Changelog
Every shipped improvement, newest first.
Switching models now tells you which settings it changed
Switching models now announces the settings it adjusted.
Every model offers a different set of resolution and duration tiers. When you switch, your current settings land on the closest tier the new model supports — and that step now raises an instant notice spelling out exactly what changed:
This model doesn't support 5s — switched to 4s
Resolution works the same way. The notice only appears once you've actually picked a tier — a freshly created node is just taking its defaults, so it stays quiet.
In short: after switching models you no longer have to re-check the panel line by line. If something moved, you'll know right away.
Full-screen voice panel, and prompts that fit your tablet
The voice panel expands to full screen. Click the expand icon in the panel's top-right and the input area grows, so a long script fits on one view — image, video and text panels already worked this way, and voice now matches.
Tablets and narrow windows: the prompt panel fits itself to the screen. It sizes to the available width and always stays fully in view — even when the node sits at the very left or right edge of your canvas, at any zoom level.
Nothing changes on a wide screen; it works exactly as it did.
Canvas videos: mute, download the master, enlarge
Select a video node — three buttons sit in the top-right of the frame.
- Download the master — you get the master file, ready to edit, publish or hand off. The version-history dialog downloads it too, so you can grab an older take without restoring it to the node first.
- Mute — sound on or off in one click.
- Enlarge — it opens right on the canvas, no fullscreen jump, no leaving the layout you're working on.
Double-click to enlarge: images and videos alike — double-click the frame, no need to find a button.
Node toolbars are back to a single always-visible row — select a node and every action is right there.
Batch video generation now runs in parallel
Submit several videos at once and they now all start together — no more waiting for the ones ahead.
Batches
Generate a batch of shots together, or run a few variants of the same prompt: they all start at once. Up to 12 run in parallel.
Everything keeps its own pace
The last step of a video render converts the result into a version that plays smoothly on the canvas. That step no longer holds up images or audio running at the same time — each goes at its own pace.
Unchanged
How long a single video takes still depends on the model itself. What changed is how many can run at once.
Director Desk is live: block the shot before you shoot it
A clapperboard button now sits in the canvas toolbar on the left. Click it and you get a director-desk node — open it and you're standing in a 3D stage.
What you can do in there
Block your actors — add figures, drag them around, turn them. Height, build and body type are adjustable. Props (tables, crates, railings) can be dropped in as stand-ins.
Place the camera — free-fly the main camera, switch focal lengths from 18mm to 135mm, and watch the live viewfinder in the corner: what you see is the shot. Happy with it? Save it as a recorded shot.
Pose them — 93 pose presets, from stand/sit/walk/run to sword, aiming, spellcasting, zombies and farm labour, grouped and labelled. Picked one? You can then tweak it joint by joint — 17 in all (head, neck, chest, spine, pelvis, plus upper arm / forearm / hand and thigh / calf / foot on both sides), each rotatable on its own, with a dot marking the ones you've touched.
Save the poses you like — up to 6, stored on your account (they follow you to another machine). One click brings a pose back, on this actor or any other. Double-click a slot to rename it.
Draw the walk — set a start and an end for any actor, add waypoints in between to route around obstacles. Waypoints are draggable right in the 3D stage. Path shape can be polyline, smooth or step, and the easing is adjustable. Hit play and the actor walks it, facing where they're going.
What you get out of it
Three things drop straight back onto the canvas:
- Clean reference still — the bare frame, to feed an image node as composition reference
- Annotated reference still — with actor labels, facing arrows and the axis line, for you or a collaborator to read
- Motion reference video — the walk rendered frame by frame to MP4, to feed a video model as a motion cue
Each of them becomes a canvas node with one click. Wire it up and keep going.
In one line
Until now "how is this shot framed" lived in your head or on a napkin. Now you can block it in 3D first, then have the model generate against that exact framing and that exact movement.
