Updates that matter when you build

• Claude Sonnet 5.5 is live. Anthropic's newest Sonnet, released on 28 September, on one OpenAI-compatible endpoint: text in, a 1,000,000-token window and up to 128,000 tokens out — twice Sonnet 5's reply ceiling. Replies stream over SSE, and failed requests are not charged.
• 24 of 24 content probes written in full. No refusals and nothing toned down, including all nine requests for an intimate scene between adults — where Sonnet 5 refused all of its direct and in-story requests.
• Six effort settings. reasoning_effort runs from none to max, with medium as the default; none skips the thinking pass and xhigh and max return their reasoning. tool_choice takes auto or none, and temperature and top_p are not accepted.
• Anthropic's list price. $2 per million input tokens, $0.20 cached, $2.50 cache write and $10 output — the same list price as Sonnet 5.

Two adult-tuned versions of Alibaba's Wan 3.0 are now live on SpicyAPI. Both animate one opening frame into 2 to 30 seconds of video at 480p, 720p or 1080p, with generated audio and an optional last frame.

Which one to use. For most clips we recommend starting with the standard Wan 3.0 or Wan 3.0 Prime. They are uncensored on SpicyAPI too, cost less per second (at 480p, $0.045 for Wan 3.0 against $0.057 for Wan 3.0 Spicy, and $0.0612 for Wan 3.0 Prime against $0.09 for Prime Spicy), keep clothing on when the prompt asks for it, and also offer text to video and reference to video. Switch to a Spicy version when you want a model that leans adult on its own; on Prime, Prime Spicy also delivered the intimate two-person tier that the standard Wan 3.0 Prime softened in our test.

What we tested. We ran our six-tier content ladder on both models at 480p: a strap sliding off, topless, full nudity, two adults in an intimate scene, consensual bondage and impact play, and cinematic gore. Every tier came back as written on both — nothing refused, covered or swapped for another person. On Wan 3.0 Prime Spicy that includes the intimate two-person tier, which the standard Wan 3.0 Prime softened in our earlier test.

Know this before you ship. Both models lean adult on their own. With clothed opening frames and prompts asking the person to stay dressed, Wan 3.0 Spicy undressed the subject in four of six clips and Prime Spicy in two of four. Review whole clips rather than poster frames, and use the standard Wan 3.0 when the wardrobe has to stay on.

Details. Any whole number of seconds from 2 to 30. Aspect ratio 16:9, 9:16, 1:1, 4:3 or 3:4, or leave it out to follow the opening frame. last_image_url pins how the clip ends. In our first batch a five-second 480p clip took about one and a half minutes on Wan 3.0 Spicy and about two to two and a half minutes on Prime Spicy. Billing is per requested second.

Seedream 5.0 Flash from ByteDance is now live on SpicyAPI with three endpoints. Layer decomposition — one image in, a clean base plate plus up to 16 transparent PNG layers out — is available on Flash and on Seedream 5.0 Pro.

Seedream 5.0 Flash. The lighter Seedream tier. 1K, 1.5K and 2K all cost $0.029, so the default is the sharpest one (2K). Prompts up to 5,000 characters, fifteen aspect ratios from 1:3 to 3:1, JPEG or PNG. Mature content needs no extra switch and no separate endpoint: lingerie, boudoir and other adult scenes render like any other prompt, in text-to-image and in edits.

Layer decomposition. Send one image, and optionally name what to separate ("the title text and the parrot"). You get a base plate with those elements removed and the background filled in, plus a transparent PNG for each element. Move, resize, recolor or swap any of them. Typical uses: changing the headline on a poster, swapping a product background, lifting one person out of a group photo, or preparing layers for animation. After any generation you can also send the result straight to "Split into layers".

Flash or Pro for layers. Pro also returns each layer's name and its position on the base plate, so the web result view becomes a canvas where you can show or hide layers, drag them, and export the visible stack as one PNG. Flash returns independent cutouts without positions, shown as a strip you can download layer by layer or as one zip. Pro also accepts more revealing input photos, which Flash declines.

Billing. Every output image counts, base plate included, up to 17 per request. We hold 17 images when the request is accepted and refund the difference once the real count is known. A typical run produces 3 to 9 images, so $0.06–$0.18 on Flash.

Read more. Seedream 5.0 Flash API, tested · Splitting an image into layers with an API

Five speech models are now live on SpicyAPI, joining Gemini 2.5 Pro TTS. None of them adds a content filter: in our tests all five read NSFW scripts in full.

