Add Voice Agent cost calculator and GitHub Pages deployment

.github/workflows/cost-calculator-pages.yml

@@ -0,0 +1,68 @@
+name: Cost calculator Pages
+
+on:
+  pull_request:
+    paths:
+      - "tools/cost_calculator/**"
+      - ".github/workflows/cost-calculator-pages.yml"
+  push:
+    branches: [main]
+    paths:
+      - "tools/cost_calculator/**"
+      - ".github/workflows/cost-calculator-pages.yml"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  validate:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "24"
+      - name: Check calculator
+        working-directory: tools/cost_calculator
+        shell: pwsh
+        run: |
+          node --test calculator.test.mjs
+          if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
+          node --check app.mjs
+          if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
+          node --check server.mjs
+          if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
+
+  deploy:
+    if: github.ref == 'refs/heads/main' && github.event_name != 'pull_request'
+    needs: validate
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      pages: write
+      id-token: write
+    concurrency:
+      group: github-pages
+      cancel-in-progress: false
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/configure-pages@v5
+      - name: Stage browser assets only
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = "Stop"
+          $site = Join-Path $env:RUNNER_TEMP "cost-calculator-site"
+          New-Item -ItemType Directory -Path $site | Out-Null
+          foreach ($file in @("index.html", "styles.css", "app.mjs", "calculator.mjs", "prices.json")) {
+            Copy-Item -LiteralPath (Join-Path (Join-Path "tools" "cost_calculator") $file) -Destination $site
+          }
+      - uses: actions/upload-pages-artifact@v3
+        with:
+          path: ${{ runner.temp }}/cost-calculator-site
+      - name: Deploy GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@v4

samples/README.md

@@ -18,6 +18,7 @@ independent shared MCP reference implementation, start with the
 
 | Sample | What it demonstrates |
 | --- | --- |
+| [Voice Agent Cost Lab](../tools/cost_calculator/README.md) | Standalone browser list-price calculator in `tools`: South India model prices, labeled Central India HD/Voice Live references, managed/BYOM model selection, prompt sensitivity, and CSV export. No Azure credentials required. |
 | [Add voice to an existing customer agent](existing_agent_voice_design/README.md) | Design guide: use a Hosted Agent as the Foundry Voice Agent conversation engine to connect managed voice to an existing customer agent or API endpoint. Includes a high-level architecture diagram; not a runnable sample. |
 | [Voice Live to Voice Agent migration](voice_live_to_voice_agent/README.md) | Compare the same Azure Speech -> GPT-4.1 mini -> Azure TTS cascade and shared local functions before and after migration. Includes dynamic session context and mid-session instruction-update examples. |
 | [Vehicle voice agent with local tools](vehicle_local_tools/README.md) | Web cockpit with cloud voice interaction and locally executed simulated climate, media, and navigation functions. Includes an offline dashboard, local policy enforcement, and a tool execution log. |

