A new API gateway called Velokey launched on Hacker News this week with a simple pitch: one endpoint, one bill, and access to nearly every frontier generative model at once. That means text-to-video engines like Seedance 2.0, Kling V3, Veo 3.1, and Sora 2 sit behind the same key as image models (GPT Image 2, Seedream 5.0, Grok Imagine) and the usual large language models from OpenAI, Anthropic, Google, and DeepSeek. The differentiator is an OpenAI-compatible request format, so teams already calling the OpenAI SDK can point it at Velokey and start generating video without rewriting their client.
Velokey is not the first tool to promise this. The "one API for all the video models" space is already crowded, and that crowding is exactly why the launch is worth a closer look. For creators and developers building on top of generative media in 2026, the real question is no longer "which model" but "which gateway," because the gateway decides your billing, your failover behavior, and how much plumbing you own.
What Velokey Actually Is
Velokey is a paid, pay-as-you-go aggregation layer. You do not download anything and you do not host a model. Instead you send a request naming the model you want, and Velokey routes it to the underlying provider, meters the usage, and returns the result. The company advertises three capabilities beyond raw access: smart model routing with automatic failover, a single console for tracking usage and cost across every model, and a stated policy of not retaining prompts or outputs for training. New accounts get $0.50 in free credits to test with, per the Velokey API documentation.
The current catalog spans three media types. Video covers Seedance 2.0, Kling V3, Veo 3.1, Sora 2, Wan 2.7, PixVerse V6, and Vidu Q3. Image covers GPT Image 2, Qwen Image 2.0, Seedream 5.0, and Grok Imagine. Text covers the GPT, Claude, Gemini, and DeepSeek families. Launch updates and support are posted through the project's official account on X. As with any single-developer Show HN launch, treat the roadmap as early: the value is in the pattern it represents more than any one feature.
The One-API Pattern, and Why Builders Reach for It
Every frontier video model ships its own API, its own auth, its own request schema, and its own billing portal. A studio that wants to A/B two models for a client shot ends up managing two vendor relationships for a single deliverable. The unified-API pattern collapses that. Here is the typical integration path, which Velokey follows closely:
- Point your existing client at the gateway. Because the endpoint is OpenAI-compatible, you change the base URL and the API key, not your request-building code.
- Name the model in the request. Swapping Kling V3 for Veo 3.1 is a string change, not a new SDK, a new credential, or a new billing account.
- Read one usage dashboard. Video is billed per second, images per image, and text per token, but all of it rolls up to a single invoice and a single spend view.
- Lean on failover. If one provider is rate-limited or down, routing can fall back to a comparable model instead of failing the whole job.
The payoff is speed of iteration. When a new model lands, you evaluate it by changing a parameter rather than onboarding a vendor. That is the same reason developers adopted unified LLM routers over the last two years, and the video and image space is now following the pattern that platforms like Replicate pioneered for hosted model inference.
Velokey vs the Unified-API Field
Velokey enters a category with several established players, each with a different bias. Some optimize for the widest model catalog, some for the lowest per-call price, and some for a hosting-and-deploy platform rather than a thin gateway. The table below frames where Velokey sits relative to the broader field. Independent roundups such as the Apiframe comparison of video generation APIs and the WaveSpeed 2026 guide are useful cross-references before you commit.
| Dimension | Velokey | Broad unified-API gateways | Hosted-model platforms |
|---|---|---|---|
| Primary interface | OpenAI-compatible endpoint | Single REST API, custom schema | Model-specific SDKs + hosted endpoints |
| Media coverage | Video, image, and text in one key | Usually video-first, some add image | Very broad, community + official models |
| Billing model | Pay-as-you-go, per-model units | Pay-as-you-go or flat monthly | Per-second compute or per-run |
| Failover routing | Automatic, built in | Varies by provider | Not the core promise |
| Best fit | Teams already on the OpenAI SDK | Video-heavy apps wanting breadth | Custom or fine-tuned model hosting |
The honest read: Velokey's edge is OpenAI compatibility plus text-image-video in one account, which lowers the switching cost for teams that already speak the OpenAI dialect. Its risk is the one every new aggregator carries, which is dependency on upstream providers it does not control.
What a Unified Video API Enables
For a creator or a small studio, the practical unlock is cheap comparison. Suppose you need a six-second product shot. With a unified key you can render the same prompt through Google's Veo and through Kling, put both in front of the client, and only pay per second of what you generate, all under one invoice. There is no second signup, no second card on file, and no reconciling two dashboards at month end.
For a developer shipping a product, the unlock is resilience and future-proofing. A single-provider integration breaks when that provider has an outage or deprecates a model. A gateway with failover degrades gracefully, and adopting next month's model is a config change instead of a sprint. That is how you keep a generative-media feature current without a rewrite every quarter.
Pricing, Free Credits, and the Trade-Offs
Velokey prices each media type in its native unit. Published examples include Claude Sonnet 4.6 at $3.00 per million tokens, GPT Image 2 at $0.006 per image, and Seedance 2.0 at $0.29 per second of video. New users get $0.50 in free credits, which is enough to run a short clip or a handful of images before committing a card.
The trade-off of any aggregator is margin and trust. A gateway typically adds a small markup over calling each provider directly, and you are trusting a middle layer with your prompts and outputs. Velokey's stated no-training-retention policy addresses part of that concern, but for regulated or sensitive workloads you should verify the data-handling terms in the docs before you route production traffic through it.
Frequently Asked Questions
Is Velokey open source or self-hostable?
No. Velokey is a hosted, paid API service, not an open-weight model or a self-hosted router. You call its endpoint and it meters usage. There is no model to download.
Do I have to rewrite my code to use it?
Not if you already use the OpenAI SDK. Velokey exposes an OpenAI-compatible endpoint, so in most cases you change the base URL and API key and keep your existing request-building code.
Which video models does Velokey support?
At launch the video catalog includes Seedance 2.0, Kling V3, Veo 3.1, Sora 2, Wan 2.7, PixVerse V6, and Vidu Q3, alongside image and text models in the same account.
How is video billed?
Video is billed per second of output, images per image, and text per token. Velokey lists Seedance 2.0 at $0.29 per second as an example, and rolls all media types into one pay-as-you-go invoice.
How is this different from calling each model's API directly?
Direct integration means separate keys, schemas, and bills per provider. A unified gateway gives you one key, one dashboard, one invoice, and automatic failover, at the cost of a small markup and trusting a middle layer.
Is it safe to use for client work?
Velokey states it does not retain prompts or outputs for training, which is a good sign, but for client or regulated work you should read the current data-handling and uptime terms in the documentation before routing production traffic.
What to Do Next
If you build with generative video or image models, the fastest way to judge Velokey is to spend the $0.50 in free credits: render one prompt through two different video models and compare cost, latency, and output quality side by side. If you are shopping the whole category rather than one tool, read at least one independent roundup, decide whether you value breadth or the lowest per-call price, and confirm the failover behavior matches what your product needs before you migrate any traffic.
