Video-generation gateways

A video-generation gateway is a hosted inference aggregator that resells many third-party video models (Veo, Kling, Seedance, Hailuo/MiniMax, Luma, Wan, LTX-Video, Pika, and others) behind a single account, a single API convention, and a single bill. You send a prompt and parameters to a gateway-owned endpoint; the gateway runs the underlying model on its own infrastructure and returns a hosted result URL. This skill covers fal.ai, Replicate, and WaveSpeed as the representative gateways, and the production decisions common to the category.

This skill is about routing and operating, not about art direction. Prompt craft, shot design, and motion direction for a specific model belong to model- or craft-specific skills. Here the questions are: which gateway, going through a gateway at all, how the async job contract works, how to discover a model's schema, how to pin a version, how to cap spend, how to fail over, and whose license governs the output.

When a gateway is the right route (and when it is not)

Use a gateway when (production heuristics):

• You need several video models behind one integration — e.g. Veo for premium shots, a Kling/Seedance tier for volume, and an audio or upscale model in the same pipeline — without signing and maintaining a separate contract, key, billing relationship, and SDK per vendor.
• You want one async job contract (submit → poll/webhook → fetch) instead of learning each vendor's bespoke long-running-job protocol.
• You are model-shopping or A/B testing and want to swap model IDs without re-plumbing auth, storage, and result handling.
• You want the gateway to host input uploads and output media on a CDN so you are not building that yourself.
• Your volume does not yet justify a direct enterprise contract, or the model has no first-party public API and is only reachable through resellers.

Go direct (or reconsider) when:

• You need a first-party capability the gateway does not expose — a parameter, resolution, region, private-model hosting, or SLA the reseller has not surfaced. Gateways expose a curated subset of each model's surface.
• You need contractual data-processing terms, HIPAA/regional residency, or a signed DPA that the gateway cannot pass through from the underlying vendor.
• You are at very high, steady volume where a direct committed-use contract beats per-call reseller margin.
• You need the provider's own moderation posture rather than the gateway's added moderation layer (see Content safety below).
• Latency is dominated by gateway queueing/cold-start overhead you cannot tune, and the direct API offers provisioned throughput.

Out of scope for this skill: direct model-provider APIs (Google Vertex/Gemini Veo, Kling's own API, Runway, etc.), local/self-hosted inference, training or fine-tuning, and image-only gateway use.

The three reference gateways at a glance

Volatile facts below verified 2026-07-10; re-verify model IDs, prices, and limits before shipping — the catalogs and rate cards move frequently.

Dimension	fal.ai	Replicate	WaveSpeed
Positioning	Generative-media-first (image/video/audio/3D), latency-focused	General model marketplace (community + "official" curated models), broadest catalog	Speed/latency-optimized inference for media
Catalog size (self-reported)	1,000+ models	100+ official models plus a large community catalog	700–1,000+ models
Submit endpoint	https://queue.fal.run/{model-id}	https://api.replicate.com/v1/predictions (community) or .../v1/models/{owner}/{name}/predictions (official)	https://api.wavespeed.ai/api/v3/{model-id}
Async job states	IN_QUEUE → IN_PROGRESS → COMPLETED	starting → processing → succeeded/failed/canceled	created → processing → completed/failed
Sync option	Queue is the primary path; short models also have a run endpoint	Prefer: wait header (default 60s) for sync	Poll or webhook; result endpoint
Billing basis	Per-output (per-second or per-video for video); GPU-per-second fallback	Per-output for official models; per-second GPU time for community models	Per-model per-run, scales with resolution/duration/count
Balance model	Prepaid credits	Prepaid credit (hard cap at $0 for new accounts)	Prepaid credits (non-expiring)
Charged on failure?	No — pay only for successful output	Community models bill for GPU time even on model errors; official models bill per output	Pay per use; check per-model behavior
Schema discovery	Model-endpoints API with expand=openapi-3.0	openapi_schema on the model version object	Per-model playground / API examples

Sources: fal queue and pricing docs; Replicate HTTP API and pricing docs; WaveSpeed predictions and pricing docs (see Sources).

