Updates that matter when you build

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.

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.

Six endpoints join the swap family, three of them for video. What separates them is how much actually travels.

Images

• spicyapi/face-swap-pro-v1/edit — only the facial features move. The target's hairstyle, hairline, head silhouette, pose, wardrobe, lighting and background all stay as they were. $0.16 per image.
• spicyapi/face-swap-v1-plus/edit — moves the face and reconciles skin tone between the two images as part of the transfer; the engine blends either the face alone or the whole head, whichever reads better. $0.16 per image.
• The original spicyapi/face-swap-v1/edit and spicyapi/head-swap-v1/edit stay at $0.013 per image.

Video

• spicyapi/face-swap-video-v1/video-edit — carries the features from a single reference image through every frame. The performer's hairstyle, head silhouette, movement, wardrobe and lighting are untouched. $0.0325 per output second.
• spicyapi/head-swap-video-v1/video-edit — hair, head shape and identity change. $0.064 per output second.
• spicyapi/character-swap-v1/video-edit — the whole figure changes, clothing included. $0.064 per output second.

The two multi-person video endpoints are deliberately two-step. Call the matching video-analyze endpoint first, at $0.08 per request: it follows every person it can see through the clip and returns one preview still each, in a fixed order — the first image is position 0, the second is position 1. You then pass the positions you picked, with one to three reference photos for each. Nothing is left to guess about which person the model will act on.

Three xAI models, one chat endpoint each, all speaking the OpenAI wire format with streaming.

• 40% below xAI's own published price, on every tier, checked against their pricing page on 19 September 2026.
• Only 4.3 accepts reasoning_effort: none. On 4.5 and 4.6 that value is refused upstream — the other four levels work on all three.
• 4.5 caches more cheaply than 4.6 ($0.18 against $0.30 per 1M cached input tokens) while matching it everywhere else.
• Prompts of 200,000 tokens or more bill at a higher long-context rate, exactly as xAI does. Reasoning tokens count toward max_tokens, so budget for them.
• These routes take text and images. They do not accept file or PDF parts.

Five image endpoints, spanning a 4x price range, so you can match the tier to the job rather than paying top rate for a thumbnail.

• bytedance/seedream-4.0/text-to-image and bytedance/seedream-4.0/edit — $0.03 per image, both Uncensored. The edit endpoint takes an existing image and a written instruction.
• krea/2-large/text-to-image — $0.06 per image, Filtered.
• krea/2-medium/text-to-image — $0.03 per image, Softened.
• krea/2-medium-turbo/text-to-image — $0.015 per image, Filtered. The cheapest image endpoint in the catalogue after the swap family.

Two things worth knowing before you pick one. The Krea tiers differ in what they will accept, not only in price: Filtered refuses mature requests outright, while Softened returns a toned-down result and still bills for it — so on a Softened route, check the output rather than the status code. And image size is not a modifier applied after the fact: the example runs on each model page show the exact size each composition was written for, because a prompt built for a tall frame does not simply resize into a wide one.

heartmula/transcribe/speech-to-text transcribes the sung lyrics in an audio file and returns them as text on the task. It is tuned for vocals over music rather than for plain speech.

• One input: audio_url, which must resolve directly to an audio file. There is nothing else to send — no language hint, no timestamp request, no diarisation switch.
• The result is text on the task. No media asset is produced, so there is no file to download and nothing enters media retention.
• No line timings, no confidence values, no speaker labels.
• $0.05 per transcription, regardless of how long the audio is.

It is also the first endpoint in the catalogue whose output is text rather than a file. If your integration assumes every completed task carries a downloadable asset, this is the one that will find that assumption — worth a look at your result handler before you wire it in.