, pipeline) {
-  if (model.managedNativeOnly) return hosting === "managed" && pipeline === "native";
+  if (model.managedPipeline) return hosting === "managed" && pipeline === model.managedPipeline;
   return ["native", "hybrid"].includes(pipeline) ? Boolean(model.realtime) : !model.realtime;
 }
 
@@ -63,9 +63,9 @@ export function meterRate(snapshot, name, unit = "tokens") {
   return meter.retailPrice * factor;
 }
 
-export function getRates(snapshot, model, hosting, pipeline = "cascade") {
-  if (model.managedNativeOnly && !supportsScenario(model, hosting, pipeline)) {
-    throw new Error(`${model.name} uses a managed native-audio tier reference; no BYOM or alternative-pipeline rates are verified.`);
+export function getRates(snapshot, model, hosting, pipeline = model.managedPipeline || "cascade") {
+  if (model.managedPipeline && !supportsScenario(model, hosting, pipeline)) {
+    throw new Error(`${model.name} uses a managed ${model.managedPipeline} tier reference; no BYOM or alternative-pipeline rates are verified.`);
   }
   const rates = { textInput: null, textCached: null, textOutput: null,
     audioInput: null, audioCached: null, audioOutput: null, speechInput: null, speechOutput: null,

tools/cost_calculator/calculator.test.mjs

@@ -5,7 +5,7 @@ import { DEFAULTS, MODELS, PROMPTS, GPT_LIVE_METER, supportsScenario, getRates,
 
 const snapshot = JSON.parse((await readFile(new URL("./prices.json", import.meta.url), "utf8")).replace(/^\uFEFF/, ""));
 const model = id => MODELS.find(m => m.id === id);
-const book = (id, hosting = "managed", pipeline = model(id).realtime ? "native" : "cascade") =>
+const book = (id, hosting = "managed", pipeline = model(id).managedPipeline || (model(id).realtime ? "native" : "cascade")) =>
   getRates(snapshot, model(id), hosting, pipeline).rates;
 const close = (actual, expected) => assert.ok(Math.abs(actual - expected) < 1e-12, `${actual} != ${expected}`);
 
@@ -30,38 +30,55 @@ test("every supported model/hosting/pipeline combination resolves without a miss
   }
 });
 
-test("Azure Realtime uses Pro and Fast uses the explicit Lite assumption with native meters", () => {
+test("Azure Realtime uses Pro native audio and Fast uses Lite realtime text plus HD", () => {
   for (const [id, tier, expected] of [
     ["azure-realtime", "Pro", [4, 0.4, 16, 32, 0.4, 64]],
     ["azure-realtime-fast", "Lite", [0.11, 0.04, 0.44, 4, 0.04, 12]],
   ]) {
     const m = model(id);
     assert.equal(m.tier, tier);
     assert.equal(Boolean(m.tierAssumption), id === "azure-realtime-fast");
-    const { rates, sources, usesGlobalReference } = getRates(snapshot, m, "managed", "native");
+    const pipeline = m.managedPipeline;
+    const { rates, sources, usesGlobalReference } = getRates(snapshot, m, "managed", pipeline);
     assert.equal(usesGlobalReference, false);
     const keys = ["textInput", "textCached", "textOutput", "audioInput", "audioCached", "audioOutput"];
     keys.forEach((key, index) => {
       close(rates[key], expected[index]);
       assert.equal(sources[key].armRegionName, "centralindia");
       assert.ok(sources[key].meterName.startsWith(`Voice Live API ${tier} - LLM `));
     });
-    const result = calculate({ ...DEFAULTS, nativeText: 150 }, "native", rates);
+    const result = calculate({ ...DEFAULTS, nativeText: 150 }, pipeline, rates);
     const [textInput, textCached, textOutput, audioInput, audioCached, audioOutput] = expected;
-    close(result.list, (5000 * textInput + 15225 * textCached + 150 * textOutput +
-      360 * audioInput + 1620 * audioCached + 720 * audioOutput) / 1e6);
+    if (pipeline === "hybrid") {
+      close(rates.speechOutput, 25);
+      assert.equal(sources.speechOutput.meterName, "Voice Live API Lite - Standard Speech Audio Output Tokens");
+      assert.equal(sources.speechOutput.armRegionName, "centralindia");
+      close(result.list, (5000 * textInput + 15225 * textCached + 150 * textOutput +
+        360 * audioInput + 540 * audioCached + 720 * 25) / 1e6);
+      assert.ok(result.rows.some(r => r.key === "speechOutput"));
+      assert.ok(result.rows.every(r => !["audioOutput", "speechInput", "liveMinute"].includes(r.key)));
+      const noCache = calculate({ ...DEFAULTS, cachePercent: 0 }, pipeline, rates);
+      close(noCache.list, (20225 * textInput + 900 * audioInput + 150 * textOutput + 720 * 25) / 1e6);
+      const reasoning = calculate({ ...DEFAULTS, reasoning: 100 }, pipeline, rates);
+      close(reasoning.list - result.list, 100 * textOutput / 1e6);
+      const silent = calculate({ ...DEFAULTS, silencePercent: 100 }, pipeline, rates);
+      close(silent.list, (5000 * textInput + 15225 * textCached + 150 * textOutput) / 1e6);
+    } else {
+      close(result.list, (5000 * textInput + 15225 * textCached + 150 * textOutput +
+        360 * audioInput + 1620 * audioCached + 720 * audioOutput) / 1e6);
+      assert.ok(result.rows.every(r => !["speechInput", "speechOutput", "liveMinute"].includes(r.key)));
+      const noCache = calculate({ ...DEFAULTS, cachePercent: 0 }, pipeline, rates);
+      close(noCache.list, (20000 * textInput + 1980 * audioInput + 720 * audioOutput) / 1e6);
+    }
     close(result.listPerMinute, result.list / 1.5);
     assert.deepEqual(result.missing, []);
-    assert.ok(result.rows.every(r => !["speechInput", "speechOutput", "liveMinute"].includes(r.key)));
-    const noCache = calculate({ ...DEFAULTS, cachePercent: 0 }, "native", rates);
-    close(noCache.list, (20000 * textInput + 1980 * audioInput + 720 * audioOutput) / 1e6);
   }
 });
 
