Updates that matter when you build

Four new places to decide which model to use, all built from runs on this platform. Leaderboards, reviews and the arena sit under Rankings in the site header; the prompt library sits under Studio.

• Leaderboards score each model on two separate scales. The Spicy Index (0–100) is preliminary for now: it scores capability only, and output quality joins it once our quality tests and arena votes are in. The Freedom Score (0–100) measures how far each model verifiably goes on our content test ladder. The two never mix, and price, speed and reliability sit beside them as raw measurements. There are boards for video, image, LLM, uncensored, best value and speed; as of 26 September, 53 models are ranked from 1,672 test runs, and scores are recalculated daily.
• Reviews: eight so far — MiniMax H3, MiniMax H3 Spicy, Seedance 2.5, Seedance 2.5 Spicy, Wan 3.0, Seedream 5.0 Pro, Z-Image Spicy and Grok 4.7. Each shows real outputs, a test report, what a usable result costs and a clear pick.
• Blind arena: two anonymous image or video results from one prompt. Pick the one you would keep, then see which models made them. Signed-in votes set Bradley–Terry ratings with 95% intervals, refreshed hourly; models whose intervals overlap share a rank. Ranks appear once enough signed-in votes are in; until then the models are listed by maker.
• Prompt library: 176 prompts for video and image models, each shown with the output it actually produced. Filter by model, task, scene or camera move, then copy the prompt or open it in the playground with every setting filled in.

The methodology is public and versioned. We sell access to every model we rank, so the method page also explains how that is kept out of the scores.

Four changes to how tasks behave. Two may need a change on your side: if you rely on identical requests returning identical results, send a seed; if you branch on error codes, handle 40310.

A failed run is tried again on another route first. Many models here are served on more than one route. Since 23 September, when a run fails after we accept it and another route serves the same model with the same settings, we run it there before reporting a failure, as long as the task has time left. You are charged once, for the result you get. Failures that would repeat anywhere, such as a copyright refusal, are not retried.

Refusals explain the cause and the fix. When a model blocks a result over possible copyright or a real person's face, errorMessage now says which it was and what to try:

• Video flagged for copyright is most often the generated soundtrack: set generate_audio: false, drop any reference audio, and leave real songs and artists out of the prompt.
• Images flagged for copyright usually show a well-known character, logo or artwork, in the prompt or in a reference image.
• Some models do not accept real people's faces in reference images. The message says so and suggests another model or an illustrated character.

errorCode stays content_rejected, the charge is refunded, and the message follows your API error language like every other error.

A random seed when you leave seed out. Since 24 September, on models that take a seed, a request without one is given a random seed, returned in the task's input. Two identical requests no longer return the same result. To keep runs comparable, send the returned seed back; a fixed seed does not guarantee an identical result. A few models vary the seed without returning it. A seed you send is always used as given.

40310 for accounts whose email is not verified. Keys on such accounts used to get a generic 403. Since 25 September they get business code 40310, with a message saying to open the verification link or send a new one from the Console. The key starts working as soon as the address is verified.

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.

Three chat endpoints join the catalogue, and the reasoning controls across the existing text models now match what each publisher actually supports.

• zai/glm-5.3/chat — $1.05 in / $0.195 cached / $3.30 out per 1M tokens, 25% below Z.ai's published price. Content label: Light, meaning restrictions are lighter but some text-only prompts may still be limited.
• zai/glm-5.3-flash/chat — $0.1275 / $0.0255 / $0.425 per 1M tokens, Uncensored.
• deepseek/v4.1-flash/chat — current price $0.30 / $0.006 / $1.20 per 1M tokens, Uncensored.

Reasoning effort is now per model, not one list copied everywhere. Claude 4.5 and 4.6 carry the levels Anthropic publishes for each; Gemini gained its minimal tier; DeepSeek accepts none, which we measured actually turning reasoning off rather than merely lowering it. Where a publisher does not offer a level, we do not advertise it — Grok 4.5 and 4.6, for instance, reject reasoning_effort: none upstream while 4.3 accepts it.

One budgeting note that catches people out: reasoning tokens count toward max_tokens. A prompt asking for 150 words can still burn several thousand tokens on a high effort setting, so size the limit for what the model will actually emit, not for the answer you want to read.

Three new families run the open-weights MiniMax H3 with up to three of your own LoRA files per request, each with its own strength from 0 to 4.

• MiniMax H3 LoRA (minimax/h3-lora/text-to-video, /image-to-video, /reference-to-video): 3 to 15 seconds at 480p, 540p, 768p or 1080p, with stereo sound generated in the same pass. Reference-to-video takes up to nine images, three clips and three audio clips.
• MiniMax H3 Singularity LoRA (minimax/h3-singularity-lora/image-to-video, /reference-to-video): the community's HDR-trained fine-tune of H3, built for fast action, dance and wide shots where faces must hold. Your H3 LoRAs load on it unchanged.
• MiniMax H3 Image LoRA (minimax/h3-image-lora/text-to-image, /edit): H3 as a still-image model at 1k or 2k, with editing from up to nine references. Design a key frame here, then animate it with the same LoRA.