• Gemini 3.8 Flash TTS — Google's newest speech model. Up to 5,000 characters, 30 voices, WAV out, and it acts out bracketed cues such as [whispering] or [sighs]. $0.00032 per second of audio.
• Gemini 3.8 Flash-Lite TTS — the faster, cheaper sibling. Same 30 voices; it sometimes reads bracketed tags aloud, so spell sounds out. $0.00021 per second of audio.
• Seed Speech 2.0 by ByteDance — 41 preset voices across ten languages, a plain-language note on delivery, MP3 out. $0.072 per 1,000 characters.
• Qwen3 TTS by Alibaba — open-source, nine voices including Japanese, Korean and two Chinese dialect speakers. $0.068 per 1,000 characters, 100-character minimum.
• Chatterbox by Resemble AI — an open-source English voice with an expressiveness dial. Up to 1,000 characters for a flat $0.036 per request.

Priced before you run. Seed Speech 2.0 and Qwen3 TTS are the first models billed by the character: the count is known up front, so the quote is the final price. The Gemini models bill per second of audio returned, with a hold that is released once the audio is measured.

What they don't do. No voice cloning on any of these endpoints, and no two-voice dialogue in one request — generate each line with its own voice and join the files.

Google's Gemini 2.5 Pro text-to-speech is now live on SpicyAPI as google/gemini-2.5-pro-tts/text-to-speech.

• Input: a script of up to 5,000 characters. Cues in square brackets such as [sighs] are performed, not read aloud.
• Voice: one of 30 prebuilt voices, each with a fixed character — from bright and firm to breathy and warm.
• Delivery: an optional style (whisper, empathetic, vocalsmile, newscaster, promohype or deadpan), one of 8 English accents, and free-text instructions on pace, mood and who's listening — none of which is spoken aloud.
• Output: one WAV file, 24 kHz, 16-bit mono.
• Price: $0.00045 per second of audio returned (about $0.027 per minute), 10% below Google's list price. A hold for the longest possible clip is placed while the task runs and released once the audio is measured; failed requests are not charged.

How far it goes. We add no content filter. In our tests it read NSFW scripts in full — direct lines, narrated fiction, breathy vocal cues — without skipping words or cutting the clip short. There's no two-voice dialogue on this endpoint and no voice cloning: one voice per request, from the 30 prebuilt options.

Try Gemini 2.5 Pro TTS →

Anthropic's newest Opus, released on 22 September, is now on SpicyAPI as anthropic/claude-opus-5.5/chat, one OpenAI-compatible endpoint at /v1/chat/completions.

• Price: $3.60 per million input tokens, $0.18 cached, $4.50 cache write and $18 output, 10% below Anthropic's list price.
• Window: 1,000,000 tokens in, up to 128,000 out.
• Thinking is always on: reasoning_effort takes low, medium or high, with medium as the default.
• Tools: up to 128 definitions; tool_choice accepts auto or none, because the model rejects a forced tool call.

How far it writes. We add no content filter of our own. On launch day we sent it 24 adult-content probes across three prompt styles and it refused none. Under an adult-fiction system prompt it wrote every intimate scene in full; asked directly, it usually wrote a toned-down version instead, where Claude Opus 5 refused outright.

Try Claude Opus 5.5 →

Kling's current video lineup is now on SpicyAPI, billed per second in US dollars with no package to buy up front.

• Kling O3 (kling/o3/text-to-video, /image-to-video, /reference-to-video, /video-edit): Kling 3.0 Omni. Cast up to seven reference photos into one shot, edit a clip you already have (a 3 to 10 second source), and turn on smart multi-shot cuts and native sound, up to 4K.
• Kling 3.0 (kling/3.0/text-to-video, /image-to-video, /character-animation): the flagship, with a first and last frame, up to 4K and switchable sound. Motion Control copies the moves, expressions and gestures from a 3 to 30 second clip onto your own character.
• Kling 3.0 Turbo (kling/3.0-turbo/text-to-video, /image-to-video): every clip comes with native sound, at 720p or 1080p.
• Kling 2.6 (kling/2.6/text-to-video, /image-to-video, /character-animation): Kling's first native-audio model, 1080p in 5 or 10 seconds, with its own Motion Control.
• Kling 2.5 Turbo (kling/2.5-turbo/text-to-video, /image-to-video): budget 1080p in 5 or 10 seconds, with a closing frame, a negative prompt and guidance.
• Kling 2.1 (kling/2.1/image-to-video) and Kling 2.1 Master (kling/2.1-master/text-to-video, /image-to-video): the previous generation, from 720p image-to-video to Master's high-fidelity 1080p.
• Kling Avatar 2.0 (kling/avatar-2.0/talking-avatar): one portrait and a voice track make that person talk or sing for as long as the audio runs, up to about five minutes.