The agent brings a storyboarding playbook to shot breakdowns
Shot breakdowns
When you ask the canvas agent to break a scene into shots, it now brings a storyboarding playbook:
- Camera values — what focal length, height and angle this shot wants
- Composition hazards — which framings fall apart most often in the render
- Cross-shot continuity — what has to carry over from one shot to the next
Transformation and suit-up shots
For armour plates flying into place, suit-ups and form changes, five new rules:
- Two similar-looking pieces of gear: give each a semantic name, and state explicitly that B's features do not belong on A
- Don't write the whole transformation as one shot — split it into fast cuts, each showing one local area
- A time code on the block isn't enough; inside the shot, spell out what happens in each second
- Naming a movie doesn't give you motion — write how each plate flies, lands and locks
- Don't write a shot shorter than 1 second; it gets stretched and eats the shots after it
What you need to do
Nothing. It's already live.
The canvas agent now writes video prompts as flowing prose
When you ask the canvas agent to write a Seedance prompt, its output looks different now.
- Bracketed blocks:
[GLOBAL][SCENE][CAMERA], filled in slot by slot - Flowing prose: one continuous passage that still carries everything those slots used to
Why prose holds up better
Filling slots slides very easily into adjectives — drop one word per slot and it looks finished. But adjectives don't produce pixels:
- "tense atmosphere" doesn't make the shot tense
- "a magenta light" can come back orange-red — the model slides toward whatever prior sits closest
Prose forces you to finish the sentence. Same lighting example, written as three things: where the light comes from, what it lands on, and what the skin looks like once it does. Write all three and the colour holds.
What you need to do
Nothing. It's already live.
The old format is still there
The bracketed template hasn't been deleted — it's kept as a compatibility format.
- Bracketed blocks:
Wan 2.7 now supports end frames: give it a first and a last image
When generating with Wan 2.7, the start/end frame mode now takes two images:
- The first sets how the video opens
- The second sets how it ends
The model works out the transition in between. When you need a shot to travel from one definite state to another, this is far more precise than giving a first frame and describing the rest in words — the same character going from seated to standing, a camera pulling from close-up to wide, daylight turning to night.
How to use it
In a video node pick Wan 2.7 → switch to start/end frames → drop an image into each slot. Using only the first one still works; that's ordinary image-to-video.
Two notes
- Keep the two images at similar dimensions — a big mismatch tends to warp the transition
- An end frame alone won't work; the first frame is required
HappyHorse 1.1 is here — 9 aspect ratios, up to 15s, audio included
A new model has landed in the video node's model list: HappyHorse 1.1, Alibaba's in-house video model.
What it does
- Text to video — write a prompt, get a clip
- Image to video — hand it a first frame and let it move
- Multi-image reference — attach up to 9 reference images to lock character and object consistency
Things worth knowing
- 9 aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4, 4:5, 5:4, 9:21, 21:9 — including 21:9 ultra-wide and 9:21 ultra-tall, which most models don't offer
- Any length from 3 to 15 seconds, not just a few fixed steps
- Audio comes built in — every clip ships with synced sound, no toggle, no surcharge
- 720p / 1080p
Pricing
16.8 credits/sec at 720p, 21.6 credits/sec at 1080p. A 5-second 720p clip costs 84 credits.
Two notes
- In image-to-video mode the output ratio follows your input image, so the aspect selector doesn't apply — that's the model's own rule
- This version doesn't do video editing; it takes image references only, not video
Batch download: review the list, then name every file
Select several assets on the canvas and hit Download — instead of packing immediately, you now get a list.
Decide the names before anything is packed
Rename any single file, or apply a rule to the whole batch:
- Index + name (default), name + index, name only, index only, type + index, prompt excerpt
- Prefix, separator (
_-space, none), start number, digits, append timestamp - Hit Apply to all when you want the rule to overwrite the ones you edited by hand too
While you type a name, the actual filename is shown underneath — the suffix added for duplicates, the characters your filesystem won't take — so you see it before downloading, not after unzipping.
Order, exclusions, zip name
- Drag the handle on the left to reorder; numbering follows
- Click × to exclude a file; the footer has Restore if you didn't mean to
- Name the zip yourself
Hover a thumbnail to see it large, so you can check which one you're renaming.
