Wan 3.0 Is on Every Platform. Is It the Same Product?

Wan 3.0 is no longer only a model launch. It is becoming a distribution event.

Alibaba Cloud released the model earlier this month with text, image, audio, video, document, and public-web-page inputs, plus videos up to 30 seconds. On August 24, availability announcements spread across creation platforms and API routes, including fal, Runway, and OpenRouter. That reach matters to a small team: you may be able to add the model through a tool you already use, choose a regional endpoint, or replace a blocked route without rebuilding the whole product.

But a shared model name does not make those routes interchangeable.

One provider may rewrite the prompt by default. Another may expose the rewrite. One may return the seed and output dimensions; another may not. Uploads, safety checks, regional execution, queue behavior, result lifetime, retry semantics, prices, terms, and provenance can differ. Even a technically valid video can fail your product because it changed a product claim, lost a reference detail, invented text, shipped without captions, or arrived after the customer had left.

This guide is for AI app builder users, nontechnical founders, and small product teams adding generated video to onboarding, product demos, marketing, education, or creator workflows. The central judgment is: treat each model-provider combination as a separate product release until it earns the same acceptance receipt.

You will leave with precise terms, a frozen evaluation set, a complete hypothetical product scenario, an accepted-deliverable cost formula, a machine-readable provider receipt, failure drills, a decision matrix, and a 48-hour rollout. This is not a Wan 3.0 benchmark and does not claim that one provider is best. It is a way to discover whether a route is fit for the promise you actually sell.

Start with what launched, and keep the claims bounded

Wan 3.0's current capability surface is unusually broad. Alibaba Cloud's official API reference documents text-to-video, first-frame and first/last-frame generation, reference-based generation, and inputs made from images, video, audio, files, or a public web link. It describes asynchronous jobs that typically take one to five minutes and outputs up to 30 seconds. The official repository presents the same family of modes and says a request can draw on as many as 20 multimodal assets.

Those are product-surface facts, not independent quality results. Statements such as “pixel-perfect consistency,” “cinematic realism,” or “production-grade” are vendor descriptions. Alibaba's own launch article says audio texture and on-screen text accuracy are still improving. It also lists Model Studio prices of $0.05, $0.10, and $0.20 per generated second for 480P, 720P, and 1080P. Prices and preview behavior can change, so record the page and check time rather than copying the figures into a permanent margin model.

The distribution wave adds availability, not evidence that every route is behaviorally equal. fal's Wan 3.0 API schema, for example, exposes defaults for generated audio, prompt expansion, enhanced reasoning, safety checking, duration, aspect ratio, and resolution. Its output schema includes a seed, duration, actual prompt, and video metadata. That is a concrete serving contract. It is not proof that another platform uses the same defaults, model revision, safety layer, scheduler, or post-processing.

The timely founder question is therefore not “Can I access Wan 3.0?” It is: which route can deliver my exact customer job, with evidence I can explain, pay for, recover, and repeat?

Separate model identity, route identity, and product identity

Four terms prevent most portability confusion.

Model identity is the model family and, where available, its precise revision or deployment identifier. Wan 3.0 is a family name. It may not disclose the exact weights, build date, or serving configuration behind every third-party listing. Route identity is the complete provider path: account, region, endpoint, model ID, API version, defaults, moderation, prompt transformation, scheduler, file service, output encoder, and terms. Two routes can advertise the same model while producing different requests and receipts. Generation attempt is one submitted job. It can finish successfully at the API layer and still produce an unusable asset. Attempts matter for spend and reliability, but they are not customer value. Accepted deliverable is an output that passes the product's frozen checks and can be used in the promised channel. It includes the final video plus any required caption file, transcript, alt description, provenance record, claim approval, and source-asset record.

Layer	Stable question	Evidence to retain
Model	What model or revision was requested and reported?	Provider model ID, revision if exposed, request time
Route	What service actually handled the job?	Provider, region, endpoint/API version, policy and defaults
Attempt	What happened to this request?	Idempotency key, task ID, states, timestamps, charges, error
Deliverable	Did the result fulfill the user promise?	Output digest, acceptance checks, reviewer, derivatives, decision

Do not collapse these layers into a badge that says “Powered by Wan 3.0.” A badge cannot tell support whether the provider rewrote the prompt, finance whether a failed job was charged, privacy whether a reference video crossed a region, or the user whether their exported clip still contains an unverified claim.

