10 Best Image-to-Video APIs in 2026

Most image to video API roundups still rank models on leaderboards that have already moved. Two of the three highest-rated models today appear in almost none of the lists published this summer, and one widely recommended option has a removal date printed on OpenAI's own deprecation page: September 24, 2026.

So this article does the thing those lists leave out. It states the date the data was taken, and it shows you how to re-check it against the API yourself when the board changes again.

Last updated: August 2026

How an image-to-video API works

You send a still image and a prompt, and the model returns a video. The image decides what the scene looks like; the prompt decides what happens in it.

That's the difference from text-to-video, where the model has to invent the subject as well as the motion. Image-to-video keeps the product, the character, and the framing you already have — which is why it fits so many real use cases, especially when the image already exists.

Two things are worth knowing before you compare models:

• Generation is asynchronous. You submit a job, receive a taskId, and poll until it finishes, rather than getting the file back in the response.
• Most current leaders generate native audio. That changes both the output and the price.

How the models are ranked

The ranking below follows the Artificial Analysis image-to-video leaderboard, read on August 13, 2026, then cross-checked against published capability and pricing data.

Elo comes from a blind vote: people see two videos generated from the same image and pick the one they prefer, with 2,350 to 12,568 comparisons per model across 29 models. It tells you which model people prefer — not whether it holds your brand colours, respects a 9:16 crop, or fits your budget. Use it to narrow the list, not to decide.

This board changed repeatedly through 2026 and will change again. When you need to refresh it, repeat three steps:

1. Open the leaderboard and record the date.
2. Check the top five against each provider's deprecation page and drop anything scheduled for removal.
3. Test with your own images rather than trusting the benchmark, because the top-ranked model isn't necessarily the best one for your specific use case.

The ranking, August 2026

Not every model is available on every platform, and this list can change at any time. Rather than trusting a roundup table, check the model list endpoint for what's actually supported.

Beyond the top ten, AVIS also carries other strong image-to-video models including Seedance 2.0 Fast, Seedance 2.0 Mini, Seedance 1.5 Pro, Seedance 1.0 Pro Fast, Seedance 1.0 Lite, and Wan 2.7. Each has its own advantage in speed, quality, or cost, so they're worth considering depending on the use case. Full models, specs, and rates are on the AVIS pricing page.

What to check before you integrate

Check the model's lifecycle first

Don't choose on rank or output quality alone. Check the model's lifecycle before you start integrating:

• Is this the latest version?
• Does the provider have a newer model that replaces it?
• Is the current model already in a deprecation window?
• If it's removed, is there a compatible replacement?

The best model today isn't necessarily the best one to build on for the long term.

A lifecycle check takes a few minutes and can save you from migrating an entire integration later. OpenAI's deprecations page is the clearest example of why: it states that on March 24, 2026 developers were notified that the Videos API and the sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, and sora-2-pro-2025-10-06 aliases would be removed on September 24, 2026, with no replacement model listed.

Read specs from the API, not from roundups

Roundups rarely get updated when a model ships a new version, so their numbers go stale fast. Worse, different sources publish completely different figures for the same model:

These aren't small differences. 4K carries four times the pixels of 1080p, and a 20-second ceiling instead of 15 changes both your cost and how you design the workflow.

So when you compare models, don't copy specs from other articles. Read them from the API where you can.

Broadly, the current generation runs about 3 to 20 seconds per clip, 480p to 4K, with 16:9 and 9:16 as the common ratios. That's enough for an overview, but not enough to start an integration.

What to verify directly from the API:

• Whether the model accepts image input and returns video output.
• Minimum and maximum clip duration.
• Supported resolutions.
• Supported aspect ratios.
• Whether the model generates audio.
• Whether the model is still active or has been retired.

On AVIS, you can check all of it through:

• The model list endpoint — GET /api/v1/ai/models returns the model catalog.
• Modality filters — use inputModalities and outputModalities to isolate the image-to-video set.
• capabilities.video.params — the per-model limits, including duration, resolution, and aspect ratio.
• isActive — whether the model is still live and accepting requests.

The point is that you're reading data from the API instead of a blog post written months ago. When a model changes, you re-query and get the current answer.

Control cost per second of output

Video bills per second of output, which makes it one of the easiest modalities to overspend on. With rates across this list running from $0.05 to $0.50 per second, the same clip can cost ten times more depending on the model. Higher resolution, audio, and re-running generations to tune a prompt all push it further.

Without controls, a single misconfiguration — the wrong, more expensive model, or a loop left running — can burn a month's AI budget in a few days.

Four habits keep it in check:

• Prototype at 480p. Only run the final at the resolution you actually need.
• Set a spending limit on the API key. AVIS supports daily, weekly, monthly, and total limits per key, which caps the damage when code misbehaves or requests repeat unintentionally.
• Track what you actually spend. The usage endpoint returns usdCost and durationMs per generation, so you know what each request really cost rather than what the rate card implies.
• Verify the model before production. Don't let a staging environment quietly run a more expensive model or resolution than you intended.

With video generation, cost control isn't just picking a cheap model. It's capping the spend, controlling the inputs, and knowing exactly what each request costs.

Frequently asked questions

What's the best image-to-video API right now?

On the Artificial Analysis leaderboard read August 13, 2026, Seedance 2.0 leads at 1,198 Elo, with MiniMax H3 and Gemini Omni Flash within six points. Rankings move quickly, though, so always check the date on any board before relying on it.

How much does one clip cost?

Rates currently run about $0.05 to $0.50 per second, depending on model, resolution, and audio. A 5-second clip can cost anywhere from a few dozen cents to several dollars. If you're still experimenting, start at 480p and check your real usdCost rather than the rate card.

Why does the API return a taskId instead of the video?

Video generation takes longer than an HTTP request should stay open. The API returns a taskId so you can track the job, then you poll until the video is ready or the job fails. This is how video generation APIs generally work — it isn't specific to one provider.

Should I worry about a model being discontinued?

Yes, and it's the check most comparisons skip. Before you start integrating, check the model's lifecycle: whether a newer version exists, whether the current one is in a deprecation window, and whether a removal date has been published.

Keep the model name in configuration rather than hard-coded. When you need to move to a new model, that's a config change instead of a rewrite of the integration.

Bottom line

On current data, Seedance 2.0 leads the Artificial Analysis image-to-video leaderboard as of August 2026. But MiniMax H3 and Gemini Omni Flash are close enough that the gap on the board isn't sufficient to tell you which model suits your product.

Rather than hunting for one absolute "best" model, start from a shortlist and test it with the actual images and use cases you'll ship.

More importantly, keep three habits:

1. Record the date of every leaderboard you rely on.
2. Check lifecycle and deprecation before building the integration.
3. Read specs and costs from the API, not from roundups.

Finally, keep the model name in config so moving to a new model is a one-line change. AVIS Gateway is built for that: you reach many video models through the same API key and the same integration, instead of rebuilding every time you switch.