GPT-Image 2.5 arrives in two versions: Flare for fast generation and Sunburst for precise edits. Both add Extra High and Max quality, sizes up to 4K and transparent backgrounds, and work with your own key or with credits. Credit prices now round up to the nearest 50 credits, with a minimum of 150 credits per image.

Credit packs can now be purchased in the app. Sign in, buy credits, and generate with supported image models without adding your own provider key. If you use your own provider keys, nothing changes: those generations are billed by that provider and never use credits.

Generate without bringing an API key. Sign in, buy credits, and use supported image models through hosted generation. If you use your own provider keys, those runs are billed by that provider and never consume credits. The new Credits page shows your live balance, the cost of a pending generation, and detailed local usage history with images, models, providers, and per-run totals. When you use your own provider keys, the Usage view estimates provider costs in dollars instead. Purchases now handle cancellations cleanly and show failures in a dialog. Credit prices follow the selected model, size, and quality. Chat text replies and automatic conversation titles work again after their upstream text models were retired. Hosted image delivery is also safer: a storage failure no longer deducts credits.

You can now choose the interface language. A new Language section in Settings switches between 50 languages on the spot - no restart, and a half-typed prompt survives the switch. Your choice is remembered the next time you open the app. The translations themselves got a pass too: hundreds of labels that were still English in every language are now translated, and wording that had drifted apart between screens has been reconciled. Dropdown menus open in the app’s own dark styling instead of the plain system popup, so a list matches the panel it belongs to. Scrollbars stay out of the way until you scroll. Zooming a scene with the scroll wheel now stays anchored under the pointer, so the spot you point at is the spot you zoom into. This update also fixes a blank window when opening Chat before any model was configured, and a language list that would not scroll past its eighth entry.

Selections on the scene canvas now identify themselves: layers and groups show their name on the selection, in a label that stays readable at any zoom level and reveals more text as you zoom in. The selection outline and resize handles also keep a constant size at any zoom. Zooming the canvas is easier: hold Cmd or Ctrl and scroll, or pinch on a trackpad. Generations and flow runs started through the local MCP server now appear in the task UI while they run, instead of only when they finish. This update also includes interface polish and stability improvements.

Scenes now have a full layer stack. Group layers together, nest groups inside other groups, collapse the ones you are not working on, and select several at once. Drag to reorder layers and groups, including mixed selections. Undo and redo arrive for scene editing, so any step can be taken back. Groups gain the controls layers already had — hide, duplicate and lock — and dragging a selected group on the canvas moves everything inside it. Double-click a layer or a group to rename it in place. Table columns can be resized and keep their width across the app. Lumiere also reopens at the window size you left and on the page you were last working in. This update fixes reference images when editing with OpenAI GPT-Image, restores the step details of flows in scenes, and brings interface polish and stability improvements.

Seedream 5.0 Pro joins Lumiere through fal.ai, with 1K/2K output, flexible aspect ratios, and reference-image editing. This update also adds flow-driven scene layers, fixes background removal in packaged builds, and includes interface polish and stability improvements.