Filenames follow your node titles
Single downloads and batch downloads both give you the title you see on the canvas: your own title if you renamed the node, otherwise the node's own name; uploaded files keep their original filename.
Prefer naming by prompt? It's still there — pick Prompt excerpt as the naming rule.
Seedance 2.5 now outputs 1080p
Seedance 2.5 now has three resolution tiers: 480p · 720p · 1080p.
- All six aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4, 21:9
- Reference images, reference video and audio all work at 1080p too
- Available from the CLI and from Claude / Cursor via MCP
Pricing
1080p costs 2.46× what 720p does. For example, 5 seconds at 16:9 is about 356 Xins (the same clip at 720p is about 145). The price updates live as you switch tiers, so you see it before you run.
Two things worth knowing first
The format is 10-bit HEVC. Finer detail than 8-bit, but some players and editors can't open it — the built-in Windows player and older editing software may show a black screen or refuse the file outright. VLC and QuickTime play it fine; for editing, transcode to H.264 first.
Generation time varies a lot. The same 1080p job can take a few minutes or twenty-odd, depending mostly on upstream queue depth rather than how many seconds you asked for. Jobs run in the background — close the tab and come back.
Canvas Agent is now open to everyone — no request needed
Canvas Agent is no longer in closed beta. Every signed-in account can use it — including brand-new ones.
Getting started
Open any canvas; the Agent panel is on the right. Just say what you want:
- "Break this script into 6 shots and generate an image for each"
- "Make a character sheet for this character, four angles"
- "Look at these clips on the canvas and flag the ones where the face drifts"
It lays out a plan for you to look over before it starts, and the work runs in the background — close the browser and pick it up later.
What it can drive
Generating images, video and audio; editing images; looking at images and video; reading the canvas; creating nodes and edges; tidying the layout — most of what you can do by hand on a canvas, you can hand to it instead.
You pick its brain model yourself (Claude, GPT, Gemini, DeepSeek, Kimi and others) from the top of the Agent panel.
Spending stays visible
Every step the Agent takes is itemised in your billing details, there to check any time. If your balance runs short it stops and tells you, rather than pushing on regardless.
Node tools: hover to open, laid out as a grid
- Hover to open — no click needed. With a node selected, move your cursor onto the dot at its top-right and the tools open on their own. Move away and it closes by itself, or click the ← at the top left.
- The dot grows into the panel rather than popping a separate window beside it — your cursor never has to cross a gap to reach the menu.
- Tools are laid out as a grid, so the panel is about half its old size and every tool is a shorter trip away.
- Tool names appear at the top of the panel: whichever icon you rest on, its name is spelled out up top.
- Touch devices still open by tapping; Apple Pencil hover works too — on the same iPad, finger taps and pencil hover each do the right thing.
- Clicks are ignored until the open animation finishes, so you can't trigger a tool while your hand is still on its way over.
Zoom out to 5% — see the whole canvas at once
- Minimum zoom goes from 25% down to 5%. On a canvas with many nodes you can now take in the whole layout in one screen instead of panning around.
- Below 14% it switches to a minimal overview: nodes become plain blocks, leaving just structure and position. A line at the bottom of the screen tells you how to get back — zoom in and node content returns exactly as it was.
- Animations now stop as you zoom out. At 20% and below, generating effects, selection glow and edge flow all stop; which nodes are still generating stays visible as a static marker, so nothing is lost. The further out you go, the calmer the canvas — and your machine stops rendering animation you can't see anyway.
- The zoom ruler gained ticks for the low end and now marks where the minimal overview begins, so you can see the boundary before you hit it.
- Also fixed a grey haze the background dot grid produced at very small zoom levels.
Batch download now keeps your node titles
- Batch downloads now use your node titles. Select a batch, download once, and the files come out named like
Xiaoyu_Episode4_Outfit.png— no more guessing which random string is which. - Nodes you never renamed no longer share one name — the start of the prompt becomes the filename, e.g.