Freeze the user job before comparing providers

A provider test becomes meaningless if every person submits a different prompt and picks the nicest result. Start with one bounded product job.

Write it in customer language: “Turn an approved product page and three brand images into a 15-second vertical launch video, with accurate product color, no invented feature claims, intelligible English narration, captions, and a downloadable MP4 within four minutes.” That is stronger than “make a good promo video.”

Then freeze the job envelope:

Inputs: exact files, public page snapshot, prompt, negative instructions, language, and permission record.
Expected facts: product name, price, availability, dimensions, colors, and features the video may state.
Creative freedoms: camera movement, background, pacing, decorative props, and music style the model may invent.
Forbidden outputs: unsupported claims, competitor marks, unlicensed people or music, unreadable required copy, unsafe scenes, and private source material in public artifacts.
Delivery contract: aspect ratio, duration band, resolution, codec, audio, captions, file-size ceiling, deadline, and download lifetime.
Acceptance owner: who can approve factual claims, brand fidelity, accessibility, and final release.

Use at least 20 fixtures before a limited pilot: ordinary cases, dense pages, sparse pages, portrait and landscape references, non-English copy, difficult product colors, small text, silence, fast motion, and a provider error. A customer-facing marketing product should add claim-sensitive and rights-sensitive cases. A playful internal storyboard tool can use a lighter set because a failed draft has a smaller consequence.

The frozen set is not a public benchmark. It is evidence about your job. Do not publish a general “Provider A beats Provider B” ranking from it unless the sampling, repeated runs, revisions, settings, judges, and uncertainty genuinely support that claim.

Preserve the effective request, not only the user's prompt

Video APIs often transform an input before generation. A provider may expand a short prompt, infer a duration, analyze a document, fetch a public page, normalize media, apply safety filters, or add generation instructions. Those steps can improve quality, but they also change what you are evaluating.

fal's current schema explicitly exposes enable_prompt_expansion, enable_thinking, and an actual_prompt output. Alibaba Cloud's API accepts a structured media array and distinguishes reference media from strict first/last-frame control. The two surfaces already show why a generic adapter that sends prompt + files is insufficient.

For every attempt, preserve:

The user's original request and the version of your prompt template.
Ordered source-asset IDs, roles, digests, permissions, and upload destinations.
Every explicit parameter and every default you rely on.
The effective or expanded prompt when the route exposes it.
Safety outcomes and policy version, without storing sensitive content unnecessarily.
Model and provider identifiers, API/client version, region, and request time.
The final seed and output metadata where available.

Treat smart duration as a product choice, not a convenience. If a customer bought a 15-second placement, an automatically chosen 23-second clip is not equivalent even if it looks better. Treat prompt expansion the same way: it can be allowed for exploratory drafts, while a regulated, claim-sensitive, or precisely scripted job may need it disabled or reviewed.

Do not promise deterministic reproduction merely because a seed exists. The provider may revise the model, scheduler, moderation, encoder, or infrastructure. A seed is one useful field in a replay record, not a warranty that tomorrow's pixels will match.

Make the asynchronous job a visible product state machine

Video generation is not a single request-response event. Alibaba's reference describes a create-and-poll flow and says the returned task_id is valid for 24 hours; it specifically warns against submitting duplicates when polling should be used. fal recommends queues and webhooks for long-running requests. That operational difference belongs in the product design.

Use explicit states:

accepted → queued → processing → generated → validating → ready

and terminal alternatives:

rejected, failed, expired, cancelled, needs_review, or delivery_failed.

The UI should tell the customer which state they are in and whether leaving the page is safe. A timeout in your frontend is not evidence that the provider failed. Retrying blindly may create two chargeable videos. Assign an idempotency key at your product boundary, map it to exactly one active provider job, and reconcile delayed webhooks or polling responses before creating a successor.

Download the finished asset into storage you control before a provider URL expires. Record a digest, media metadata, and the source task. Verify that the file opens, duration and dimensions match the receipt, audio is present when required, and the exported derivative is the same object your reviewer approved. A 200 response from a result endpoint does not prove that the customer received the file.

