Black Forest Labs’ FLUX 3 image model opened for API access on October 1. Generation and editing now live behind a single endpoint, the official Playground lets anyone try it directly, OpenRouter listed it the same day, and Krea announced it was integrating the model as well. Commercial weights are already available through BFL, while the open-weight release is still described only as coming “in the next few weeks.”
That’s just over two months after FLUX 3 first surfaced. When BFL unveiled FLUX 3 on July 23, it was pitched as a unified multimodal model that jointly learns image, video and audio generation within a single architecture, with robotic action prediction thrown in as well. Only the video capability shipped at the time; the image model was promised “within weeks.”
What it can do
Per BFL’s API documentation, the image model offers four output tiers — 768-square, 1K, 2K and 4K — defaulting to 1K. The 4K tier runs to roughly 16 megapixels, and the official demo image measures 5456×3072. Aspect ratios range across 15 presets from 21:9 to 9:21, plus an automatic option.
Key capabilities include:
- Up to 10 reference images, supplied as URLs or base64;
- Instruction-based editing, with text-to-image and image editing sharing the same endpoint;
- Layout control via integer coordinates on a 0-to-1000 grid, positioning each element by specifying top, left, bottom and right, then referencing it by name.
The documentation also lists several constraints. Reference images larger than 4 megapixels are downscaled to 4 megapixels before processing; 4K generations can take several minutes; and pixels outside an editing frame “generally” stay unchanged — language BFL stops short of guaranteeing, so pixel-level consistency isn’t promised. BFL has not disclosed parameter count, weight size or GPU requirements.
How the pricing works
On BFL’s pricing-page calculator, the default example lists a per-image price of $0.048, struck through in favor of a promotional $0.024. A third-party price breakdown puts the full schedule at $0.048 for 1K, $0.10 for 2K and $0.607 for 4K, with reference images and prompts included in the per-image price; everything is half off during the October 1-8 launch window. OpenRouter’s listed starting price is $0.0205 per image, consistent with that discount.
Rough math: for something like e-commerce product shots where 1K is good enough, 1,000 images during the promotion run about $24. Scale up to 4K print-grade assets and the same 1,000 images cost just over $300 at the promotional rate — more than $600 at list price.
Where it sits next to domestic options
As of launch day, FLUX 3’s image model hadn’t appeared on any public text-to-image or image-editing leaderboard, and BFL hasn’t published its own benchmark scores. On Artificial Analysis’s text-to-image leaderboard, the previous-generation FLUX.2 holds an Elo score of 1021, versus 1197 for the closed-source leader, GPT Image 2.5.
The closest point of comparison for developers in China is Alibaba’s Qwen-Image-2.1, open-sourced in late September. At just 7 billion parameters, it also supports editing with up to 10 images and tops Artificial Analysis’s open-weight text-to-image leaderboard with a score of 1034, as well as leading the open-weight editing leaderboard. Its license does carry separate terms for commercial use, though, which require a standalone agreement.
The trade-off is straightforward: FLUX 3 is currently only reachable through overseas endpoints like BFL or OpenRouter, billed per image in dollars, while Qwen-Image-2.1 can be self-hosted, trading away per-call fees for GPU costs and commercial-licensing paperwork. Whether FLUX 3’s open weights actually ship — and what license they carry — will decide whether it even makes the shortlist for self-hosted deployment in China.
Sources: Black Forest Labs API documentation and pricing page, OpenRouter model page, CocoLoop, Artificial Analysis leaderboards, Decrypt; per-resolution pricing and promotional discounts cross-checked against a third-party price breakdown.