tools/cost_calculator/README.md

@@ -0,0 +1,166 @@
+# Voice Agent Cost Lab
+
+Standalone browser calculator with **South India / USD / pay-as-you-go**
+model-price references.
+Compare managed Voice Agent / Voice Live tiers and **bring your own model
+(BYOM)** using list-price estimates.
+BYOM models and GPT Live use **South India Global** prices from Azure OpenAI.
+**Azure HD voice uses Central India rates** as requested. No South India
+Voice Live tier or BYO speech meters were found, so managed tiers and BYO STT
+also use the approved, explicitly labeled **Central India references**.
+No Azure credentials, packages, build step, or live model calls are needed.
+
+## Run
+
+With Node.js 20 or later, from the repository root:
+
+```powershell
+node tools\cost_calculator\server.mjs
+```
+
+Open **http://127.0.0.1:8095**. To change the port, set `$env:PORT = "8096"`
+before running the command. The server listens only on the loopback interface.
+Stop it with Ctrl+C. Opening `index.html` as `file://` will not work because
+the page loads modules and the local price snapshot.
+
+Choose hosting, pipeline, and model; change call assumptions or deployment
+rates. The tool shows per-call, per-minute, call-volume list prices, component costs, and
+comparisons at **0, 1K, 5K, 10K, and 20K** system-prompt tokens. Export CSV
+includes list-price amounts, assumptions, overrides, snapshot date, and
+component reference regions, processing scopes, and meter IDs.
+The CSV also distinguishes managed-tier pricing from AOAI Global economic
+references when a model has no published managed tier.
+Changes are in memory only and reset on reload.
+
+## Deploy
+
+This is a static page: calculations run in the browser and `prices.json` is
+bundled with the page. The Node server is for local preview only.
+
+The [Pages workflow](../../.github/workflows/cost-calculator-pages.yml)
+checks calculator changes in pull requests and publishes the site when those
+changes merge into `main`. It can also be run manually from `main` in the
+Actions tab. Pull requests and other branches cannot deploy.
+
+The repository must have **Settings > Pages > Source > GitHub Actions**
+enabled. The deployment publishes exactly `index.html`, `styles.css`,
+`app.mjs`, `calculator.mjs`, and `prices.json` at the site's root, excluding
+the local server, tests, documentation, and all other repository files.
+No dependency installation or build command is needed. All asset URLs are
+relative, so the page works under a repository subpath.
+
+For this repository, the published address is
+**https://microsoft.github.io/voice-agent/** after the first successful
+deployment. The site is public; do not add confidential data or credentials.
+
+**Azure Static Web Apps** is an alternative, especially when access control
+is required. Set the app location to this tool's directory, leave the API
+location empty, and deploy the existing assets without a build. Configure
+authentication and route access rules if the page must not be public.
+
+Deploy only the five browser assets above; no server, credentials, or live
+Azure connection is needed. Pricing updates require updating the snapshot
+and redeploying.
+
+## Price basis
+
+`prices.json` is a **2026-10-22 snapshot**, not a live price feed, from the
+[Azure Retail Prices API](https://prices.azure.com/api/retail/prices).
+It records original billing units, exact USD prices, meter IDs, filters, and
+effective dates where returned. Rates per 1K are multiplied by 1,000 for the
+calculator's USD-per-million fields.
+The GPT Live meter is hourly and is divided by 60 for USD-per-minute pricing.
+
+- Managed models use Central India's Pro, Standard, or Lite **Voice Live
+  meter references**, not
+  direct OpenAI model API prices. GPT Realtime 2.1 uses the Pro text-output-2
+  meter; `gpt-realtime` uses the original output meter.
+- When no managed tier is published, use the exact model's **AOAI Global
+  inference prices + Central India BYO speech** as an economic reference.
+  This fills GPT-5.4 mini/nano, GPT OSS 120B, and GPT-6 Sol reference scenarios.
+  The UI and CSV explicitly label these **not confirmed managed Voice Live
+  tariffs**. A model price does not prove managed Voice Agent availability.
+- Azure HD uses the **Central India** Voice Live standard-speech output meter. This is not
+  standalone Speech's character-billed HD price.
+- BYOM uses **Central India Voice Live BYO standard-speech input/output**
+  references plus the selected model's **South India Global** deployment
+  prices. GPT-4.1, GPT-5.1, GPT-5.2, GPT-5.4, GPT Realtime 1.0/1.5/2.1,
+  GPT Realtime 2.1 mini, GPT OSS 120B, and GPT-6 Sol are prefilled from the
