Changelog

📦 New

• GPT Image 2.5. Four nodes: Flare and Sunburst, each as text-to-image and as an edit. The edit nodes take up to 10 reference images, so a real product can land in a generated scene with the label still readable. Five quality tiers from low to max, priced live on the card.
• Clone a voice once, use it everywhere. Record a sample right in the browser. Clone it with Fish Voice Clone once your consent is checked. It shows up by name in Fish TTS on any workflow after that, no recloning needed. Also reachable through the API and MCP (prepare_voice_clone needs the opt-in write scope).
• Two new video models. Gemini Omni Flash 1.1 generates from text, a first frame, reference images, or an existing video, at 360p, 720p, 1080p, or 4K. Hailuo 03 Max Turbo renders faster than the clip plays, at half the per-second price of Hailuo 03 Max.
• Songs with real lyrics. MiniMax Music 2.6 sings section-tagged lyrics you write, like Verse and Chorus. Lyria 3 Pro writes its own lyrics off one prompt and an optional mood image. ElevenLabs Music moved to v2.5 and picked up a lyrics port.
• Motion graphics, for agents. Ask for a reveal in the Edit composition box, or build one over MCP or the API, and it can now use shape clips (rect, ellipse, line, arrow, path), masks, crop, rounded corners, pop-out cutouts, styled text with a typewriter reveal, and music that ducks under the voiceover on its own. Editor controls for these come next; today they are agent-built.
• Camera keyframes, in the editor. Stage scenes got real UI for this one: keyframe the camera on its own lanes and move or edit the children directly on the timeline, no scene graph hand-editing needed.
• New free canvas tools. Color Palette turns colors into a swatch sheet, a prompt string, a list, and JSON at once, free to run. Feed it a mood board or a website’s brand colors, or just type hexes. Painter brushes a mask on an image for regional edits. GLSL Shader runs your own fragment shader over an image and turns its uniforms into sliders and color pickers. Video Trim cuts an exact window for 0 credits.
• Agents can post for you. New MCP tools, opt-in: list_social_accounts, publish_post, get_post_status. They reach LinkedIn, Threads, Bluesky, Mastodon, and YouTube (YouTube uploads may stay private until Google finishes reviewing Wireflow), with the same ban-safety limits the app enforces.

🧠 Improved

• Blueprint pages show themselves working. Every public Blueprint page now has a "See it run" gallery of real outputs pulled straight from the flow, so you can judge a blueprint by its results before you fork it.
• ElevenLabs v4 and v4 Turbo are now options on the ElevenLabs TTS (Custom Voice) node, with inline tags like [whispering] and [chuckles].
• Instagram Feed format in the Video Editor. Pick Instagram Feed as an output format and set the name, handle, and caption right there.

📦 New

• Marble 3D worlds. Turn a single image into an explorable 3D world from World Labs, and explore it inside Wireflow: orbit, zoom, and fly through the world right on the node card. Frame the shot you want with the camera and that framed view becomes a real output you can wire downstream, into a video model or the compositor. Previz for filmmakers, walkthroughs for archviz. A still world now renders on demand, so an open board no longer spins your laptop fans.
• Meshy 7. Text or image to 3D model, as one node that routes to the right endpoint based on what you connect. Captures trim to the visible content automatically, and the Rodin card learned the same trick: frame a shot with its camera and export it as a PNG with transparency.
• Three new video models. MiniMax Hailuo 03 Max, the top tier of the Hailuo family. Seedance 2.5, now at the top of the video lane. And FLUX 3, connected directly to Black Forest Labs, with keyframes, first/last frame control, and extend-video. FLUX Video Upscale joins it for cleaning up finished clips.
• Teams. Create a team, name it, invite people by email, and belong to more than one, the way Figma does it. Invited collaborators land on the board they were invited to, and shared workflows open properly for everyone.
• Color Grade node. Grade your footage upstream of the composition, with a realtime preview and draggable dials living on the card itself. No round-trip to an external tool.
• Document node. A markdown brief that lives on the canvas next to the work it describes. Double-click for a fullscreen reader, deep-link straight to a document, and @-mention any node to link it, with the mention staying live through renames.
• @ references in prompts. Type @ in a Seedance prompt to reference another node's output by name. The picker drops in below, handles highlight in the prompt, and the model receives the actual media.
• Texting Story, animated. The iMessage-style template can now hold photo and video bubbles inside the conversation, bubbles can open on motion, and the video runs exactly as long as the chat does instead of a fixed duration.
• Connect your AI agent. Wireflow now speaks MCP: connect Claude, Cursor, or any MCP-capable agent and let it browse your workflows, run them, and preview a video cut frame-by-frame before paying to render it. Writing workflows is a separate opt-in scope, and agents can file feedback when something blocks them. See wireflow.ai/mcp.
• Bigger uploads. Files upload directly to storage now, so the documented 25MB actually works, and an Import node paints the file you dropped instead of a placeholder.

