Updates that matter when you build

Wan 2.6 Flash, the faster and lower-cost tier of Alibaba's Wan 2.6, now has a second endpoint on SpicyAPI: reference to video.

• Wan 2.6 Flash, reference to video (alibaba/wan-2.6-flash/reference-to-video): up to five references, up to three of them short clips and the rest still images, or stills alone. Name them in the prompt as character1, character2 and so on, and the model casts them into a new scene of 2 to 10 seconds at 720p or 1080p, in five aspect ratios.
• Sound is a switch. generate_audio is off by default; turn it on and the clip comes back with generated voices, effects and ambience at twice the per-second rate.
• Still images are free. Reference clips add their seconds to the bill, counted up to five seconds in total.

There is no content filter in front of the endpoint, and prompt expansion is off by default so your prompt reaches the model as written. None of our test requests was refused.

Billable seconds are the output length plus the reference clips, each rounded up to whole seconds, with the clips counted up to five seconds in total.

Qwen Image 2.1 Pro, the hosted Pro tier of Alibaba's Qwen Image 2.1, can now be called on SpicyAPI, and two models already in the catalog change today.

• Qwen Image 2.1 Pro, text to image (alibaba/qwen-image-2.1-pro/text-to-image): one to six images per request in fifteen aspect ratios at 1k, 1.5k or 2k. Our requests took 14 to 43 seconds.
• Qwen Image 2.1 Pro, image editing (alibaba/qwen-image-2.1-pro/edit): one to ten reference images and a plain-language instruction. Leave the aspect ratio out and the result keeps your source's shape at about 2K.
• Qwen Image 2.1 Turbo now costs $0.0144 an image on both endpoints, down from $0.0192.
• Wan 2.6 Flash image to video has a new generate_audio switch. Off, the default, the clip stays silent at the same per-second price as before; on, it comes back with a generated soundtrack at twice the rate.

There is no content filter in front of Qwen Image 2.1 Pro, and prompt expansion is off by default so your prompt reaches the model as written. Not one of our test requests was refused.

For both Qwen models every resolution tier costs the same, each image returned is billed, and the number of reference images does not change the price.

Qwen Image 2.1 Turbo, the 8-step checkpoint of Qwen Image 2.1 that Alibaba's Qwen team released with open weights on 9 October 2026, can now be called on SpicyAPI. The page that showed it as coming soon now runs it.

• Text to image (alibaba/qwen-image-2.1-turbo/text-to-image): one to six images per request in fifteen aspect ratios at 1k, 1.5k or 2k. Our requests took 13 to 39 seconds.
• Image editing (alibaba/qwen-image-2.1-turbo/edit): one to ten reference images and a plain-language instruction. Leave the aspect ratio out and the result keeps your source's shape at about 2K.

There is no content filter in front of either endpoint. Prompt expansion is off by default: with it on, Alibaba's rewrite tones adult prompts down, and in our tests an edit asking a clothed portrait to reveal more came back unchanged.

Every resolution tier costs the same, each image returned is billed, and the number of reference images does not change the price.

You can now continue a video you already have. Upload a clip, say what happens next, and SpicyAPI generates a new segment from the last frame of your clip and joins both into one MP4. It works on 31 video model families, in the API and on the site.

• How it works. Send your clip as video_url (MP4, MOV or WebM, up to 60 seconds) and a short prompt describing what happens next. On most models the prompt is optional. You get back the joined MP4 and, separately, the new segment on its own. You can extend the joined result again as many times as you like.
• Optional end frame. Some models accept last_image_url, a frame you want the new segment to end on. Seedance 2.5 reads up to the last 30 seconds of your clip as reference, so the new segment follows the motion more closely.
• Pricing. You pay only for the new segment, at that model's image-to-video price. Your original clip is not charged. Seedance 2.5 also charges for the seconds of reference it reads, at the same price as its reference-to-video. If joining fails, the whole job is refunded.
• A clean seam. Color, brightness and volume of the new segment are matched to your clip, and the frame rate follows your clip. No black bars are added. To line the seam up, the picture may be cropped or adjusted by a few percent at the edges.
• On the site. In Playground, Studio, Assets, the task history in the console and the short film editor, every video result has a one-click Extend button. Your clip is filled in, the same model is selected, and the price sits next to the button.
• 4K HDR clips. Phone footage in 4K HDR has a length limit: about 15 seconds at 60 fps and about 40 seconds at 30 fps. 1080p HDR is not affected.