A girl in white before the snow mountain.png. - Real duplicates get a counter (
_2,_3), so nothing overwrites anything else in the same batch. - Markdown exports from text nodes follow the same naming.
- Batch downloads now use your node titles. Select a batch, download once, and the files come out named like
Subtitle removal paused · Common actions moved to the top
Subtitle removal is paused
Video subtitle removal relies on an external service that is currently down. Rather than let you start a job, wait, and get a failure back, we have taken the feature offline for now.
It is hidden from the video node's tool list, and the canvas agent will no longer call it either. It will return once the service is fixed — we will post here when it does.
Videos you have already generated are unaffected.
Tool list: the everyday actions moved to the top
Download, Set as cover, Expand and Info now sit at the top of the list. These are the "take the result away" actions you reach for most often, and they used to sit below a stack of editing tools.
Everything else keeps its order.
Node tools now live behind one small button
The toolbar is now a single small dot
Selecting a node used to park a wide toolbar right above it. On image nodes that's up to 9 buttons — about 380px wide, while a portrait node is only 280px. The toolbar was wider than the picture.
Now it collapses into a small dot at the node's top-right corner. Click it and the tools open as a labelled list beside the node, clear of the image. It closes again once you pick something.
No more guessing what an icon means
The old toolbar was icons only — you had to hover to find out what each one did. Every entry now carries its name: Redraw, Erase, Enhance, Outpaint, Remove background, Multi-angle, Panorama, Lighting, Annotate, Crop, Grid split, Info, Download, Set as cover, Expand.
Tools that were buried under "More" are one click away
Outpaint, remove background and multi-angle used to sit behind a second "More" menu. They're in the main list now.
Image, video, audio and document nodes all work this way
Text nodes keep their formatting bar as is — bold, italic, headings and lists need to stay one click away while you are writing.
Zooming never puts it over your artwork
The dot keeps the same size at any zoom level, and its distance from the node scales with the node, so it never ends up on top of your image when you zoom out.
A third desk pet: Xiaoyu, as a figure
There are three canvas pets now.
Xiaoyu
The new one is Xiaoyu, styled as a collectible figure — one plastic material, one lighting setup, oversized white hoodie and black shorts.
She has the same full animation set as the other two: idle, blink, glance around, wave, beckon, cheer, slump, being picked up, landing back down — plus a heads-down typing pose for when work is in progress.
How to switch
Right-click the pet on your canvas → Switch pet, then click the one you want. The swap is instant and remembered per device.
Gugu Gaga and Doro are both still there, and the default is unchanged.
Switching models now keeps your settings
Switching models no longer resets your settings
Changing the model on a video node used to snap resolution straight back to that model's default — your carefully chosen 4K or 1080P, gone. The rule now:
- The new model also has your current tier → kept exactly as is
- You picked 480P, the new model starts at 720P → lands on 720P
- You picked 4K, the new model tops out at 1080P → lands on 1080P
Duration works the same way. Your prompt and the audio toggle never change.
Portrait stays portrait
When the new model does not support your aspect ratio, it used to become 16:9 every time. Now it snaps to the closest shape instead: 3:4 lands on 9:16, still portrait.
MiniMax H3 keeps 768P
The cheaper 768P tier that shipped today used to jump back to 2K (40% more) if you switched away and back. It stays put now.
Grok Imagine 1.5 keeps 1080P
Switching to it used to force 720P with no way back. Text-to-video and image-to-video now both hold 1080P.
The panel only shows options that actually do something
The "Adaptive" aspect ratio is gone for Gemini Omni and Grok Imagine 1.5 — both always render 16:9, so that button never had any effect.
MiniMax H3: a cheaper 768P tier, and reference videos now work
A 768P tier, 29% cheaper than 2K
H3 now has a 768P resolution option. 50 Xins for 4s (2K is 70), 100 for 8s (2K is 140). Draft at 768P, finish at 2K.
Reference videos and audio now work
H3's reference tray takes three kinds of material: up to 9 images, 3 videos, and 3 audio clips, 12 files total.
