← Back to blog
GeneralAbout 6 min read

Sixty-Six Image Models, Thirty-Three Video Models. Read It Like a Menu.

Published Nov 7, 2026
Sixty-Six Image Models, Thirty-Three Video Models. Read It Like a Menu.

OpenRouter's catalogue currently lists 66 models that output images and 33 that output video. If you have ever tried to pick one by scrolling a leaderboard, you know the feeling: the ranking tells you what other people ran, not what you should run.

So I pulled the whole shelf through the public models API and read it as a catalogue instead. Prices below are the image-token rate converted to dollars per million tokens, or the published per-second video rate where one exists. Nothing here is a benchmark score. It is the shape of the market.

The shelf has eight families, not sixty-six products

Count the entries and you get 66. Count the lineages and you get roughly eight: OpenAI's GPT Image line, Google's Gemini/Nano Banana line, ByteDance's Seedream, Qwen's Image series, Black Forest Labs' FLUX, Recraft, Krea, and a long tail of one-offs from Sourceful, Microsoft, Meta, Tencent and xAI.

Recraft alone accounts for about sixteen of the 66, because it sells style consistency as a product: per-request style IDs, utility variants, vector output. That is not padding, it is a different business model — you are buying a consistent look, not a model. Same story with FLUX, which appears as FLUX 3 Image, FLUX 3 Image Klein, and four FLUX.2 tiers, because open weights and API hosting are two different products sharing a name.

Practical consequence: when you compare "66 image models," you are mostly comparing eight vendors' pricing strategies. Treat the variants as tiers within a family and the decision gets much smaller.

A photographic contact sheet on a light table with one frame circled

Images: five price bands, and what each one buys

Budget band, under $3 per million image tokens. Tencent's Hy Image 3.5 Preview at $1.6, Recraft V4.1 Flash at $1.7, FLUX 3 Image Klein at $1.9, Meta's Muse Image at $2.4. This is where you do exploration and thumbnails. Klein is the interesting one because it is open-weights: the API is a convenience, not the product. Muse is notable for being agentic — it plans before it renders, which matters when your prompt is a paragraph rather than a phrase.

Workhorse band, $4–11. Qwen Image 2.1 Turbo at $4.4, Seedream 5.0 Flash at $4.3, Krea K3 Edit at $7.2, Seedream 5.0 Lite at $8.4, FLUX 3 Image at $9.8, Qwen Image 2.1 Pro at $11.0. Most production traffic should live here. K3 Edit is the one to note: it takes one to eight reference images and is built for targeted edits rather than fresh renders, which is what most real requests actually are.

Premium band, $30. GPT Image 2.6 Sunburst, GPT Image 2.6 Flare, Gemini Nano Banana 2.1. All three land on the same number, which is not a coincidence — it is a price umbrella. You are paying for text rendering, dense layouts, and consistency across edits. If your output has words in it that must be spelled correctly, this band is where that gets solved.

Specialist band, $38–120. Microsoft MAI-Image 2.6 at $38 is positioned for precision. Gemini 3 Pro Image at $120 is twenty times the budget band, for conversational editing with reasoning attached.

The vector exception. Recraft's V4.1 Vector at $19 and Pro Vector at $50 exist because SVG is a different deliverable, not a different quality tier. If you need something a designer can open and edit, the price comparison to raster models is meaningless.

Videos: 33 models, four jobs

Video pricing mostly isn't published in the catalogue, because it is billed per second by each provider. Where it is, the ladder is instructive: Gemini Omni 1.2 Flash runs $0.023 per second at 360p, $0.070 at 720p, $0.104 at 1080p, $0.209 at 4K. A fifteen-second 720p clip is about $1.05.

That ladder tells you how the category is structured. Group the 33 by the job they do, and the choice becomes legible.

General text-to-video and image-to-video. Gemini Omni 1.2 and 1.1 Flash, Seedance 2.5, Wan 3.0 and 3.0 Prime, Hailuo 3 and 3 Max, Runway Gen-4.5, Kling 3.0 Pro and Standard, Veo 3.1 and its fast/lite tiers, Grok Imagine Video 1.5. This is the crowded part of the market, and the differentiators are clip length, native audio, and reference handling rather than raw fidelity. Omni's pitch is a fifteen-second take with audio, first/last frame control, and multi-turn editing via a previous job ID.

Editing existing footage. Runway Aleph 2.0 takes a video and a text instruction and changes something inside the shot. FLUX Video Edit does the same. FLUX Video Upscale enlarges. This is the least glamorous and most economically useful category, because reshooting is what actually costs money.

Talking humans. HeyGen's Avatar IV and HeyGen Video, plus the general models used with a portrait. Corporate training, localisation, product explainers. Boring, enormous market.

Batch and draft tiers. Seedance 2.0 Mini and Fast, Veo 3.1 Fast and Lite, Grok Imagine Video 1.5 Lite, Wan 3.0 Prime. The existence of a "fast" tier in five families at once is the real signal: the industry has accepted that you generate a cheap draft, look at it, then pay for a final.

How to read the price table without getting fooled

Three traps, in order of how often they catch people.

Image tokens are not images. A "$30 per million image tokens" rate does not tell you the cost of a 1024×1024 picture until you know how many tokens that resolution consumes, and that varies by model. The honest comparison is a fixed test image at a fixed quality, priced end to end. Anyone quoting per-image costs without naming the resolution is guessing.

Draft prices are not production prices. A 360p video at $0.023 per second is a real offering and a terrible basis for budgeting a campaign. Budget at the resolution you will ship.

The auto routers are not models. openrouter/auto and its siblings accept text, image, file, audio and video and route for you. Useful when you don't care; useless when you are trying to attribute cost or reproduce a result.

What I would actually pick

If I were building an image pipeline this week: Seedream 5.0 Flash or Qwen Image 2.1 Turbo for exploration, Krea K3 Edit for revisions, GPT Image 2.6 Sunburst for anything with type in it, and FLUX 3 Klein when the client wants the weights. That is four models, three price bands, and it covers roughly everything a content team asks for.

For video: Omni 1.2 Flash as the default because fifteen seconds with audio covers most social formats in one call; Aleph 2.0 for fixing a shot instead of reshooting it; a fast tier for drafts. Then budget at 1080p and accept that the 360p number in your spreadsheet was fiction.

The wider point is that catalogue size is not choice. Sixty-six models with eight lineages and three price bands means the decision space is small, and most of the entries exist to serve procurement, not craft. Pick the band by what the output must survive — a thumbnail, a poster, a billboard — and the model name sorts itself out.

Related articles