🧠 Improved

• Sign-in works on work email. Corporate email scanners (Microsoft 365, Mimecast) were consuming magic links before you could click them, so business inboxes always saw "link no longer valid". The link now lands on a confirm page a scanner cannot trip, and the email also carries a code you can type to finish in the tab you started in.
• Canvas quality wave. Standard zoom shortcuts (Cmd +/-, Shift+1 to fit, Cmd+0 for 100%), duplicate preserves a selection's internal wiring, port navigation with hover cards and jump-to-source, copy and paste a node's inputs onto another node, one-click "test with another model", and dragging a group keeps every node in formation. The workflow grid is virtualized, so accounts with 1,000+ workflows scroll smoothly.
• Boards arrange themselves. When an agent or an API call adds nodes, the board lays itself out instead of piling nodes in a heap, the layout engine understands groups, and an agent can declare stages and get an organised board back.
• Live collaboration is actually live. "Collaborate live" now connects reliably, and every open canvas hears server-side changes as they happen, including your own solo boards updated by an automation or an agent.
• Billing you can trust. Running a workflow asks before re-running paid nodes that already finished. A failed run delivers the paid results that did complete and refunds the rest. A free node crashing can no longer eat a paid run's credits, cost estimates account for batch fan-out everywhere, and refunds land with whoever actually paid.
• Runs stop serving stale results. Every way of starting a run now notices when upstream inputs changed, not just the play button, so you never pay to render yesterday's data.
• Top-up packs estimate videos too. Credit pack cards now translate to "about N videos" alongside images, so you know what you are buying.

🔄 Changed

• One place to get help. Support is one consistent entry point per surface now, in the bottom-right corner, and it no longer covers the run panel.
• Compositor keeps maturing. Content-driven sizing (hug, fit-to, anchor) is now authorable in the visual editor, stack layers bring Auto Layout-style flow, corners can round per-corner, and you can design a static card in the compositor and composite it over picture.

📦 New

• Automations calendar. Scheduled runs appear as slots on a month grid you can drag between days. Cadence scheduling plans irregular, human-looking timing ("4 to 5 times a week, weekday mornings, never two close together") instead of a rigid cron, with a credit forecast for the whole schedule. Open to all paid plans, and you get notified if a scheduled run does not happen.
• Bluesky, Mastodon, and Threads. Three new native connectors join LinkedIn and YouTube, with scheduled posting on all five. Mastodon asks which instance you are on, and Threads tokens refresh automatically before they lapse.
• Layerize Image. One flat image in, separate movable layers out: each object becomes its own transparent layer with a real name ("laptop", "coffee mug"), partly hidden objects are reconstructed whole, and the background comes back filled in. Wire it straight into the Compositor. 16 credits per run.
• Reddit research. Browse a subreddit or search Reddit from the canvas and pull a post's title and body as text for downstream steps, like feeding an LLM. Free to run on the same daily search allowance, joining TikTok, Meta Ads, and Reels.
• Texting Story template. An animated iMessage-style story ad: full-frame chat, an opening hook, a reveal clip, beat-synced sound effects, and a CTA card, rendered vertical at 1080x1920. It can also end on the product: a closing act with a paste-in moment and numbered replies.
• Google Drive nodes, finished. Fetch from and save to Drive with a real folder picker, search, a run button, and a link to the saved file. They were previously configurable only through the API.
• MiniMax Hailuo 03. New video model as one node over three endpoints: it routes to text-to-video, image-to-video, or reference-to-video based on what you connect.
• Compositor sizing that follows content. A text box can hug its content instead of a fixed size you have to guess, a container can auto-wrap another layer plus padding, and a layer can anchor to another layer's edge. All opt-in, existing layers untouched. Available today when authoring compositor JSON through the API or an agent; editor controls come next.
• An API agents can drive. MCP tools now create, clone, and edit workflows (opt-in write scope, always private drafts) and schedule recurring runs. The public API catalog grew from 8 to 19 documented operations, agents get the full schema for authoring a composed video, and API keys can be created with specific scopes instead of full access.

🧠 Improved

• Cancel means cancel. Cancelling a run now stops paid nodes from firing at the next layer (a cancelled 60-item batch used to keep billing to the end), an API key can cancel its own runaway run, and Retry either works or is not offered.
• Runs honor your inputs. Running a template through the API silently ignored the caller's documented inputs on most published templates and ran the author's demo values instead. It now refuses with a clear error rather than spend your credits on the wrong thing, and requesting specific nodes runs only those nodes and what feeds them, not the whole graph.
• More refunds you don't have to ask for. Ask a FAL image model for 4 images and get 1, and the difference comes back automatically. Refunds on shared workflows go to whoever actually paid. Billing history now shows accurate credits used per run.
• Find the right node faster. The Add Node menu ranks models by real usage from the last 90 days, you can star favorites to pin them, keyboard navigation works throughout, searching a node by name puts it first instead of 28 lookalikes, and every node's menu now links to its documentation.
• Shared workflows open properly. An invited collaborator used to land on an empty new-workflow screen; they now get the board they were invited to, and view-only collaborators can duplicate it into their own account.
• Blueprints got clearer. Blueprint pages render the actual node graph, list the outputs a blueprint really produces, and can be re-published, and a placed blueprint node links back to its source.
• Mobile canvas shows your balance. The credit balance and upgrade button now render on mobile widths instead of being hidden.

🔄 Changed

• Security hardening. Run history on published templates and apps is now private to whoever paid for the run, retry and resume are restricted to the payer, signed private URLs are redacted from public blueprint snapshots, and API keys are scoped at creation.
• Honest numbers. The typical credit rates advertised across the site had drifted from reality; the headline numbers and plan cards now quote real measured averages.