Measure these operational outcomes separately:

queue delay;
generation time;
validation time;
end-to-end ready time;
provider failure rate;
duplicate-attempt rate;
late-success reconciliation rate;
delivery failure rate;
user abandonment before ready.

A beautiful median result can still be a poor product route if the tail regularly misses the customer's moment.

Evaluate the deliverable in the channel where it will ship

Do not review only the provider preview. Export the actual 9:16, 1:1, or 16:9 derivative, compress it the way your app does, play it on a phone, and inspect it with sound both on and off.

Use a layered acceptance suite:

Check	Test question	Example blocker
Source fidelity	Are required people, products, colors, and layouts preserved?	Product color changes materially
Claim fidelity	Is every objective statement supported by the approved fact sheet?	Invented battery life or discount
Narrative	Does the sequence communicate the intended job without a prompt explanation?	Call to action appears before the product
Motion	Are cuts, physics, hands, faces, and object continuity acceptable?	Product duplicates between shots
Audio	Is speech intelligible and music/dialogue appropriate?	Brand name is mispronounced
Text	Is required copy accurate and readable at delivery size?	Price or URL is malformed
Safety and rights	Are consent, likeness, trademark, music, and policy checks satisfied?	Unapproved real-person likeness
Accessibility	Are captions, transcript, and necessary visual description present?	Meaning exists only in uncaptioned speech
Technical delivery	Do codec, dimensions, duration, size, and playback pass?	Social channel rejects the file
Provenance	Can the team connect output to sources, transformations, and approval?	No source or reviewer record

The FTC's advertising-substantiation policy says advertisers should possess a reasonable basis for express and implied objective claims before dissemination. The practical product lesson is simple: a source page is not permission for a video model to improvise. Extract an approved fact sheet, compare spoken and on-screen claims against it, and hold anything unknown for a human.

Accessibility is also part of the asset, not cleanup after export. W3C's audio and video guidance covers captions, transcripts, description of meaningful visual information, and accessible players. Your provider does not need to generate every derivative, but your product promise must name who or what does. If a generated voice says the only essential information and no captions survive the export, the video is not ready for many users.

Measure cost per accepted deliverable

The headline price per generated second is only one input. A 30-second 1080P attempt at Alibaba's published launch rate would be $6, but one accepted customer video may require drafts, failed jobs, higher-resolution reruns, caption correction, claim review, and storage or delivery work.

Use this operating formula:

accepted-deliverable cost = all attempt charges + supporting model/API charges + review labor + correction labor + storage/delivery cost + refund/recovery cost

Then divide total cost by accepted deliverables, not submitted attempts.

Suppose 100 customer jobs create 170 paid attempts. Eighty videos pass automatically, 12 pass after human correction, five are refunded, and three remain unresolved at the cutoff. Your denominator is 92 accepted deliverables, not 100 jobs or 170 generations. Keep the unresolved cohort visible rather than quietly treating it as failure or success.

Track correction minutes alongside money. A cheaper route that saves $0.80 per attempt but adds six minutes of review can be more expensive for a small team. Also separate draft and final tiers: low-resolution generation may be appropriate for story selection, while only selected drafts earn a final-resolution attempt. Alibaba's published price ladder makes that strategy possible on its route, but test whether the lower-resolution draft predicts the final output well enough for your workflow.

Do not assume a failed request is free, a safety rejection is charged, or a cancelled queue job is refundable. Record the provider's actual billing event and reconcile invoices against attempt IDs. Unknown billing semantics are a pilot blocker when your product offers fixed-price generation.

Recheck rights, privacy, region, and provenance for every route

Multi-input video can send more sensitive material than a text prompt: unreleased product pages, customer screenshots, employee voices, real-person reference video, pitch decks, and brand assets. A convenient new provider expands the data path.

Alibaba Cloud's regional deployment documentation separates the selected data-storage region from the service deployment scope and says regional domains differ in model availability, limits, and operational features. Its Model Studio terms place responsibility on the customer to obtain needed rights and independently evaluate output; they also describe cross-border processing and say customer content is not used to improve models without separate consent. Those statements apply to that contractual route. They must not be copied onto a marketplace, reseller, or unrelated provider without checking its own terms.