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

From 23 September 2026, both Wan 2.2 LoRA image-to-video endpoints are 15% off, with no end date set.

Prices are per second of output.

• Nothing to change on your side. The discount is applied when a request is priced, so the estimate before you submit, the charge afterwards and the price cards all show the new figure. Resolutions, durations, content behaviour and request shape are unchanged.
• This gives back most of the 20 September increase. The condition behind that increase has since moved in our favour, and we are passing it on. At $0.0204 the price is still 2% above where it was before the increase, not below it.
• No end date is set. If that ever changes, the date will appear on the price card.

Alibaba open-sourced Qwen-Image-2.1 today, and it is in the catalogue the same day. Four endpoints: generation and reference editing, each in a plain edition and in a LoRA edition that loads weight files you point it at.

• Bring your own weights. Up to three LoRA files per request, each with its own strength, given as a direct HTTPS link to a .safetensors file or as a Hugging Face repository. No upload step, no allow-list, no approval queue.
• Nothing sits between the prompt and the model. All four endpoints are unrestricted: no content filter in front of the model, no maturity flag to set, no declaration to attach to a request.
• Fifteen aspect ratios, three resolution tiers. From 3:1 and 21:9 strips down to 1:3 and 9:21 columns, at 1k, 1.5k or 2k.
• Editing takes up to ten reference images. Each one is priced into the estimate you see before you submit, so send the references that are doing work and drop the rest.

moonshot/kimi-k2.6/chat is no longer in the catalogue. A request that names it now comes back 404, and its web address permanently redirects to Kimi K3.

• Replace the model ID with moonshot/kimi-k3/chat. Nothing else in your request changes: same OpenAI wire format, same streaming behaviour.
• Tasks created before the retirement are unaffected. You can still read them with the key that created them.
• A retired endpoint is never quietly rerouted to a different model. We would rather a request fail loudly than charge you for a model you did not choose — a substitution you cannot see is worse than an error you can.

Since 20 September 2026, both Wan 2.2 LoRA image-to-video endpoints cost 20% more per second of output.

Nothing else about them changed: same resolutions, same durations, same content behaviour, same request shape.

Why we are telling you rather than letting you find it on the invoice. These two were priced against an assumption that turned out not to hold for them specifically, and we corrected it rather than keep a price we cannot sustain. A rate that quietly stops covering its own cost does not stay available for long, and an endpoint that disappears is worse for you than one that costs 20% more.

One visible side effect: the Spicy endpoint no longer displays a struck-through reference price. We show one only where we are genuinely below the reference, and at $0.024 we no longer are. The reference price itself did not move — we did.

Five official SDKs are live, alongside a CLI, an MCP server, an Agent Skill and a key-less proxy. All of them implement the same four steps: quote, create with an idempotency key, wait, collect.

• Almost nothing enters your dependency tree. Python and Go use the standard library only, PHP needs no library either, and Java brings exactly one: Jackson.
• The whole asynchronous flow, not just the HTTP call — a wait budget on task creation, the three-step upload for your own images and video, and webhook signature verification.
• @spicyapi/cli browses the catalogue, quotes a price before you spend and downloads finished results from a terminal. @spicyapi/mcp gives Claude Code, Codex, Cursor or any MCP client the same abilities, asking you before anything is billed. The Agent Skill installs with npx skills add Spicy-API/spicy-skill and teaches a coding assistant the correct order of operations.
• @spicyapi/proxy keeps the key on your server. A key compiled into a browser, mobile or desktop app is a public key; this package is the server-side hop that holds it instead.

All five are MIT-licensed and open source. Guides for each: docs.spicyapi.ai

Two ways to stop writing polling loops. Both are live.

A wait budget on task creation. Append ?wait=<seconds> to createTask and we hold the connection open until the task reaches a terminal state, answering with the full task record — the same shape recordInfo returns.

• Budget runs out: you get the ordinary 202, identical to omitting it, and poll from there.
• You disconnect: nothing changes for the task. It keeps running and bills as usual.
• The ceiling is 60 seconds. Larger values are clamped rather than rejected, and a malformed value is ignored rather than rejected — a typo in an optimisation parameter should not fail a generation that would otherwise succeed.

Use it on endpoints that finish in seconds, which is most image and edit work. Video models routinely run for minutes, so it will simply time out and you end up polling anyway.

Account-level webhook endpoints. Register a URL once together with the events it wants, and every matching task is delivered to it — you never pass a callback URL on a task again. An account holds up to 16 endpoints, each with its own signing secret. A 410 Gone from your server disables that endpoint immediately rather than retrying an address whose owner has retired it. Endpoints are API-only for now; the console has no screen for them yet.

Details: tasks · webhooks