Image Generation API from China: GPT Image 2.5, Gemini, Grok

Availability note — 2026-09-25: The catalog lists seven image models (gpt-image-2, GPT Image 2.5 Flare and Sunburst, two Gemini image models and two Grok Imagine entries) and no video models; Doubao Seedream is not listed. The video section below is kept as a historical example only — check the live catalog before integration.

Text-to-image and video generation models are now good enough to sit inside real products — marketing pipelines, e-commerce listings, game assets, short-form video. The catch for developers in China is familiar: the strongest models live behind billing pages that want a foreign card and endpoints that are unreliable or unreachable from Mainland networks.

This guide covers the seven image models in the catalog as of 2026-09-25 — gpt-image-2, GPT Image 2.5 Flare and Sunburst, Gemini 3 Pro Image and Gemini 3.1 Flash Image (the line the community searches as "Nano Banana"), and the two Grok Imagine entries — through one gateway key, paid in RMB. No video model is listed today; see video generation availability before planning a clip.

The Access Problem, Briefly

Every major image or video model has its own gate:

• gpt-image-2 hangs off an OpenAI platform account, which needs a card most Mainland banks won't issue.
• Gemini 3 Pro Image and Gemini 3.1 Flash Image sit behind Google endpoints that are blocked or unstable from China, plus Google Cloud billing.
• Grok Imagine requires an xAI account and billing relationship most China-based teams don't have.
• The video models each come with their own console, account and top-up flow — running four of them means four integrations and four balances to watch.

Router One provides a China-reachable gateway for the models in its current catalog, with one key, a wallet you can top up by card or Alipay, and per-request cost traces. The catalog supports image generation but lists no video models. The dedicated landing pages cover the surface area — image generation API and video generation availability — this post is the hands-on walkthrough.

What's in the Catalog (September 2026)

Image models are unit-priced: a flat USD amount per generated image, no token math. The seven image ids, checked against the live catalog on 2026-09-25 — live prices are always on router.one/models:

The August 2026 snapshot's Seedream and video rows now sit under Historical Video Examples below; none of those models is listed today.

Setup: One Key for Everything

Create an API key in the Router One dashboard and top up — card and Alipay share one hosted checkout, and USDT/USDC works on six chains; no foreign credit card required. The same key you may already use for chat models drives the media endpoints.

export ROUTER_ONE_KEY=sk-your-router-one-key

Generating Images

The image endpoint follows the OpenAI Images API request shape — POST a model ID and a prompt:

curl -X POST https://api.router.one/v1/images/generations \
  -H "Authorization: Bearer $ROUTER_ONE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2",
       "prompt": "Flat-lay product photo of a ceramic teapot, soft morning light"}'

Because the shape is OpenAI-compatible, existing OpenAI SDK image code works after changing the base URL and key. Model choice notes from real usage:

• gpt-image-2 — strongest prompt adherence and in-image text rendering; the default for product shots and UI mockups.
• GPT Image 2.5 Flare (gpt-image-2.5-flare) — the fast tier of the GPT Image 2.5 line, first seen 2026-09-11 (catalog release date 2026-09-08); the same request shape as gpt-image-2, at the flat per-image price on the GPT Image 2.5 Flare model page.
• GPT Image 2.5 Sunburst (gpt-image-2.5-sunburst) — the high-fidelity tier of the same line, listed the same day; same request shape, with its per-image price on the GPT Image 2.5 Sunburst model page.
• Gemini 3 Pro Image ("Nano Banana") — multimodal prompting and photorealism; strong at restyling reference images.
• Gemini 3.1 Flash Image — the Flash entry on the same Gemini line, added to the catalog in August 2026; same prompting, same flat price, worth A/B-ing against the Pro entry on your own prompts before you standardise — the Nano Banana 2 API guide has the full request shapes.
• Grok Imagine — two catalog entries, grok-imagine-image and grok-imagine-image-quality. Both cost the same per image, so the choice is about output, not budget.

Full request and response fields are in the Images API docs.

Historical Video Examples

No video models are listed as of September 5. The examples below document the earlier endpoint shape and require an available model before they can run.

The August 2026 snapshot listed these media models. None of them is in the catalog as of 2026-09-25; the rows are kept for reference only:

Video generation is asynchronous: you submit a task, get a task ID back, and poll until the clip is ready. Don't wrap the submit call in a tight client timeout — poll instead.

# 1. Submit the task
curl -X POST https://api.router.one/v1/videos/generations \
  -H "Authorization: Bearer $ROUTER_ONE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "viduq3-turbo",
       "prompt": "Drone shot over terraced rice fields at dawn, mist in the valleys"}'

# 2. Poll the task until it completes
curl https://api.router.one/v1/videos/generations/{task_id} \
  -H "Authorization: Bearer $ROUTER_ONE_KEY"

wan2.6-i2v is the image-to-video model: it animates a still you pass in, so it needs an image_url and rejects a prompt-only request with a 400.

curl -X POST https://api.router.one/v1/videos/generations \
  -H "Authorization: Bearer $ROUTER_ONE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "wan2.6-i2v",
       "prompt": "Camera pushes in slowly, steam rising from the cup",
       "image_url": "https://example.com/still.jpg"}'

Generation time varies by model and load — keep the task ID and check status rather than holding a connection open. Duration and resolution are locked per model (see the table above), so a duration or size field in the body changes nothing about the clip you get; the other five models are text-to-video and reject an image_url outright. Endpoint details are in the Videos API docs.

Keeping Media Spend Under Control

Media generation is where API bills surprise people — one enthusiastic batch job is suddenly hundreds of images. The gateway's standard controls apply to media models:

• Per-request traces. Every generation lands in Dashboard → Logs with model, cost, and status. Finance can see exactly what a campaign's asset pipeline cost.
• Per-key budgets. Give the image pipeline its own key with a spend cap, separate from chat traffic. A runaway job stops at the cap instead of draining the wallet.
• Unit pricing. Because each image is a flat price, cost forecasting is multiplication, not token estimation.

FAQ

Is Sora available? No. As of 2026-09-05, the catalog contains no video models: Sora, Vidu, PixVerse, MiniMax Hailuo, Wan, Veo, and Grok Video are unavailable. The examples above describe the August snapshot. Check the live catalog before trying a video request.

What about Midjourney? Midjourney and Doubao Seedream are absent as of 2026-09-05. The current image lineup contains GPT Image 2, two Gemini image models, and two Grok Imagine models, plus — first seen 2026-09-11 (catalog release date 2026-09-08) — GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. Copy exact IDs from the live catalog.

Is billing token-based? No — media models are unit-priced per image or per video. Chat models on the same key remain token-metered. Both appear in the same cost trace and usage dashboard.

Can one key serve chat, image, and video? One key can authenticate supported chat and image requests. The presence of /v1/videos/generations does not establish that a video model is available; as of 2026-09-05 none are listed. Match each exact model ID to its supported endpoint before use.

How do I pay in RMB? Top up the wallet by card or Alipay on one hosted checkout — amounts convert at the checkout rate. Stablecoins (USDT/USDC on six chains) also work. See the payment walkthrough.

Conclusion

The historical examples above show how the media APIs are shaped. As of 2026-09-05, use the current image catalog for GPT Image 2, the GPT Image 2.5 Flare and Sunburst tiers first seen 2026-09-11 (catalog release date 2026-09-08), Gemini image models, and Grok Imagine; Seedream and video models are unavailable. Verify the model ID and price before starting a batch.

Start with the image generation API page for the endpoint surface, check video generation availability before planning a clip, and use router.one/models for live pricing.