Changelog
Every model we switch on and every product change, on the day it shipped. Prices are in credits — the same number you see in the workspace.
RSS feedWhen Claude or ChatGPT starts more generations than LUVI runs at once, the extra ones now wait in line instead of failing. Each gets its id straight away and starts by itself as soon as a slot frees. Nothing is charged while it waits, and cancel_generation removes one that has not started yet.
You can now delete your account from Account → Settings. The account is disabled at once and permanently removed after a 14-day grace period, during which the confirmation email restores it in one click.
Every active model has a public page at /models — parameters, starting price in credits, the prompting guide we use ourselves — grouped by family and category, with no sign-in needed. The MCP tools are documented at /docs/mcp.
The Workspace works on a phone: compose, pick references from the library, generate. Videos play through a phone-sized 720p rendition instead of the 4K master, loop, and iPhone .mov uploads are accepted and remuxed on ingest.
MiniMax H3, Gemini Omni, Wan, Hailuo, Qwen-Image and Vidu now carry a prompting guide, distilled from each vendor's own documentation. It shows on the public model pages, powers the prompt rewrite toggle in the workspace, and is what Claude and ChatGPT read through get_prompt_manual before generating.
This page. Every model we switch on and every product change lands here on the day it ships, with the price in credits and a link to the model page. Subscribe to the RSS feed to follow along.
Every generation is now charged from what the provider actually billed, not from the estimate. The number you see before you run is a hold; the deduction that lands is worked out from the provider's settled price, and it is never above the hold.
Black Forest Labs' joint audio-video line is live again after the provider had closed it: text-to-video, image-to-video, keyframes, first-and-last frame and extend. One render carries a multi-shot sequence with its own dialogue, ambience and music. Pricing was re-verified against the provider's real bill before the switch was flipped.
Every uploaded image is stored as JPEG or PNG. HEIC from an iPhone, WebP, GIF, AVIF, TIFF, BMP and SVG are converted on the server, in every browser, with orientation and colour profile preserved. Nothing to do on your side.
Alibaba's Wan 3.0 joins in six rows: text-, image- and reference-to-video for both Wan 3.0 and the Prime tier. It generates picture and sound together, runs up to 30 seconds, and takes reference images, videos and audio in one request. Prime starts from 2,800 credits for a 5-second clip.
OpenAI's two GPT Image 2.5 tiers, each with text-to-image and edit. These rows are charged at the provider's settled price rather than the pre-run estimate, so what you pay follows what the provider actually billed — never more than the hold you saw.
Three monthly plans replace ad-hoc top-ups: each includes a credit allowance, storage and a bonus on the first month. Plan pricing is applied inside the same formula that prices every generation, so the number on a model page is the number you pay.
Claude and ChatGPT can now tell you what a generation cost: get_generation_status, get_asset and list_library carry the real charge in credits, and list_transactions returns the same ledger as Account → Billing.
Google's caption-driven video model in five rows: text-to-video, image-to-video, reference-to-video, video edit and video extend. Sound is generated in the same pass, edits are one sentence long, and a clip can be extended up to 40 seconds. A 10-second clip starts from 2,200 credits.
MiniMax's H3 family generates voice, sound effects and music together with the picture, with lip-synced dialogue in eleven languages and reference images, videos and voices in one request. H3 Max is the fast tier and starts from 1,216 credits for an 8-second clip; H3 from 2,080.
Redeem a promo code in Account → Billing. A code attaches its plan, its discount applies to every generation for as long as it runs, and any bonus credits arrive immediately.
Deleting a folder no longer deletes it. It moves to the trash for 24 hours, its files stay exactly where they were, and Restore brings the folder and its parents back. After 24 hours the folder is removed and its files return to the project root.
xAI's typography-led image model, text-to-image and edit, at up to 2K. It plans a layout like a designer, so posters, packaging and app screens with real copy are its home turf. From 120 credits per image.
Google's first Omni line on LUVI: text-to-video, image-to-video, reference-to-video and conversational video edit, sound included. A 10-second clip starts from 2,500 credits.
ByteDance's Seedance 2.5 arrives with text-to-video, image-to-video, first-and-last frame, reference-to-video, edit and extend. It follows timecoded beats and generates its own audio. A 5-second clip starts from 3,030 credits.
Eighteen audio models in one release: text-to-speech from ElevenLabs v3, Gemini TTS, MiniMax Speech and Seed Audio; music from MiniMax Music and Suno; speech-to-text. Text-to-speech is billed by the character — ElevenLabs v3 from 200 credits per 1,000 characters — and results play inline in the library.
Two ways to get a local file into LUVI from a chat: get_upload_link opens the upload dialog locked to a project and folder, and create_upload_ticket lets Claude's code execution push the bytes straight to storage — files up to 1 GB, no size cliff.
From Claude or ChatGPT you can now move existing generations and uploads into folders with move_to_folder, and audio results come back with an inline player instead of a link.
OmniHuman 1.5 turns a photo and an audio track into a talking head; sync lipsync v3 re-syncs the mouth in an existing video to new audio; VEED lipsync dubs a video to a new voice. All three are priced per second of input, from 130 credits per clip on VEED lipsync.
ByteDance's Seedream 5.0 Pro for stills: text-to-image, edit and sequential sets from one prompt. From 144 credits per image.
Alibaba's joint audio-video family in text-, image- and reference-to-video plus video edit. Its timecoded shots are honored, so a storyboard renders as a storyboard, and it lip-syncs dialogue in seven languages. A 5-second clip starts from 2,800 credits.
A photo and a voice track become a talking-head video, billed by the length of the audio. Standard and fast tiers, 480p or 720p. From 1,650 credits per clip.