Reading the table: the three overlap heavily on what you can call — the same headline models (Veo, Kling, Seedance) appear on all three — so differentiation is mostly in the operational contract: billing granularity, whether failures cost money, how you discover a schema, and how versions are pinned. Treat the catalog as necessary but not sufficient; the operational contract is where integrations succeed or fail.

The async job contract (the core of every gateway)

Video generation is slow (seconds to minutes), so every gateway is fundamentally an asynchronous job queue. Do not build on a synchronous request that blocks until the video is done — you will hit timeouts and lose work. The universal shape:

1. Submit a job → receive a job/request ID immediately.
2. Await completion by webhook (preferred) or polling (fallback).
3. Fetch the result (a hosted media URL) and download/persist it before it expires.

fal.ai queue

• Submit: POST https://queue.fal.run/{model-id} with header Authorization: Key $FAL_KEY. Returns request_id, response_url, status_url, cancel_url, queue_position.
• Status: GET https://queue.fal.run/{model-id}/requests/{request_id}/status (append ?logs=1 for runner logs). States: IN_QUEUE, IN_PROGRESS, COMPLETED.
• Result: GET https://queue.fal.run/{model-id}/requests/{request_id}.
• Cancel: PUT .../requests/{request_id}/cancel — removes queued jobs immediately; in-progress jobs get a best-effort cancel signal.
• Documented guarantee: "requests in the queue are never dropped."
• Verified 2026-07-10 from fal queue docs.

Replicate predictions

• Submit (community): POST https://api.replicate.com/v1/predictions with a version field. Submit (official): POST https://api.replicate.com/v1/models/{owner}/{name}/predictions (no version needed).
• Sync shortcut: add Prefer: wait (or Prefer: wait=5) to hold the HTTP connection up to 60s and return output inline for fast models — not suitable for multi-minute video.
• Deadline: Cancel-After header (5s–24h) auto-cancels the prediction itself (distinct from Prefer: wait, which only bounds the HTTP wait).
• Status/result: poll the URL in urls.get until succeeded or failed; output is null until done. Web debug view at urls.web.
• Verified 2026-07-10 from Replicate HTTP/create-prediction docs.

WaveSpeed predictions

• Submit: POST https://api.wavespeed.ai/api/v3/{model-id} → returns a task/prediction ID.
• Result: GET https://api.wavespeed.ai/api/v3/predictions/{id}/result. States: created, processing, completed, failed. outputs holds result URLs; timings.inference gives duration in ms.
• Data retention: predictions available for 7 days (verified 2026-07-10); persist outputs promptly.
• Verified 2026-07-10 from WaveSpeed get-result docs.

Webhooks vs polling (production heuristic)

Prefer webhooks for video: a minutes-long poll loop wastes requests and risks rate limits, and holding a connection open is fragile. Fall back to polling only when you cannot expose a public HTTPS endpoint (local dev, locked-down networks). If you poll, back off — image tasks ~2s, video tasks ~5s, then widen the interval (WaveSpeed's documented guidance, and a sound default everywhere). Never poll a single job multiple times per second.

Design every webhook handler to be idempotent: gateways retry deliveries and may deliver more than once for the same job ID.

Webhook security and reliability

Webhook payloads arrive from the public internet — verify them.

fal.ai signs webhooks with ED25519. To verify (verified 2026-07-10):

1. Read headers X-Fal-Webhook-Request-Id, X-Fal-Webhook-User-Id, X-Fal-Webhook-Timestamp, X-Fal-Webhook-Signature (hex).
2. Fetch the JWKS from https://rest.fal.ai/.well-known/jwks.json and cache it up to 24h.
3. Reject if the timestamp is outside ±300s (5 min) of now.
4. Reconstruct the signed message (request ID + user ID + timestamp + SHA-256 of the raw body, newline-joined) and verify the signature against the JWKS ED25519 public keys.
5. Optionally IP-allowlist using webhook_ip_ranges from https://api.fal.ai/v1/meta.

fal delivery/retry (verified 2026-07-10): initial delivery timeout 15s; on failure it retries 10 times over 2 hours. The success payload carries status: "OK", results, and a seed; a request error carries status: "ERROR"; a serialization failure sets payload: null with a payload_error. The request_id is stable across retries; gateway_request_id reflects the latest attempt.