Kling reviews content itself: suggestive shots sometimes pass, anything more revealing is refused, and a refused request is never charged. For adult work, use one of our Spicy families.

Kling O3, Kling 3.0, Kling 2.5 Turbo and Kling Avatar 2.0 are priced below Kling's own per-second list price.

Three new image families run open models with up to three of your own LoRA files per request, each with its own strength. Every image is one flat price, whatever the aspect ratio.

• Z-Image Turbo LoRA (alibaba/z-image-turbo-lora/text-to-image, /edit): Alibaba's 8-step Z-Image Turbo, a few seconds per image, nine aspect ratios at about one megapixel, plus image-to-image with a strength control.
• FLUX.1 Dev LoRA (black-forest-labs/flux-1-dev-lora/text-to-image): Black Forest Labs' FLUX.1 [dev], the base with more than a thousand community LoRAs on Hugging Face, with a guidance control.
• Qwen Image 2512 LoRA (alibaba/qwen-image-2512-lora/text-to-image, /edit): Qwen-Image-2512 at up to 1536 pixels on the long side, and Qwen-Image-Edit-2511 editing with one to three sources and edit LoRAs such as new camera angles.

A LoRA is a direct link to a .safetensors file: no upload, no allow-list, no review. All five endpoints are uncensored, and loading LoRAs adds no separate charge.

Eight single-purpose endpoints went live on 26 September. Each does one job on an image or clip you already have, at a flat price. Call them like any other model, or use them in Spicy Studio, where they sit under Tools and on your results.

What decides the bill:

• Video upscaling bills on the source clip's length, rounded up to the whole second, with a 3-second minimum.
• Talking avatar and lip sync bill on the returned video, which runs as long as your audio (3-second minimum). Lip sync loops a clip that is shorter than the audio.
• Sound effects bill on the duration_seconds you ask for (1–30, default 8), not on your clip, so set it to the clip's length. It makes effects and ambience, not dialogue or music.
• Audio and video inputs must be direct links to the file, and uploading them is the most reliable route. Inputs outside the allowed length are rejected before anything is charged.

xAI released Grok 4.7 on 21 September, and xai/grok-4.7/chat has been callable here since 22 September, on the same OpenAI chat format and SSE streaming as the other Grok endpoints.

• 40% below xAI's published price on the standard tier ($2 / $0.50 / $6), and 60% below its long-context tier ($4 / $1 / $12).
• Uncensored on this route. On 22 September, nine creative-writing probes at level 4 of our five-level content ladder, asked three different ways, came back as nine complete scenes with no refusal; on 24 September every other level we ran passed too, the top one included. We add no filter of our own.
• The system role is honoured, checked on purpose, so a frame set at the top of a thread carries through it.
• Reasoning bills as output. The high default costs more than it looks; for dialogue, low or medium reads better and costs less.

Since 24 September both MiniMax H3 Max endpoints accept resolution: "480p" alongside 768p.

Prices are per second of output and match MiniMax's own published rates. 480p is a native tier, not a downscale: the cheap way to check motion and framing before you pay for 768p.

Five content labels moved. Another round of content tests on the routes we serve finished on 24 September, and these labels changed:

• Softened means a request beyond the model's own limits comes back toned down instead of failing, and it is billed like any success, so check the output rather than the status code. Filtered means the request is refused with content_rejected and refunded.
• If you picked H3 Max image-to-video for uncensored work after our 21 September post, minimax/h3/image-to-video (from $0.025 a second) and both Wan 2.5 endpoints now carry the Uncensored label.

Four families join the video catalogue, priced per second of output. Every one carries a measured content label rather than an assumed one.

• minimax/h3-max/image-to-video and text-to-video — $0.08 per second. Image-to-video is Uncensored; text-to-video is Softened, meaning a mature prompt returns a toned-down result rather than an error, and is still billed. Check the output, not the status code.
• alibaba/wan-2.6/text-to-video and image-to-video — $0.09 per second, both Uncensored. reference-to-video is $0.065 per second and Filtered: it refuses mature requests outright.
• alibaba/wan-2.6-flash/image-to-video — $0.0225 per second, Uncensored. The cheapest per-second video in the catalogue by a wide margin.
• lightricks/ltx-2.5/text-to-video and image-to-video — $0.09 per second, both Uncensored.
• vidu/q3-pro/image-to-video at $0.054 and vidu/q3-turbo/image-to-video at $0.042 per second, both Uncensored.

Labels come from real measurement on the route we actually serve, so they change when the route does. A model with no label is one we have not measured yet — we would rather say nothing than guess on your behalf.