Video generation API comparison Async jobs OpenAI Gemini

Video Generation API Comparison 2026: Dimensions That Matter

Compare video generation APIs in 2026 with dated tables for job shape, edits/extensions/characters, persistence, and ops—without fabricated rankings.

AveMujica API Jul 15, 2026 8 min read

Last checked: 2026-07-15.

A useful video generation API comparison in 2026 evaluates job mechanics, limits, media durability, safety, and operability under your constraints. Public demos emphasize cinematic quality; production systems fail on expired content URLs, unbounded duration inputs, weak tenant isolation, and opaque failure states. This article lists comparison dimensions, dated candidate facts from official docs, and how an OpenAI-compatible gateway changes the shopping list—without declaring an unsupported winner or inventing prices, quality scores, or model availability.

Comparison rules of engagement

Re-read first-party docs on the decision date (here: 2026-07-15) and again before contract signature.
Compare dimensions with evidence links; do not paste influencer benchmark tables as truth.
Measure latency and success rate in your region with your prompts.
Separate model quality experiments from platform requirements (SSO keys, team quotas, audit logs).
Record model ids exactly as enabled in your gateway; marketing names drift.
Do not invent a ranked leaderboard.

Primary documentation hubs reopened for this revision:

OpenAI video generation guide
OpenAI Videos API reference
OpenAI deprecations — Videos API / Sora 2 shutdown 2026-09-24
OpenAI pricing
Google Gemini video generation (Gemini Omni Flash + Veo families)
Runway developer docs
Luma docs
Kling / ByteDance / Volcengine first-party API docs for the specific product line you contract

If a provider only offers a closed partner API, mark “integration uncertainty” as a first-class risk.

Dated candidate snapshot (capability inventory, not a ranking)

Candidate (as of 2026-07-15)	Official job shape	Wait / notify	Edit / extend / characters (per vendor docs)	Persistence notes from official docs	Gateway / ops note for this repo
OpenAI Sora 2 (`sora-2`, `sora-2-pro` and dated snapshots)	OpenAI Videos: create / retrieve / content	Poll ~10–20s; webhooks `video.completed` / `video.failed`; Batch for offline `POST /v1/videos`	Edits via `POST /v1/videos/edits`; extensions via `POST /v1/videos/extensions`; characters via `POST /v1/videos/characters`; remix deprecated for new work	Interactive downloads short-lived (~1 hour in OpenAI guide); Batch outputs have separate retention	Adapter exists for Sora-style create/retrieve/content; remix path still present; edits/extensions/characters not fully normalized. API shutdown 2026-09-24—exclude from greenfield defaults
Google Gemini video (Gemini Omni Flash; Veo 3.1 / related Veo ids in Google model lists)	Google AI video generation surfaces (long-running / conversational depending on model)	Provider-specific operations / polling; not identical to OpenAI Videos JSON	Google documents conversational editing for Omni Flash; Veo materials highlight extension/frame control style features—verify current method names	Follow Google retention and download guidance	Gemini/Vertex task adapters exist for Veo-style models in this codebase; confirm enabled aliases before promising
Runway API	Proprietary Runway job API (see Runway docs)	Vendor SDK / task status (confirm webhooks in current docs)	Mode matrix is vendor-defined (t2v/i2v/etc.)	Vendor URL expiry rules	No first-class `relay/channel/task/runway` package in this tree as of check date—treat as external or custom channel
Luma API	Proprietary Luma job API	Vendor status model	Vendor-defined	Vendor URL expiry rules	Same as Runway: confirm channel enablement; do not assume OpenAI Videos parity
Kling-style / other Asian video APIs	Often proprietary REST with task ids	Poll task status	Vendor-defined	Vendor CDN expiry common	Kling-style routes can appear when those adapters are enabled

Fill an extra column “enabled on our gateway?” before procurement. Empty cells mean “not verified in our lab,” not “feature absent forever.”

Dimension 1: API shape and async semantics

Pattern	What you implement	Watch-outs
OpenAI-style videos	create / retrieve / content (+ edits/extensions/characters upstream)	status enums, `expires_at`, remix deprecation
Proprietary job API	vendor SDK + optional webhooks	dual auth, dual logs
Sync-looking HTTP	single long request	LB timeouts, weak progress
Hybrid	create + webhook + poll fallback	need both paths tested

Questions for each vendor:

Are statuses documented and stable?
Is progress a percent, a stage string, or absent?
Are webhooks signed? What happens if webhooks fail—can you poll?
Is content a separate authorized GET?
For OpenAI: are you building against edits rather than deprecated remix?

This repository normalizes toward OpenAI-like task objects with statuses queued, in_progress, completed, and failed, plus progress, seconds, size, and expires_at. Gateway adapters may still talk Kling-style or other routes upstream.

Dimension 2: Duration, resolution, audio, and other multipliers

Video cost and risk scale with duration and resolution more aggressively than chat tokens. Compare from official docs:

allowed duration range and units (string seconds vs number)
default duration if omitted
max resolution / aspect ratios
whether native audio is included
seed / reproducibility controls (if documented)

In this codebase, MaxTaskDurationSeconds = 3600 is an upper validation bound for task duration inputs. That is a safety ceiling, not a product recommendation. Your comparison sheet should include “product max seconds” separate from “provider max seconds.”