+  published South India meters. Global describes processing scope, not a
+  promise of India-only inference.
+- **GPT-6 reasoning is explicitly GPT-6 Sol**, using South India Global
+  short-context Standard prices: **$2 input / $0.20 cached input / $10 output
+  per 1M tokens**. The generic, unspecified GPT-6 placeholder is replaced
+  with the exact priced variant. Reasoning tokens use its output rate.
+- No cached OSS price was found, so replayed OSS text uses its full input
+  rate. You can explicitly override that conservative assumption.
+- **GPT Live (`gpt-live-1`) voice sessions:** Azure OpenAI Media's `Live 1 Gl Unit`
+  meter lists **$3/hour = $0.05/minute** for South India Global, effective
+  2026-09-01. Meter ID: `2e5b0a5c-3723-50a8-ba75-a5473f702fc1`. This is the
+  South India Global reference, **not a Central India rate or an
+  availability guarantee**. A 90-second session costs **$0.075 for voice**,
+  before separately billed backend model usage.
+- Other missing deployment prices require **manual rates**. No price is inferred
+  from the supplied reference comparison table or substituted from another
+  model or provider. Cross-region Voice Live references are shown explicitly.
+  GPT Live uses a separately priced backend deployment regardless
+  of the hosting selector, plus the Azure OpenAI session-minute reference.
+- **Blank means unknown**, not zero. Missing model or speech prices block
+  complete totals until the required list rates are entered.
+
+Selection is an economic scenario, **not proof of model, transport, or Voice
+Agent availability in South India**. In particular, a BYOM native-audio
+scenario requires confirmation of API support and deployment prices.
+Provisioned-throughput and dedicated hosting charges are not modeled.
+
+References: [Voice Live tier mapping and usage](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live),
+[Speech list pricing](https://azure.microsoft.com/en-us/pricing/details/speech/),
+[Azure OpenAI pricing](https://azure.microsoft.com/en-us/pricing/details/azure-openai/),
+[Voice Agent billing](https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/voice-agent-pricing).
+
+To refresh the snapshot, rerun its two recorded OData filters with
+`currencyCode='USD'`, follow every `NextPageLink`, and review the exact meter
+names and units before replacing the file. Do not silently mix regional,
+Global, Data Zone, Batch, or fine-tuning meters.
+
+## Assumptions and formulas
+
+Defaults preserve the supplied **90-second call**, **4 user + 4 agent
+turns**, **45 seconds of speech each**, **20 audio tokens/sec** in both
+directions, **100% reused-context caching**, and **150 agent text tokens**
+for text-response pipelines. User transcript length was not supplied:
+the calculator explicitly assumes an editable **150 user text tokens**.
+Native output text and reasoning tokens default to zero and can be changed.
+These defaults are modeling inputs, not measured usage.
+
+Each call starts cold; the system prompt costs `P * input_rate` on its
+first turn. Its `P * (turns - 1)` reused tokens are charged at the selected
+cache hit rate. With equal turn lengths, each side's total conversation
+tokens are replayed `(turns - 1) / 2` times across later turns. Cache misses
+use the full input rate. Native mode reuses caller and assistant audio.
+Realtime-to-HD mode reuses caller audio and assistant text. Cascade and
+GPT Live's reasoning backend reuse user/assistant text. Hidden reasoning is
+billed as output but not replayed as visible transcript history.
+
+For each component, `list cost = quantity * USD_per_million / 1,000,000`.
+GPT Live voice sessions instead use `call_minutes * USD_per_minute`, in
+addition to the turn-based backend model estimate. Real delegation patterns
+can differ from this simplified assumption.
+
+Per-minute list prices divide by the full call duration, not just speaking
+time. Call-volume estimates multiply the per-call list price by the number
+of calls.
+
+The original reference totals are not regression targets: they do not supply
+all token prices or context-accounting details and may use different rates.
+This calculator recomputes costs transparently rather than fitting those totals.
+Actual caching eligibility, audio token rates, silence billing, tool calls,
+telephony, storage, network, tax, capacity charges, and model quality are not
+captured. In particular, the documented OpenAI input estimate is typically
+~10 audio tokens/sec; 20 is retained here because it was explicitly requested.
+
+## Checks
+
+```powershell
+node --test tools\cost_calculator\calculator.test.mjs
+```