Replicate lets you filter which events fire via webhook_events_filter (e.g. ["completed"]); start and completed are always sent. Replicate supports webhook signature verification via a signing secret — verify before trusting a payload. (Verify current mechanism in Replicate's webhook docs before shipping; verified conceptually 2026-07-10.)

WaveSpeed delivers a POST with id, model, input, outputs (on completed), urls.get, status, and error (on failure). Confirm current signing details in WaveSpeed's webhook docs before relying on them.

Schema discovery per model

Each model exposes a different input surface (resolution enums, duration caps, image/video reference slots, audio toggles, seed, negative prompt). Never hardcode assumptions — discover the schema so your integration validates inputs and adapts when a model updates.

• fal.ai: the model-endpoints API accepts expand=openapi-3.0 and returns the full OpenAPI 3.0 schema for an endpoint in the openapi field. Query by endpoint_id (e.g. fal-ai/veo3) to get JSON Schema for every input parameter. Use list/search modes to enumerate the catalog. (Verified 2026-07-10.)
• Replicate: fetch the model version object at /v1/models/{owner}/{name}/versions/{version_id} and read openapi_schema.components.schemas for the Input and Output schemas. (Verified 2026-07-10.)
• WaveSpeed: each model's page/playground exposes the parameter set and auto-generates Python/JS/cURL examples; there is no single documented cross-catalog OpenAPI export as of 2026-07-10 — treat per-model docs as the source of truth.

Heuristic: cache the discovered schema per (model-id, version) and re-fetch on a version change or on the first 422/validation error. Building a small schema-validation layer in front of the gateway catches parameter-shape breakage before you spend money on a failed job.

Version pinning and deprecation

Reproducibility is the reason to care about versions: the same prompt on a silently-updated model can produce different motion, framing, or quality.

• Replicate has the most explicit model: community models are addressed as {owner}/{model}:{64-char-version-hash}. Pinning the hash guarantees identical behavior over time; omitting it (or using an official model addressed as {owner}/{model} with no hash) means Replicate keeps you on the latest maintained version. Trade-off: pin the hash for reproducibility and to avoid surprise behavior changes; use the unpinned official-model form when you want automatic upgrades and less maintenance. Replicate documents pinning as a reproducibility best practice. (Verified 2026-07-10.)
• fal.ai and WaveSpeed address models by a stable string ID (e.g. fal-ai/veo3, or versioned IDs like fal-ai/kling-video/v2/...). Version granularity is encoded in the ID the vendor publishes; watch changelogs for new IDs and deprecations of old ones.

Deprecation is a real operational risk on gateways, not a hypothetical. Underlying models are retired or superseded (a new Kling/Veo/Seedance version lands and the old endpoint is sunset). Mitigations (heuristics):

• Track each gateway's changelog and pin to explicit versioned IDs where offered.
• Keep a config-driven model map (env/config, not inline strings) so a deprecated ID is a one-line change, not a code hunt.
• Add a catalog/health probe in CI that resolves each model ID you depend on and alerts when one stops resolving or its schema changes.
• Maintain a named fallback model per capability tier so a sunset endpoint degrades to a substitute instead of failing the pipeline.

Spend controls

Video is the most expensive generative modality per call, and gateway spend can run away fast under retries or a bad loop. Controls differ sharply:

• fal.ai: prepaid credits are the primary cap — you cannot spend past your balance. You pay only for successful outputs; you are not charged for server errors (HTTP 500+) or time spent queued. fal's own docs do not advertise per-project budget alerts as of 2026-07-10 — build your own spend accounting from the billing/platform APIs.
• Replicate: monthly spend limits were deprecated (July 1, 2025); prepaid credit is now the hard cap — work stops at a $0 balance (prepaid credit for new accounts since July 16, 2025). There are no native budget alerts post-deprecation; low-balance alerting is left to you or third-party tools. Note the failure-billing asymmetry: community models bill for GPU seconds even when the model errors, so a crash-looping community model still costs money; official models bill per output. (Verified 2026-07-10.)
• WaveSpeed: prepaid, non-expiring credits; usage-based per-model pricing where cost scales with resolution, duration, and output count. The cost estimate is shown before you run, and can be queried via the pricing API. Monthly credit lines exist for approved/enterprise accounts. (Verified 2026-07-10.)