Editing needs no separate mode
H3 has no separate "edit" entry point — attach a reference video and just say what to change in the prompt. Attach a clip with Chinese subtitles, write "replace the text on screen with English", and the layout and footage carry over.
Clip lengths are checked up front
Reference videos and audio run 2–15 seconds each, 15 seconds combined. If a clip is over, you hear about it before you're charged — not after the render.
Seedance 2.5 is live: 30-second clips you can edit and extend
Seedance 2.5 is now available on the canvas.
Up to 30 seconds per clip
The Seedance 2.0 family caps at 15 seconds; 2.5 doubles that to 30 seconds. It supports 480p and 720p, with up to 30 reference images.
Video editing: change what you don't like, in place
Video nodes now have an Edit video button in the toolbar. It opens a player — step to the exact frame you want to change, draw on it to point at the spot, then describe the change.
Every frame you capture drops a timestamp tag into the prompt automatically — no counting seconds by hand:
00:02 change the outfit to blue 00:03 switch the background to a night bar scene 00:09 make the hair dark blue with highlightsVideo extension: add to either end
Two new entries on the panel:
- Extend backwards — generates what happened before the clip begins
- Extend forwards — continues the story after it ends
You pick the output length, 4 to 30 seconds. The output ratio follows the source video.
How reference videos are priced
Correction (2026-08-10): this section previously said reference videos were often cheaper and showed a "32% off" table. That was our miscalculation. They are not cheaper — the correct explanation is below.
Seedance 2.5 charges a lower rate when your request includes video input. But video-input requests also carry a minimum billing amount, and the two cancel each other out almost exactly.
The result: attaching a reference video costs about the same as text-only generation, and only becomes more expensive once the reference runs longer than two-thirds of your output.
For a 30-second 720p output:
Credits Text-only 867 With a 4s reference 865 (same) With a 10s reference 865 (same) With a 20s reference 865 (same) With a 25s reference 951 (10% more) So attach a reference when your work needs one — there is no need to keep clips short just to save credits. The credit figure on the panel is live, so you can see the exact price before you generate.
Collapsed nodes now show their settings
- MiniMax H3 now shows its resolution ("2K") when collapsed — every other model already did; H3's slot was blank.
- More settings on the collapsed bar: Seed Audio's volume, pitch and sample rate; ElevenLabs' stability; Grok Imagine's "Pro" mode.
- A chip only appears once you change a default — a node you haven't touched stays exactly as wide as before.
- Video-edit resolution is visible too: 720p and 1080p differ by 2× in price, so you can now see which one you picked without expanding.
- When Kling has a reference image, the output ratio follows that image and the panel doesn't let you change it — so the collapsed bar no longer shows a ratio that does nothing.
- The "generate audio" toggle is gone for Gemini Omni and Grok Imagine 1.5: audio for those two is decided upstream, so the switch had no effect.
- Z-Image's "Advanced settings" (negative prompt, steps, seed) have been removed: we measured that the upstream simply doesn't read them, so changing them did nothing — better no knob than a knob that isn't wired to anything.
MiniMax H3: a 4-second option, and room for more references
- A 4-second length — 20% cheaper than 5s, handy when you are still testing a prompt.
- More reference material: up to 9 images, 3 videos and 3 audio clips, 12 files in total (was 9 combined). Mix them freely.
- Go over the limit and you get told right away, with nothing charged — no more quietly dropping the extras.
Both DeepSeek brains are now 50–63% cheaper
The new prices
In the canvas agent's brain picker, both DeepSeek tiers (per million tokens):
Brain Input Output DeepSeek V4 Flash 30 → 12 Xins 60 → 23 Xins DeepSeek V4 Pro 109 → 55 Xins 218 → 109 Xins Already live — nothing to configure. New conversations bill at the new rate.
Who can use them
- DeepSeek V4 Flash is available on the free tier. It is now the cheapest brain in the picker, with output an order of magnitude cheaper than the other free brain. For everyday work — editing the canvas, creating nodes in bulk, setting parameters — it is plenty.