Dimension 3: Modes of generation (use modern vocabulary)

Not every API is only text-to-video:

text-to-video
image-to-video / first-frame reference
video extension
targeted edit from prior output or uploaded video
character consistency assets
motion control tools
legacy remix (document only as compatibility/deprecated)

Map each mode to a product story before enabling it. Edit/extension features need read authorization on the source asset and clear billing semantics. Task objects here may still carry remixed_from_video_id; your app should persist lineage regardless of whether the upstream call was remix, edit, or extension.

For OpenAI specifically (docs 2026-07-15): prefer scorecard columns labeled edit / extension / characters, not remix-only language. Remix remains a historical/compatibility term.

Dimension 4: Latency classes and UX

Without publishing fake percentiles, classify from your tests:

Interactive draft (user waits in-page with progress)
Background render (notification on completion)
Batch (overnight queues; OpenAI Batch supports POST /v1/videos JSON lines)

Vendors that look similar on a single 4-second clip can diverge on 10+ second jobs or peak-hour queues. Capture p50/p95 yourself; put the date on the spreadsheet. Pair with channel health practices from LLM channel health monitoring.

Dimension 5: Content durability and delivery

Ask each provider—and your gateway—the same questions:

How long is download valid after completed?
Is there an official content endpoint vs a signed CDN URL?
Can content be re-fetched after the first download window?
Do Batch jobs have a different retention window than interactive jobs?
Do you need object storage for video workbench playback?

OpenAI’s guide states interactive download URLs are valid for a maximum of about one hour; Batch-generated videos have a separate documented window. Product architecture should assume your object keys are the durable record.

Dimension 6: Safety, moderation, and policy

Compare:

blocked content classes documented publicly
whether human likeness / character uploads need special enablement
copyrighted character/music rejection behavior (called out in OpenAI’s guide)
your own retention and takedown workflow

Also compare your obligations: can you delete a team’s videos from workbench storage? Are usage logs sufficient for abuse response? Multi-tenant denial tests belong in the bake-off (multi-tenant LLM gateway architecture).

Dimension 7: Pricing units and platform fit

Record cost units (per second, per job, resolution tier, credits), whether audio is bundled, edit/extension pricing vs new generation, and the pricing-page snapshot date. Never invent unit prices in marketing tables.

Prefer OpenAI-compatible /v1/videos plus stable aliases for multi-vendor clients; gateway keys and usage logs for team budgets; and video workbench personal plus /api/teams/:team_id/video-workbench when humans need team-scoped history. Before Sora shutdown, enable a second alias and persist bytes early. Keep native SDKs behind feature flags when a capability is non-portable.

Lab scorecard template

Last checked: 2026-07-15
Vendor / model id: ________________________________
Docs URL: ________________________________________
Pricing URL + snapshot: ____________________________

API shape: openai-videos / proprietary / hybrid
Statuses documented: Y/N   Webhooks: Y/N (signed? Y/N)
Content GET authZ: ______________________________________
Duration min/max (docs): ________________________________
Duration we will expose: ________________________________
Modes needed: t2v / i2v / edit / extension / characters / other
Remix only? (legacy) Y/N — migration plan: ______________
Our p50/p95 create→completed (region ____): _____________
Failure error quality (1-5 qualitative): ________________
URL/content expiry notes: ______________________________
Tenant isolation test passed? Y/N
Persistence plan: provider URL only / gateway proxy / our object storage
Cost unit (from official page): _________________________
Shutdown / deprecation date if any: _____________________
Support runbook owner: __________________________________
Go / no-go + rationale (no global rank):

Run the same three prompts and one image-to-video fixture (if required) on each finalist. Freeze seeds if the API supports them; if not, note non-reproducibility as a risk for QA.

Fair bake-off and exit cost

Freeze three prompts and one licensed image-to-video still; fix duration/size; run each case multiple times across peak and off-peak; record create errors, terminal failures, median completion, and download success. Attempt cross-tenant content download. If OpenAI is still in the mix before 2026-09-24, time a migration drill to a second alias. Reject cherry-picked frames and undated quality scores.

Exit questions: can you export prompts/task ids/MP4s; do edit/extension lineages lock you to vendor ids; can finance accept two pricing units during migration; did you calendarize the Sora/Videos shutdown? Gateways only help if product code uses aliases and you already persist bytes.

How to choose without a global ranking

Vendor lock-in OK: native SDK can be fastest if you still document expiry and moderation.
B2B team billing/audit: gateway + /v1/videos + object storage + team video workbench.
Multi-model lab: aliases and channel health first.
Regulated path: region, retention, and key custody over demo reels.
Greenfield mid-2026: do not default to Sora 2 long-term without a migration plan.

Revisit when model ids, duration tiers, pricing units, or deprecations change.

Sources

Closing checks

No universal winner: pick the API that clears your scorecard on the measurement date. Demo reels do not prove authZ, expiry, team billing, or deprecation readiness. OpenAI-compatible routes differ by gateway—confirm statuses, content auth, and whether edits/extensions/characters are normalized. Re-benchmark when defaults or deprecations change. Mixing image and video vendors is fine when logging and keys stay unified.

Selection drill: enable one durable non-Sora-only alias in staging, run create→poll/webhook→content→object-storage, and prove a second tenant cannot download the task.