Free AI Model, Nano Banana 2, Seedream v5 and Qwen Image Max

Today we’re shipping a big update to our image generation stack: a new free model tier, several new pro models, and a more consistent credit system across the whole product.

A new Free AI Model tier#

We’ve added a Free AI Model that you can use for both text-to-image and image-to-image:

• Text-to-image: powered by flux-2-dev, exposed as “Free AI Model” in the model selector.
• Image-to-image: powered by z-image-turbo, available as “Free AI Model (10 times/day)”.

Both models cost 0 credits. For logged-in users, we apply a 10 calls/day server-side limit on the free image-to-image path to keep the system healthy. The free text-to-image endpoint is also integrated into the same pricing logic, but with a cost of 0 credits.

This tier is ideal for:

• Quick experiments and exploration
• Drafting compositions and styles
• Social posts, thumbnails, and other “good enough” creative tasks

Whenever you need production-grade quality, you can switch to one of the paid models without changing your workflow.

Nano Banana 2 is live#

Nano Banana 2 is now available as a first-class model for both text-to-image and image-to-image.

Compared with the original Nano Banana, Nano Banana 2 focuses on:

• More stable composition, especially with complex prompts
• Better detail in fur, hair, and small objects
• More consistent behavior across similar prompts

You’ll find:

• A dedicated /models/nano-banana-2 page
• A side-by-side Nano Banana vs Nano Banana 2 comparison on the Nano Banana model page
• A refreshed announcement bar on the homepage pointing directly to Nano Banana 2

If you’ve already built prompts around Nano Banana but want a bit more sharpness and stability, try running them as-is on Nano Banana 2 and compare.

Seedream v5 and Qwen Image Max#

We’re also upgrading two of our heavier hitters:

• Seedream v5

  • New text-to-image and image-to-image endpoints from Bytedance
  • Available in the model lists as a v5 variant, with its own landing page
  • A good choice for high-detail, design-oriented work

• Qwen Image Max

  • Upgraded from the previous Qwen Image to a new qwen-image-max endpoint
  • Dedicated landing page with usage examples and a detailed FAQ
  • Especially strong for controlled image-to-image editing, style transfer, and batch editing

On the Qwen Image Max model page, we outline common use cases:

• Art style transfer (anime, watercolor, oil painting, sketch…)
• Batch editing product images while keeping consistent styling
• Enhancing marketing visuals and social content
• Iterating quickly through creative design directions

If you work with lots of similar images (catalogs, campaigns, etc.), Qwen Image Max is now one of the best options in the stack.

Smarter, model-aware credits#

Under the hood, we’ve unified how credits are calculated and displayed:

• Every model has an explicit MODEL_CREDIT_COST on both the frontend and backend.
• The UI shows “Free / 1 Credit / N Credits” directly on the main generate button.
• The app only blocks on balance when creditCost > 0, which means free models behave as you’d expect even with a zero balance.
• GPT-Image-1.5 keeps its quality-based pricing (low/medium/high → 0/1/3 credits), now wired through the same credit cost helper.

This reduces surprises when switching between models and makes it obvious what you’re going to pay before you hit “Generate”.

Better routing and small UX fixes#

Finally, a few quality-of-life improvements:

• The routing middleware is now more robust:
  • It correctly distinguishes between localized and non-localized paths.
  • It handles paths with dots like /models/gpt-image-1.5 without breaking.
• Model dropdowns have better width handling across devices.
• The ImageCombiner automatically switches model lists when you move between text-to-image and image-to-image, and keeps the selected model valid.

We’re not done yet. This release lays the groundwork for more fine-grained plans and limits on top of the new free tier. If you have ideas for what you’d like to see next (more free capacity, new models, or better editing tools), hit us up.