Production heuristics for cost safety:

• Keep the working balance small and top up deliberately — prepaid balance is your real circuit breaker on all three.
• Estimate before you submit: video cost scales with duration × resolution × output count (× audio, on some models). A 8s 1080p Veo clip with audio is dramatically pricier than a 5s 720p Kling clip — know the unit price. Example (fal, verified 2026-07-10): Veo 3.1 is billed ~$0.20/s at 720p/1080p without audio and ~$0.40/s with audio; Kling 2.5 Turbo Pro ~$0.07/s. Prices change — re-verify.
• Cap concurrency and retries in your own code; an unbounded retry loop on a failing model is the classic runaway on any prepaid gateway.
• Prefer gateways/models that do not charge for failures for experimentation, and prefer per-output over per-GPU-second when your job latency is unpredictable.
• Instrument cost per job (job ID → model → parameters → charge) so you can attribute and alert; do not rely on the gateway to warn you.

Retries and cross-gateway failover

Retries. Transient failures (429 rate limit, 5xx, capacity/cold-start timeouts) are expected. Retry with exponential backoff and jitter, an idempotency key you control, and a cap (both attempts and total spend). Because a submitted-but-slow job is not the same as a failed job, distinguish "submit failed" (safe to retry) from "job running slowly" (poll/await, do not resubmit — resubmitting double-charges). Honor Retry-After when present.

Cross-gateway failover. Because the same headline models appear on multiple gateways, a gateway is a substitutable route, which is the strongest operational argument for the aggregator pattern. To fail over:

• Abstract behind a capability interface (generateVideo(prompt, tier, constraints)), not a gateway-specific call, so the router picks a provider+model.
• Maintain a model map per capability tier across ≥2 gateways (e.g. "premium cinematic" → Veo on fal or Veo on Replicate; "volume" → Seedance on WaveSpeed or Kling on fal).
• Normalize at the seams: each gateway has its own auth, submit shape, status vocabulary, parameter names, and result JSON. Failover only works if you translate inputs/outputs to a canonical form.
• Watch for output drift: the same model on two gateways can differ in exposed parameters, default resolution, moderation, or even minor pipeline details. Failover preserves availability, not necessarily identical output — treat a failover result as a near-substitute, not a bit-identical one, and re-verify quality-critical shots.
• Keep separate prepaid balances and keys funded on each gateway so failover is not blocked by a zero balance on the backup.

Content-safety behavior differs by gateway

Two moderation layers stack: the underlying model's own safety behavior and the gateway's added moderation. They are not identical across gateways, so the same prompt on the same model can be accepted on one gateway and blocked on another.

• fal.ai integrates OpenAI's Omni moderation API for real-time filtering, maintains an NSFW content policy, and partners with NCMEC/Thorn (CSAM) and StopNCII (NCII). Some image models expose an enable_safety_checker toggle (default on); video-model moderation posture is model-dependent. Users remain bound by fal's Acceptable Use Policy. (Verified 2026-07-10.)
• Replicate enforces its terms and the underlying model's restrictions; moderation strictness is largely inherited from the specific model.
• WaveSpeed applies its own acceptable-use terms atop each model.

Heuristics: (1) test moderation behavior per gateway, not once — a prompt cleared on your primary may be refused on your failover route, breaking availability assumptions. (2) Handle a moderation rejection distinctly from a transient error — it is deterministic, so retrying or failing over the same prompt often just fails or double-charges; instead surface it, adjust the prompt, or route to a model with a different policy. (3) Do not assume a gateway strips or adds a watermark, C2PA/provenance signal, or safety filter the same way the direct API does — verify.

Data governance and commercial-rights passthrough

Two separate questions: where does your data live, and who may use the output commercially.

Data retention / residency:

• fal.ai: input uploads and output media go to fal's CDN (v3.fal.media/...). Retention is controllable via the X-Fal-Object-Lifecycle-Preference header (expiration_duration_seconds, or null for no expiry). Request/response JSON is stored 30 days by default. Download outputs before they expire — expired files are unrecoverable. (Verified 2026-07-10.)
• Replicate: for API-created predictions, input params, outputs, and logs are auto-removed after ~1 hour by default — you must persist outputs yourself. (Verified 2026-07-10.)
• WaveSpeed: predictions retained 7 days. (Verified 2026-07-10.)

