Shaw/live ai pricing

[lalalune]

No description provided.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
eliza-cloud-v2	Error		Apr 15, 2026 1:51am

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: cdc64e6b-1738-4f88-8963-35dd963fac15

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch shaw/live-ai-pricing

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[claude]

Code Review: Shaw/live ai pricing (#455)

Overview

This PR introduces a live AI pricing catalog — replacing hardcoded constants with a DB-backed catalog that syncs from external sources (Vercel Gateway, OpenRouter, FAL.ai, ElevenLabs). It adds an admin override endpoint, a public pricing summary endpoint, a cron refresh job, and a new music-metadata-based audio duration estimator for more accurate STT billing. A significant restructuring: voice endpoints are unified under /api/v1/voice/ and the old /api/elevenlabs/ endpoints are now thin re-exports.

---

🔴 Critical Issues

1. Failed video generation now refunds 100% of credits (generate-video/route.ts)

Previously a failed video attempt was charged VIDEO_GENERATION_FALLBACK_COST (partial charge for the API attempt). This PR changes it to reconcile(0) — a full refund. Users can now trigger video generation requests indefinitely at zero cost if the underlying provider fails. This should be intentional if so — but the prior logic existed specifically to cover provider-side costs.

- await reservation.reconcile(VIDEO_GENERATION_FALLBACK_COST);
+ await reservation.reconcile(0);

2. seedance pricing parser is not implemented but the model is in the type union (ai-pricing.ts)

case "seedance":
  throw new Error("Seedance pricing parser is not implemented yet");

The seedance model is listed in the VideoModelId union in ai-pricing-definitions.ts. If refreshPricingCatalog() is called with Seedance in the catalog, it will throw and fail the entire refresh run. Either remove Seedance from the supported models list or implement the parser before shipping.

3. /api/v1/pricing/summary has no error handling

All Promise.all calls in the GET handler will throw unhandled exceptions if any external catalog fetch or DB lookup fails. Since this is a public, unauthenticated route, a single failing FAL endpoint will return a 500 with an internal error message instead of a graceful response.

// No try/catch anywhere in this function
const videoCosts = await Promise.all(
  SUPPORTED_VIDEO_MODELS.map(async (model) => { ... })
);

Wrap each lookup in a try/catch or use .catch(() => null) and filter nulls before building the range.

---

🟡 Medium Issues

4. Eight fragile regex parsers for FAL.ai HTML scraping (ai-pricing.ts)

Models like veo, kling, hailuo_standard, wan, and pixverse are priced by parsing raw HTML from FAL.ai's pricing pages with regexes. Any HTML restructure on FAL's side will throw:

throw new Error(`Unable to parse Veo pricing paragraph: ${paragraph}`);

There's no fallback to the last known-good database price when a parse fails. The behavior should be: fail gracefully, log a warning, and keep serving the last valid price from the DB rather than blowing up the entire refresh run.

5. externalCatalogCache Map grows unbounded

const externalCatalogCache = new Map<string, ExternalCacheValue>();

This module-level Map is never pruned. In a long-running Node.js instance (Next.js on Vercel Node runtime), entries accumulate indefinitely. Add an eviction step when setting: before inserting, delete entries where expiresAt < Date.now(), or cap the map size.

6. ElevenLabs pricing is a hardcoded snapshot, not live

ai-pricing-definitions.ts includes static elevenLabsSnapshotPricing that never updates automatically. If ElevenLabs changes rates, the system silently uses stale prices until a manual code deploy. At minimum, document that this requires manual updates and add a staleness warning in the admin endpoint.

7. music-metadata audio parsing blocks the request thread for large files

parseBuffer() is called synchronously in the STT route handler for files up to 25 MB. For large files this is a measurable delay on the request. Consider moving this to a worker thread or using a streaming parser to extract duration without loading the full buffer.

---

🟢 Positive Observations

• Decimal.js for all financial math — consistent and correct, avoids floating-point drift in billing calculations.
• Admin override endpoint is properly auth-gated with requireAdminWithResponse() and records updated_by for audit purposes.
• Cron endpoint uses verifyCronSecret — consistent with other cron routes in the project.
• Dimension-based pricing with wildcard fallback is a clean design that will scale to new models without schema changes.
• 20% platform markup is centralized via applyPlatformMarkup() — easy to change in one place.
• Transaction-wrapped refresh runs with ai_pricing_refresh_runs audit table is a solid approach.
• Voice endpoint consolidation under /api/v1/voice/ with the legacy paths as re-exports is clean.

---

Minor Nits

• as PricingProductFamily and as PricingChargeUnit casts on raw DB string values (aiEntryToPrepared()) have no runtime validation. A Zod parse or explicit set membership check would prevent stale DB values from silently corrupting pricing logic.
• buildRange() in the summary route calls await on a non-async function — minor inconsistency, no functional impact.
• The 15-minute EXTERNAL_CACHE_TTL_MS and AbortSignal.timeout(15_000) are hardcoded. Consider moving to named constants with a comment explaining the rationale.
• ai-pricing.ts is ~1,286 lines. The FAL HTML parsers are good candidates to extract into a lib/services/ai-pricing/fal-parsers.ts sub-module before this grows further.

---

Summary

The architecture is sound and the live-pricing direction is the right call long-term. The three issues to resolve before merging are: (1) clarify the intended behavior of failed video billing, (2) gate or remove the Seedance model until its parser is implemented, and (3) add error handling to the public /pricing/summary endpoint so a single failing external catalog doesn't return a 500 to unauthenticated callers.