-test("Azure models are selectable only for managed native audio, never substituted with GPT BYOM rates", () => {
+test("Azure models use their correct managed pipeline, never substituted with GPT BYOM rates", () => {
   for (const id of ["azure-realtime", "azure-realtime-fast"]) {
     for (const hosting of ["managed", "byom"]) for (const pipeline of ["native", "hybrid", "cascade", "live"]) {
-      const supported = hosting === "managed" && pipeline === "native";
+      const supported = hosting === "managed" && pipeline === (id === "azure-realtime-fast" ? "hybrid" : "native");
       assert.equal(supportsScenario(model(id), hosting, pipeline), supported);
       if (!supported) assert.throws(() => getRates(snapshot, model(id), hosting, pipeline), /no BYOM or alternative-pipeline rates/);
     }
@@ -138,7 +155,7 @@ test("GPT-5.6 Luna BYOM uses exact Standard Global inference meters and BYO spee
 
 test("all selectable models have priced scenarios across both hosting modes", () => {
   for (const hosting of ["managed", "byom"]) for (const m of MODELS) {
-    const pipeline = m.realtime ? "native" : "cascade";
+    const pipeline = m.managedPipeline || (m.realtime ? "native" : "cascade");
     if (!supportsScenario(m, hosting, pipeline)) continue;
     const result = calculate(DEFAULTS, pipeline, book(m.id, hosting));
     assert.notEqual(result.list, null, `${hosting} / ${m.id}`);
@@ -325,8 +342,9 @@ test("silence scales only audio rows, with zero preserving baseline and full sil
 test("all five prompt sizes grow monotonically for every priced model", () => {
   assert.deepEqual(PROMPTS, [0, 1000, 5000, 10000, 20000]);
   for (const m of MODELS) for (const hosting of ["managed", "byom"]) {
-    if (!supportsScenario(m, hosting, m.realtime ? "native" : "cascade")) continue;
-    const costs = PROMPTS.map(prompt => calculate({ ...DEFAULTS, prompt }, m.realtime ? "native" : "cascade", book(m.id, hosting)).list);
+    const pipeline = m.managedPipeline || (m.realtime ? "native" : "cascade");
+    if (!supportsScenario(m, hosting, pipeline)) continue;
+    const costs = PROMPTS.map(prompt => calculate({ ...DEFAULTS, prompt }, pipeline, book(m.id, hosting)).list);
     if (costs.every(c => c !== null)) {
       for (let i = 1; i < costs.length; i++) assert.ok(costs[i] >= costs[i - 1]);
     }