Because outputs are transient, always download and re-host anything you need; never link end users directly to a gateway CDN URL for durable content. For regulated data, confirm whether the gateway can offer a DPA or regional processing — most reseller flows send your inputs to third-party model infra.

Commercial-rights passthrough (the critical governance point): a gateway granting you rights to its service output does not override the underlying model's license. Rights flow through the model, not the gateway.

• Replicate's terms grant you "all right, title and interest" in the Output including commercial use — but explicitly subject to Third-Party Terms determined by the model you used. Some models (e.g. permissive ones like Stable Diffusion) impose no ownership claim; others restrict commercial use. (Verified 2026-07-10.)
• The same logic applies on fal and WaveSpeed: the model card / underlying vendor license governs whether you may use a clip commercially, whether attribution is required, and whether the vendor asserts training or usage rights over your inputs.

Heuristic: before using a gateway-generated video commercially, resolve the underlying model's license and the model provider's terms, not just the gateway's ToS. Premium models (Veo, Kling) frequently carry provider-specific commercial terms and per-clip restrictions; a gateway reselling them does not relax those terms. Record, per model you ship on, the license basis for commercial use so a later dispute or model swap is auditable.

Troubleshooting

Symptom	Likely cause	Action
Submit returns fast but no video	Working as designed — it is async; you got a job ID, not a result	Poll status/result or await the webhook; do not treat the submit response as the output
Webhook never arrives	Endpoint not public HTTPS, handler slow (>15s on fal), or signature check rejecting valid payloads	Verify reachability, respond quickly then process async, confirm signature logic; fall back to polling
Duplicate results processed	Gateway retried delivery	Make the handler idempotent keyed on job/request ID
422/validation error	Wrong parameter shape for this model/version	Re-fetch the model schema (fal openapi-3.0, Replicate openapi_schema); validate inputs before submit
Charged but no usable output	Community-model GPU-time billing on error (Replicate), or output expired before download	Prefer pay-per-output models for flaky work; download outputs immediately; check retention window
Model ID stopped resolving	Version deprecated/sunset	Update the config-driven model map to the new versioned ID or fail over to the tier fallback
Prompt blocked on one gateway, fine on another	Different stacked moderation layers	Treat as deterministic; adjust prompt or route to a different-policy model — do not blind-retry
Runaway spend	Unbounded retries or a bad loop against a prepaid balance	Cap retries/concurrency, keep balance small, instrument per-job cost
Output looks different after a "no-op" change	Unpinned version silently updated	Pin explicit version hashes/IDs; add a schema/behavior probe in CI
Result URL 404s later	Gateway CDN retention elapsed	Re-host durable content; never serve end users a raw gateway URL

Complete examples (illustrative — verify IDs/prices before use)

Example A — Choose a route for a batch of short social clips

Intent: generate 500 five-second 720p vertical clips/day, cost-sensitive, some prompts near content-policy edges, must not exceed a fixed daily budget.

Reasoning: volume + cost sensitivity favors a per-second/per-run model on a gateway that (a) does not charge for failures and (b) supports a hard prepaid cap. Pick a mid-tier model (e.g. a Kling/Seedance turbo tier) rather than Veo. Put a second gateway behind the same capability tier for failover, since availability of a single model on a single gateway is not guaranteed at 500/day.

Plan: primary = fal queue.fal.run submit + ED25519-verified webhook, prepaid balance sized to the daily budget as the circuit breaker; secondary = same-tier model on WaveSpeed with its own funded balance. Router picks by health + queue depth. Per-job cost is logged (job ID → model → params → charge). Moderation rejections are surfaced to a review queue, not retried.

Expected result: clips delivered async, budget bounded by prepaid balances, a sunset or outage on one route degrades to the other. Failure modes: forgetting to fund the backup balance (failover silently fails); blind-retrying moderation rejections (wasted spend); serving raw CDN URLs that later 404.

Example B — Pin a reproducible cinematic model on Replicate

Intent: a hero shot must render identically across months for a marketing pipeline; commercial use required.