- DeepSeek V4 Pro is on the paid tier, and worth switching to for long context (1M) and harder reasoning.
Flash also moved to the official July 31 release in the same update.
Switching brains
Open the agent panel (click the desk pet at the bottom-right of your canvas, or press ⌘I / Ctrl I). The brain picker is at the top, and each brain shows its input/output price right underneath.
Reference materials, reworked: a tray, numbering, and video refs on Wan 2.7 & MiniMax H3
The video, image and audio panels now share one layout, with reference materials collected in a tray above the prompt.
Thumbnails carry numbers, and the number is the @ in your prompt. The third image shows a 3; writing
@image3in the prompt points at exactly that one. Drag a thumbnail to reorder and the numbers follow.Past twelve, they fold. The last slot becomes a counter button; opening it reveals a panel to the right holding everything, split into image / video / audio sections, with drag-to-reorder inside. Prompt and thumbnails stay on screen together.
Wan 2.7 and MiniMax H3 now take video references. Attach video nodes in reference mode — up to 3 clips. MiniMax H3 additionally accepts images, video and audio at once.
Seed Audio reference voices are now picked from the canvas. Hit the reference button and select an existing audio node instead of uploading the file again.
Each model's reference-image limit is now honoured by the panel, so nothing sits in the tray that the model would never actually use.
New model: MiniMax H3
- MiniMax H3 joins the video models: text-to-video, image-to-video (with optional last frame), and multimodal reference (image + video + audio). 2K, 5–15s, native audio, 17.5 Xins/sec. See the announcement.
Grok Imagine 1.5: up to 7 reference images, and text-to-video
Multi-image reference: up to 7 at once
Put the person, the product and the look into separate reference images, then name each one in your prompt with
<IMAGE_0>,<IMAGE_1>:Handheld UGC-style clip. The person from
<IMAGE_0>holds the skincare jar from<IMAGE_1>and talks to camera, casual phone-camera framing.Images map in wiring order — the first is
<IMAGE_0>, the second<IMAGE_1>, and so on.Name every image you attach. An image you pass but never mention gets ignored, or blends into the shot unpredictably.
Good for: one character across a whole series, talking-head product clips with the real product in frame, or borrowing the palette and texture of one image for a new scene.
Runs from text alone
Write a prompt and go — no starting frame required. Attach an image and it animates from that frame; attach nothing and it generates from the text.
1080p added
Text-to-video and image-to-video now support 1080p. Multi-image reference mode tops out at 720p.
Lower price per second
Quality Before Now 480p 11.2 Xins/s 10 Xins/s 720p 19.6 Xins/s 17.5 Xins/s 1080p — 31.25 Xins/s Already live, nothing to switch on.
It comes with sound
Picture, lip-synced dialogue and ambient effects are generated together in one pass, not dubbed afterwards. Any whole number of seconds from 1 to 15, seven aspect ratios.
Let your AI assistant tidy your canvas directly
- Name your nodes, then actually find them. Your AI assistant can now set node titles directly. Once a canvas is titled, you can search it by name instead of scrolling through hundreds of nodes.
- Wiring, cleanup and parameter changes too. It can link nodes as references, remove leftover draft nodes, and switch a node's model / aspect ratio / resolution — parameter changes are saved as a draft only: nothing is generated and no Xins are spent until you run it.
- Your finished work is protected by default. Nodes that already hold a result, and nodes still generating, cannot be deleted or edited unless you explicitly say so — "tidy up my canvas" will never quietly take your paid output with it.
- Prompt text still lives in the canvas editor. Rich-text prompts and
@图片Nreferences are position-linked, so that stays where it's handled properly. - Applies to personal canvases; team canvases are collaborative and are still edited in the UI.
To get it: MCP users restart the client; CLI users run
npm i -g @xinyuai/cli@latest.The canvas agent is now available to everyone
Now open to everyone
Open any canvas and click the desk pet in the bottom-right corner, or press ⌘I (Ctrl I on Windows), and the agent panel appears.