FLUX 3 and Wan 2.2 Spicy LoRA already had their own native extend and keep it. The new endpoints are the video-extend model of each family below.

Eighteen LoRA-ready model families joined the catalogue today, 33 endpoints in all. Each one loads your own LoRA files by direct link at no extra charge (up to three per request on most, one on P-Image), and the platform adds no content filter.

Before you plug a LoRA in:

• Match the base. A LoRA only works on the base model it was trained for. One made for a different base usually loads without an error and changes nothing, so check its model card first. Every model page lists what we tested.
• Use a direct link to the weights file. A link to a repository page, or to a file that needs a login, fails with a clear error and is not charged.
• Start at scale 1. Strengths well above 1 quickly degrade the picture.

Open any of the new families under Models for tested results, example prompts and the full price table.

Vidu Q4 Preview, the public preview of Vidu's Q4 video model (released 7 October 2026), is now on SpicyAPI with two endpoints. Clips run 3 to 16 seconds at 540p up to 4k, with dialogue, effects and ambience generated alongside the picture; write a line in double quotes and it is spoken.

• Image to video (vidu/q4-preview/image-to-video): start from one opening frame; the clip keeps its shape. In our content tests it was never refused.
• Reference to video (vidu/q4-preview/reference-to-video): 1 to 15 reference images of people, outfits, products or places plus up to 3 voice clips become one scene. It refuses nudity and sex scenes, and refused requests are not charged.

2k and 4k come back as 10-bit HEVC files, which some browsers cannot play inline. Sound on or off and the number of references do not change the price.

FLUX 3, the video model from Black Forest Labs, is now on SpicyAPI with four endpoints: text to video, image to video (with an optional last frame), video extend and video upscale.

Text to video, image to video and extend are at Black Forest Labs' list price.

Clips with sound. Videos run 5 to 20 seconds with synchronized audio, and sound on or off costs the same. Draft is a fast, low-cost preview tier at 720p: use it to try a prompt, then render the one you keep at full quality.

Extend and upscale. Extend takes a clip of up to 15 seconds and writes 5 to 20 new seconds from its last frame; you get the new segment only. Upscale enlarges a clip of up to 20 seconds by 1.5 to 3 times and keeps its sound: precise only sharpens, creative also rebuilds fine detail.

Content tests. We add no filter of our own. FLUX 3 runs Black Forest Labs' own moderation: suggestive scenes come back, anything more revealing is refused. Failed or refused requests are not charged.

Six FLUX.2 models from Black Forest Labs are now on SpicyAPI, each with a text-to-image and an edit endpoint: FLUX.2 Max, Pro, Flex, Dev, Klein 9B and Klein 4B.

Prices are per image. Aspect ratio does not change the price, and an edit costs the same whether you send one reference image or three.

What each one is for. Max is the flagship for the best detail. Pro is the everyday workhorse. Flex is strongest on typography and text inside the image. Dev is the 32B open-weight model. Klein 9B and 4B are small distilled models that return an image in well under a minute.

Content tests. We add no filter of our own. Max, Pro and Flex run Black Forest Labs' own moderation: suggestive images come back, more revealing requests are usually covered up or refused, and edits of people are refused more often than text to image. Dev and the two Klein models are open-weight with no moderation, and nothing was refused in our tests, though the smaller models often add clothing on their own.

Every model takes nine aspect ratios at 1k (about one megapixel) or 2k (2048 px on the long side), and up to three reference images for edits. Failed or refused requests are not charged.

