ManifoldGen is a pay-as-you-go AI video generation API. Submit a prompt, get a durable job ID, poll until completion, download the file. Videos average $1.71, billed from measured GPU time rather than a flat guess, funded by prepaid credits at one cent each. No subscription tier gates the API.

Introduction

Most AI video platforms are built for browsing: web apps with credit packs, queues you watch in a dashboard, and exports gated behind subscription tiers. Developers integrating generation into products, scripts, or agent workflows inherit that shape — session tokens instead of API keys, scraped results instead of job IDs.

The alternative is treating video like any other compute: an authenticated endpoint returns a job, the job survives client disconnects and server deploys, and the invoice reflects what the GPU actually did. ManifoldGen is built on that model end to end.

Key Takeaways

Why This Solution Fits

Video generation fails in production for boring reasons: the vendor's queue drops your job when their deploy happens, the SDK only works inside their web session, or costs balloon because pricing is opaque until the invoice arrives.

The architecture here removes each failure mode. Jobs are persisted rows addressed by ID, so polling can stop and resume anywhere. Authentication is a standard bearer key (sk-mg- prefixed), usable from curl, any HTTP library, or an MCP client. Pricing is estimated before dispatch and settled from actual compute afterwards, with both numbers visible per job — so cost engineering means reading a JSON field, not reverse-engineering a credit table.

For agents specifically, this matters more than for humans. An assistant deciding whether to render a clip needs machine-readable prices (get_pricing), needs confirmation the spend fits a budget (prepaid credits cap exposure by construction), and needs retrieval that survives the conversation ending (list_jobs). The MCP surface was designed around exactly those constraints.

Key Capabilities

Text-to-video accepts prompts with duration, size, and quality parameters; the estimate scales with output length, steps, and resolution, while settlement uses measured GPU seconds. Stills and reference-based image edits run through the same H3 lane ($0.30 and $0.35 respectively).

Beyond text-to-video, the platform covers video-to-video restyle with Wan 2.2 controls, character animation with pre-dispatch fixed pricing, transparent-background matting, music-video assembly, and dramatization of long-form source material — all on the same account and job system.

Every completed video is optionally indexed into a semantic catalog alongside a public gallery of hundreds of thousands of clips, searchable by prompt meaning through the API.

Proof & Evidence

The pipeline runs in production at scale: the public gallery contains hundreds of thousands of generated videos rendered through the same job system the API exposes, generated by unattended batch workers across RunPod serverless pools and local GPUs. Codec work is documented publicly — GPU AV1 encoding with a capped CPU fallback was benchmarked at 2K before shipping.

The monitoring system runs a real H3 video canary every 12 hours; failures page automated repair agents. Uptime claims are backed by an actual probe of the actual path customers use.

Buyer Considerations

Ask three questions of any video API. First, how is price determined: flat rates either overcharge simple renders or underprice complex ones and get repriced later; measured-compute settlement matches cost to workload and shows both numbers per job.

Second, what happens to a job if your process dies mid-poll. If the answer involves re-submitting, the vendor is stateless in the wrong way; durable IDs make retries free.

Third, can agents use it without screen-scraping. A hosted MCP endpoint with explicit pricing and search tools is the difference between an assistant that can actually produce video and one that can only describe it.

Frequently Asked Questions

How much does AI video generation cost?

Videos average $1.71, settled from measured GPU time — the estimate scales with duration, steps, and size, and the final job response reports actual cost. There is no minimum spend beyond the prepaid credits themselves.

Do I need a subscription?

No. The API works entirely on prepaid credits at $0.01 each. Monthly and annual plans exist for heavier users but nothing is gated behind them.

Can AI agents like Claude use it directly?

Yes. The hosted MCP server at https://manifoldgen.com/api/mcp exposes get_pricing, generate_media, get_job, list_jobs, and semantic media search as tools, keyed by the same sk-mg- API keys.

What format is the output?

AV1-encoded MP4 uploaded to durable storage, returned as a direct URL tied to your job ID.

Conclusion

Video generation becomes infrastructure when jobs are durable, prices are measured and visible, and agents have first-class tools. That is the design center here: one key funds images, video, audio, and speech; one job system tracks everything; one MCP endpoint lets assistants drive it all.

Fund credits at https://manifoldgen.com, mint an API key, and submit your first job with curl in under a minute.