Plan:

1. Resolve the exact version: GET /v1/models/{owner}/{name}/versions and pick the hash; store {owner}/{name}:{version_hash} in config.
2. Submit POST /v1/predictions with that pinned version, a webhook, and webhook_events_filter: ["completed"]; set a Cancel-After deadline as a spend guard.
3. Discover/validate inputs against openapi_schema.components.schemas.Input before submitting.
4. On completion, download and re-host the video within the ~1-hour retention window.
5. Rights check: confirm the underlying model's Third-Party Terms permit commercial use, and record that basis, before publishing.

Why structured this way: pinning the hash locks reproducibility; the schema check prevents parameter drift from silently changing output; the retention-aware download prevents a lost result; the rights check prevents shipping a clip whose model license forbids commercial use. Failure modes: using the unpinned official-model form (auto-upgrades break reproducibility); assuming Replicate's ownership grant overrides the model's restrictive license.

Example C — Discover a model's schema before integrating (fal)

Intent: integrate a video model you have not used; you do not know its duration cap, resolution enum, or whether it takes an image reference.

Plan: call the fal model-endpoints API with endpoint_id set to the model (e.g. fal-ai/veo3) and expand=openapi-3.0; read the returned openapi field's JSON Schema for the input object. Generate your request typing and a validation layer from it; cache it keyed on the model ID; re-fetch on the first 422 or a changelog-announced version bump.

Why: this makes the integration self-describing and resilient to the model adding/renaming a parameter, instead of hardcoding a guessed body that breaks on the next model update. Failure mode: skipping discovery and hardcoding parameters copied from a blog post that lags the current schema.

Sources

All URLs verified 2026-07-10.

• fal.ai queue API — https://fal.ai/docs/model-apis/model-endpoints/queue
• fal.ai webhooks (ED25519, retries, JWKS) — https://fal.ai/docs/model-apis/model-endpoints/webhooks
• fal.ai pricing / prepaid credits / pay-per-success — https://fal.ai/docs/documentation/model-apis/pricing and https://fal.ai/pricing
• fal.ai model search / OpenAPI schema discovery — https://fal.ai/docs/platform-apis/v1/models and https://docs.fal.ai/model-apis/model-endpoints
• fal.ai data retention / media expiration header — https://fal.ai/docs/documentation/model-apis/media-expiration
• fal.ai Trust & Safety (OpenAI Omni moderation, NCMEC/Thorn/StopNCII) — https://fal.ai/legal/trust-and-safety
• fal.ai Acceptable Use Policy — https://fal.ai/legal/acceptable-use-policy
• Replicate create-a-prediction (sync/async, Prefer, Cancel-After, webhook_events_filter) — https://replicate.com/docs/topics/predictions/create-a-prediction
• Replicate HTTP API reference (version formats, openapi_schema, rate limits, ~1h retention) — https://replicate.com/docs/reference/http
• Replicate model versions (pinning, reproducibility) — https://replicate.com/docs/topics/models/versions
• Replicate official models — https://replicate.com/docs/topics/models/official-models
• Replicate pricing (per-second GPU hardware, per-output official) — https://replicate.com/pricing
• Replicate billing (spend-limit deprecation, prepaid credit) — https://replicate.com/docs/topics/billing
• Replicate rate limits — https://replicate.com/docs/topics/predictions/rate-limits
• Replicate Terms of Service (Output ownership, Third-Party Terms) — https://replicate.com/terms
• WaveSpeed predictions / get-result (endpoints, states, 7-day retention) — https://wavespeed.ai/docs/get-result and https://wavespeed.ai/docs/what-are-predictions
• WaveSpeed pricing (per-model, credits) — https://wavespeed.ai/docs/how-pricing-works and https://wavespeed.ai/pricing
• WaveSpeed webhooks — https://wavespeed.ai/docs/docs-api/webhooks (confirm current signing/retry details before shipping)

Model-specific prices (Veo 3.1 ~$0.20–0.60/s by resolution/audio; Kling 2.5 Turbo Pro ~$0.07/s) are from fal's pricing pages, verified 2026-07-10, and are volatile — always re-verify against the live rate card.