For each route, answer:

Which entity contracts with us, and which subprocessors or underlying providers are involved?
Where are inputs, logs, intermediate assets, and outputs stored and processed?
How long do uploads, tasks, and results persist, and how are they deleted?
Are inputs or outputs used for training, review, abuse detection, or service improvement?
Who may upload a person's face, voice, copyrighted work, private page, or customer file?
What moderation is applied, can it be changed, and what happens to rejected content?
Who owns or may use the output, and what third-party rights remain our responsibility?
Is provenance metadata created, preserved through download and transcode, or removed?

C2PA's Content Credentials 2.4 specification provides a format for binding provenance assertions and now includes expanded live-video support. A credential can help preserve model, source, edit, and signer history. It cannot prove that a product claim is true, that a likeness was consented to, or that a platform will retain the metadata after transcoding. Keep the internal receipt even when the delivery channel strips credentials.

Scenario: a product-page-to-video feature chooses its first production route

Imagine LaunchCut, a small app that turns a public product page, a logo, and three approved images into a 15-second vertical launch video. The user can edit the script, but the app generates the shots, narration, music, and captions.

The team sees Wan 3.0 appear on several platforms and wants immediate failover. Instead of wiring three provider names behind one dropdown, it freezes 30 evaluation jobs:

12 ordinary software and consumer-product pages;
four pages with pricing footnotes;
four pages containing tables or small UI text;
four Chinese or bilingual pages;
three reference sets with difficult brand colors;
two consented real-person examples;
one simulated timeout that later completes.

LaunchCut extracts a fact sheet before generation. Only the product name, approved feature names, current price, availability region, and call to action may appear as facts. Backgrounds and camera movement are creative. Competitor logos, testimonials, performance superlatives, and invented discounts are forbidden.

Route A exposes an expanded prompt and completes quickly, but five clips turn nuanced feature descriptions into absolute claims. Route B is slower and costs more per accepted clip, yet preserves the fact sheet better. Route C produces strong video but its current document route requires a public URL that violates LaunchCut's private-draft promise.

The team does not crown a universal winner. It chooses Route B for final customer delivery, Route A for visibly labeled editable drafts with claim scanning, and holds Route C for private documents until a suitable upload and deletion contract is verified. It keeps a manual storyboard fallback when no route can meet the deadline.

The hypothetical result illustrates the decision method; it is not a real YBuild customer, a measured provider comparison, or evidence about current Wan 3.0 quality.

Use a provider acceptance receipt

Keep one versioned receipt per model-provider combination. Unknown fields stay unknown; they do not inherit a reassuring answer from another route.

yaml

video_provider_release:
  product_job: "approved product page to 15s vertical launch video"
  model_family: "Wan 3.0"
  provider: "candidate-provider"
  provider_model_id: "recorded value"
  api_schema_version: "recorded value"
  region:
    storage: "verified value or unknown"
    inference: "verified value or unknown"
  request_contract:
    source_asset_digests: ["sha256:..."]
    source_permissions: "receipt-id"
    prompt_template_version: "launchcut-v4"
    prompt_expansion: "on | off | unavailable | unknown"
    effective_prompt_returned: true
    duration_seconds: 15
    aspect_ratio: "9:16"
    resolution: "1080p"
    audio: true
    safety_policy_version: "verified value or unknown"
  job_contract:
    product_idempotency_key: "job-..."
    provider_task_id: "task-..."
    terminal_states: ["ready", "rejected", "failed", "expired"]
    late_success_reconciliation: "tested"
    result_copied_to_owned_storage: true
  acceptance:
    fixture_set: "video-acceptance-2026-08-25"
    claim_fidelity: "pass"
    source_fidelity: "pass"
    motion_audio_text: "pass"
    accessibility_derivatives: "pass"
    technical_delivery: "pass"
    output_sha256: "sha256:..."
    reviewer: "role, not secret"
  economics:
    attempt_charge: "invoice-linked"
    total_correction_minutes: 0
    accepted_deliverable_cost: "calculated"
  data_and_rights:
    retention_and_deletion: "verified contract/reference"
    training_use: "verified contract/reference"
    likeness_and_source_rights: "receipt-id"
    provenance_preserved: "yes | no | partial | unknown"
  decision: "ship | draft-only | limited | hold | reject"
  recheck_on: ["model change", "schema change", "policy change", "term change"]

The receipt is not compliance theater. Product uses it for routing; support uses task and output IDs; finance uses attempt and invoice links; reviewers use the fact sheet and fixtures; privacy uses the data path; leadership uses the final decision and recheck triggers.