Claude Haiku 5.5, released by Anthropic on 7 October, is now on SpicyAPI as one OpenAI-compatible chat endpoint: anthropic/claude-haiku-5.5/chat.

Content tests. We add no filter of our own. In 24 test requests on 8 October the model wrote 20 in full, toned 2 down and refused 2. An intimate scene between two adults came back complete in 8 of 9 attempts; the refusals were explicit, part-by-part figure studies.

Good to know. The model decides how much to think, so there is no reasoning_effort; there is no stop or temperature either. max_tokens does not cut the reply short: ask for the length you want in the prompt, and you pay for the tokens actually written. Cache hits cost $0.10 per million tokens.

Short answers came back in three to five seconds in our tests. Failed requests are not charged.

Google released Nano Banana 2.1 on 6 October 2026, an update to Nano Banana 2 (Gemini 3.1 Flash Image) with better image quality, prompt adherence, character consistency across edits and text rendering. Two endpoints are live today:

• Text to image (google/nano-banana-2.1/text-to-image): 1K, 2K or 4K in fourteen shapes from 1:8 to 8:1, PNG or JPEG, with optional web search and image search. In our tests every line of text we put in quotes came out spelled as written, and web search let a forecast poster print the correct date.
• Edit (google/nano-banana-2.1/edit): one to fourteen source images, with faces and product details kept; auto keeps the source shape.

Google's content policy applies: suggestive fashion renders, while nudity requests come back toned down in text to image and are refused in editing (refused edits are not charged).

Editing adds $0.0024 per source image; web search and image search add $0.0168 each when switched on.

Seven new families join the catalogue, all uncensored, plus a text-to-video endpoint for Wan 2.2 LoRA.

• Prefect Pony XL LoRA (prefect/pony-xl-lora/text-to-image): the Pony XL anime checkpoint with up to three of your own Pony LoRAs per request.
• Z-Image Base LoRA (alibaba/z-image-base-lora/text-to-image, /edit): the non-distilled Z-Image base that most community Z-Image LoRAs are trained on, with image-to-image and a strength control.
• Neta Lumina (neta-art/neta-lumina/text-to-image): Neta.art's anime model, prompted with Danbooru-style tags or sentences.
• Jib Mix Qwen and Jib Mix Qwen LoRA (j1b/jib-mix-qwen-image/text-to-image, j1b/jib-mix-qwen-image-lora/text-to-image): a photorealistic Qwen-Image fine-tune for natural portraits, with or without your own LoRAs.
• HunyuanVideo 1.5 (tencent/hunyuan-video-1.5/text-to-video, /image-to-video): Tencent's open video model, 5 or 8 seconds at 480p or 720p.
• LTX-2 19B and LTX-2 19B LoRA (lightricks/ltx-2-19b/..., lightricks/ltx-2-19b-lora/...): Lightricks' video model with a sound track on every clip, 5 to 20 seconds at up to 1080p, and camera or style LoRAs on the LoRA build.
• Wan 2.2 LoRA text to video (alibaba/wan-2.2-lora/text-to-video): Wan 2.2 with your own high-noise and low-noise LoRAs, now from a prompt as well as from an image.

A LoRA is a direct link to a .safetensors file trained for that base model; loading LoRAs adds no separate charge.

Two new image families run Anima Turbo, the fast 8-to-12-step build of CircleStone Labs' 2-billion-parameter anime model. Prompt it with Danbooru-style tags, plain sentences or both.

• Anima Turbo (circlestone-labs/anima-turbo/text-to-image, /edit): anime illustrations in fifteen aspect ratios at two resolution tiers, usually under ten seconds per image, plus image-to-image that turns photos into anime with a strength control.
• Anima Turbo LoRA (circlestone-labs/anima-turbo-lora/text-to-image, /edit): the same model with up to three of your own LoRA files per request, each with its own strength, for retro game looks, manga ink, characters and adult styles.

A LoRA is a direct link to a .safetensors file: no upload, no allow-list, no review. All four endpoints are uncensored, image-to-image costs the same as text-to-image, and loading LoRAs adds no separate charge.