Image & Video Generation API Pricing Compared (2026)
Media models are billed per result, so unlike text models the comparison fits in one table. On byesu, gpt-image-2 is $0.01 per image, nano-banana-2 $0.025, seedream-4.5 $0.05, and video runs from $0.015–$0.025 per second — with veo-3.1 at $0.01875/s, a fraction of Google's $0.40/s list price. Failed video tasks are automatically refunded, and one sk- token covers every model below. Official list prices were checked in July 2026 from each vendor's pricing page; vendors change prices, so always confirm on their pages before budgeting.
Image Model Pricing
| Model | byesu price | Official list price | Source |
|---|---|---|---|
gpt-image-2 | $0.01 / image | Token-based: $5/M text in · $8/M image in · $30/M image out (≈$0.006–$0.211 per 1024×1024, see below) | OpenAI pricing |
nano-banana-2 (Gemini 3.1 Flash Image) | $0.025 / image | $0.067 (1K) · $0.101 (2K) · $0.151 (4K) per image | Gemini API pricing |
seedream-4.5 (ByteDance) | $0.05 / image, flat incl. 2K & 4K | $0.04 per image | fal.ai model page |
firefly-gpt-image-2 | $0.04 / image, 1K–4K, ≤6 reference images | Credit-based via Adobe Firefly plans; no public per-image USD list price | Adobe Firefly plans |
A few honest notes. On seedream-4.5, byesu's flat price sits slightly above the vendor's per-image rate — what you get for it is one price across 2K and 4K output plus the same endpoint and token as every other model here. On nano-banana-2, Google's own per-image rates rise steeply with resolution, while byesu's per-image price stays flat. Live per-resolution pricing for every model is shown in the Generation Playground.
gpt-image-2: Turning Token Pricing Into a Per-Image Cost
OpenAI does not sell gpt-image-2 by the image — it bills tokens: $5 per million text input tokens, $8 per million image input tokens ($2/M when cached), and $30 per million image output tokens. A generated image consumes output tokens in proportion to its resolution and quality setting, so the real cost moves per request. Published OpenAI examples for a 1024×1024 image work out to roughly:
| Quality | Approx. cost per image (OpenAI, direct) |
|---|---|
| Low | ≈ $0.006 |
| Medium | ≈ $0.053 |
| High | ≈ $0.211 |
Edit and reference-image requests pay image input tokens on top, and long prompts add text input tokens — so a "per-image price" from OpenAI is always an estimate, not a list price. (The Batch API halves token rates in exchange for asynchronous processing.)
byesu removes the math: $0.01 per image, flat, whether it's a first draft or the tenth revision with reference images attached via /v1/images/edits.
Video Model Pricing (Per Second)
Video price = per-second unit price × the seconds parameter. Available durations vary by model, and failed tasks are automatically refunded.
| Model | byesu per second | Example totals (byesu) | Official list per second | Source |
|---|---|---|---|---|
gemini-veo31 (Veo 3.1, 1080p) | $0.01875 /s | 6s = $0.1125 · 8s = $0.15 | $0.40 standard (720p/1080p) · $0.10–$0.12 Fast | Gemini API pricing |
runway-gen4-turbo (image-to-video) | $0.025 /s | 5s = $0.125 · 10s = $0.25 | $0.05 (5 credits/s × $0.01/credit) | Runway API docs |
firefly-video (1080p) | $0.015 /s | 5s = $0.075 | Credit-based via Adobe Firefly plans; no public per-second USD list price | Adobe Firefly plans |
For scale: an 8-second Veo 3.1 clip is $0.15 on byesu versus $3.20 at Google's standard list rate — the kind of gap that decides whether iterating on prompts ten times per shot is viable.
Which Model Should You Pick?
| Model | Best for | Typical look | Price position |
|---|---|---|---|
gpt-image-2 | Everyday generations, icons, rapid iteration | Clean, instruction-faithful | Lowest per image ($0.01) |
nano-banana-2 | Create and edit, text rendering, up to 4K | Versatile, strong prompt adherence | Mid ($0.025) |
seedream-4.5 | High-res posters, art direction | Detailed, stylized, 2K/4K native | Premium ($0.05 flat incl. 4K) |
firefly-gpt-image-2 | Multi-reference workflows (≤6 refs), 1K–4K | GPT-style output, Firefly pipeline | Mid ($0.04) |
gemini-veo31 | Text- or image-to-video, cinematic 1080p | High visual quality, natural motion | $0.01875/s |
runway-gen4-turbo | Animating a still (first frame required) | Controlled, production-style motion | $0.025/s |
firefly-video | Short 5s 1080p clips | Adobe pipeline consistency | $0.015/s |
Rule of thumb: iterate on gpt-image-2, finalize on nano-banana-2 or seedream-4.5, and reserve video seconds for prompts an image draft has already validated.
How to Call These Models
Every model above answers on standard OpenAI-style endpoints — POST /v1/images/generations for images, async POST /v1/videos for video — with one token from the media / media-gen group; see Calling via API plus the model lists on Image Generation and Video Generation. Billing is pay-as-you-go with no subscription — top up with USDT, Alipay or WeChat Pay.
FAQ
How much does the gpt-image-2 API cost per image?
On byesu, gpt-image-2 is a flat $0.01 per image. Calling OpenAI directly, the model is billed by token — $5/M text input, $8/M image input, $30/M image output — which works out to roughly $0.006–$0.211 per 1024×1024 image depending on the quality setting, based on published OpenAI examples.
What does the Veo 3.1 API cost per second?
Google lists Veo 3.1 at $0.40/s for 720p/1080p standard output and $0.10–$0.12/s for the Fast tier. On byesu, gemini-veo31 is $0.01875/s, so a 6-second clip costs about $0.11 and an 8-second clip $0.15. Failed tasks are automatically refunded.
How does token-based image pricing convert to a per-image cost?
A generated image consumes image output tokens that scale with resolution and quality, so the bill varies per request. Published OpenAI examples put a 1024×1024 gpt-image-2 image at about $0.006 (low), $0.053 (medium) and $0.211 (high quality), and edit requests add image input tokens for every reference image. Per-result billing removes that math: one image equals one fixed price.
What happens if a video generation task fails?
On byesu, failed video tasks are automatically refunded. The video price equals the per-second unit price × the seconds parameter, charged when the task is created and returned in full if the task ends in a failed state.
Which image generation API should I choose?
Pick gpt-image-2 for fast, low-cost everyday iteration; nano-banana-2 for strong creation plus editing up to 4K; seedream-4.5 for high-resolution 2K/4K art direction; and firefly-gpt-image-2 when a workflow needs up to 6 reference images. All of them sit behind one byesu token, so switching models is a one-string change.