Run the failure drills that demos hide

Before production, force at least these failures:

Prompt transformation drift: change the provider or disable expansion and check whether prohibited claims reappear.
Reference-order drift: swap two source assets and confirm your adapter preserves positional meaning.
Late success after timeout: let the client time out, then deliver the webhook. Verify that no duplicate attempt is created.
Expired result: delay download until the provider URL is unavailable. Confirm owned storage or a clear recovery path.
Safety disagreement: send a consented but policy-sensitive fixture. Confirm the user sees a useful state without exposing internal moderation detail.
Private source leak: attempt to pass a private page through a public-link field. The product should block it rather than make the page public.
Claim mutation: use a page with a qualified price or limited-region feature. Reject any video that removes the qualification.
Caption loss: transcode and publish the final derivative. Verify captions or a separate accessible alternative still travels with it.
Provider revision change: rerun the frozen set after a model ID, default, client, or policy changes.
Failover under load: route bounded jobs to the alternate and compare accepted outcomes, not only successful HTTP responses.

This is product-specific pre-deployment evidence, not a claim that the fixture set covers every harm. The NIST Generative AI Profile treats testing before deployment, ongoing monitoring, incident learning, and documented limits as connected practices. Preserve the failed cases and rerun them after a material route change instead of deleting embarrassing outputs from the evaluation set.

Do not average away a critical failure. A route that produces nine impressive clips and one unapproved medical claim, leaked draft, or duplicated charge does not receive a 90% launch grade. Define hard blockers separately from scored quality.

Choose a posture instead of pretending every route is interchangeable

Posture	Use when	Product behavior
Ship	Route passes hard gates and accepted-cost/reliability targets	Customer delivery allowed; receipt retained
Draft-only	Creative value is high but claims, rights, or consistency require editing	Clearly label editable draft; no direct publish
Limited	Evidence is sufficient for a narrow input/channel/region	Enforce the boundary in routing and UI
Hold	Important terms, retention, billing, revision, or failure behavior remain unknown	Keep testing; do not accept customer promise
Reject	Route fails a hard gate or cannot support recovery	Remove from production routing

A second provider is not automatically a fallback. It earns a limited or ship posture with the same frozen job, even if its API integration is easy. Conversely, portability does not require pixel-identical output. It requires equivalent fulfillment of the product promise within declared tolerances.

This framework fits customer-facing generated video, product demos, ads, onboarding, education, creator tooling, and internal brand production. It is intentionally heavier than needed for disposable personal ideation. It also does not certify copyright, privacy, advertising, accessibility, or AI-law compliance; qualified owners must decide those obligations for the actual market and content.

A 48-hour rollout for a small team

Hours 0–4: Name one product job, one owner, hard blockers, creative freedoms, delivery requirements, and the current primary route. Capture the route's model ID, schema, defaults, region, pricing page, terms, and task lifecycle. Hours 4–12: Build 20–30 frozen fixtures and an approved fact sheet per fixture. Record source permissions. Add difficult cases, one timeout, one safety case, one private-source case, and one output-expiry case. Hours 12–24: Run the primary and one candidate route. Preserve effective requests, task events, outputs, charges, and correction time. Review final delivery derivatives on the target device and channel. Hours 24–32: Complete the provider receipts. Separate hard blockers from scored dimensions. Calculate cost per accepted deliverable and ready-time tails. Resolve unknown retention, training-use, rights, and billing fields or mark the route hold. Hours 32–40: Run late-success, duplicate-prevention, result-copy, caption, private-source, and failover drills. Confirm support can find one attempt from the customer job ID. Hours 40–48: Choose ship, draft-only, limited, hold, or reject. Release only the enforced boundary. Schedule rechecks for any model, API schema, default, safety policy, price, term, or region change.

Wan 3.0's expanding distribution is genuinely useful. It can reduce integration friction and give small teams more routing options. The mistake would be treating availability as substitutability. A model name tells you what to evaluate; the provider acceptance receipt tells you what you can responsibly promise.

References