What it does: read what is already on your canvas, create generation nodes on request, set the model and parameters, and wire up reference images. Once the nodes are prepared, you press “Generate all” to actually start them — whether and when anything is spent stays your call.
Two brains are free
- Gemini 3.5 Flash
- DeepSeek V4 Flash
Both are available without paying. Ten stronger brains (Claude Opus 5 / Sonnet 5, Gemini 3.1 Pro, the GPT-5.6 family, Kimi K3, DeepSeek V4 Pro and more) are on the paid tier and are marked as such in the picker.
It tells you the truth about what it did
- While nodes are still generating, the task card reads “Media still generating · N still running”, and only says complete once they all finish.
- A plan step that could not be ticked automatically is marked ❓ “not auto-confirmed”, with a note that the work may well have been done — we just could not confirm it automatically. No guessing.
- Image models do not all offer the same quality tiers; the agent only picks from the tiers a given model actually supports, so the tier and price on the confirm card are exactly what runs and what gets charged.
Tell us when it is wrong
Every reply has a 👎 and a ⚠️ button. If a run goes sideways, or it does something you did not ask for, press one — that is the only way we can see which specific run went wrong.
Canvas agent: an unticked step now tells you what actually happened
The canvas agent is in whitelisted beta — these changes are visible only on accounts that have it enabled.
An unticked step no longer leaves you guessing
Each step in the plan is ticked automatically as it runs. The tick is matched by the tool name written on that step — so if the agent reaches the same outcome with a different tool, that step never gets ticked.
Previously it just sat there blank, and you could not tell whether the work had been skipped or done without being ticked.
Such a step is now clearly marked ❓ “not auto-confirmed”, with a line under the plan: it may well have been done — we just could not confirm it automatically, so the result and the canvas are the better check.
Why not simply say “not run”
Because that would usually be wrong. What normally happened is that the agent used a different tool to achieve the same thing. Saying “not run” would send you off to redo work that is already finished.
We state only what we actually know: this step could not be auto-confirmed.
Also
Failed or stopped runs are unaffected — there, a step that never ran is already explained by the run itself.
Canvas agent: the task card now waits for your media to finish
The canvas agent is in whitelisted beta — these changes are visible only on accounts that have it enabled.
The task card waits for your media
The agent prepares the nodes; what actually starts them is you hitting Generate all.
During that window the card now reads “Media still generating · N still running”, and only switches to “Task complete” once every node from this round has finished. Same when a run fails or is stopped — as long as something is still running, the card will not claim it is done.
Only this round counts
The count covers only the nodes this round of the agent created. Anything you started by hand elsewhere on the canvas is excluded, so the number always means “what is left in this task”.
Stopping one of them
There is deliberately no stop button on the card: these are canvas generations, so stop them on the node itself.
Canvas agent: a steadier upstream, and specs that always match the price
The canvas agent is in whitelisted beta — these changes are visible only on accounts that have it enabled.
A steadier upstream
Every brain model, plus the agent’s ability to look at images, watch video and listen to audio, moved to a different upstream.
Nothing to change on your side: model picker, past conversations and nodes you already created all behave the same.
Specs that always match the price
Image models do not all offer the same quality tiers — SeeDream 5.0 Pro goes up to 2K, Nano Banana 2 reaches 4K.
The agent now picks only from the tiers a given model actually supports. If it ever asks for one that is not offered, it is stopped right there, told which tiers exist, and picks again.
Which means: the tier and the price on the confirm card are exactly what runs and exactly what gets charged.
Also
- Thinking traces are not surfaced for the Gemini brains in this build.
- Nano Banana 2 no longer lists a 512 tier (it had no price, and was not selectable anyway).
A second desk pet: meet Doro, and you can switch
Your canvas pet is no longer the only one.
Switching pets
Right-click the pet on your canvas → Switch pet, then click the one you want in the popup.
You pick by looking at it, not by reading a name. The swap is instant and remembered per device, so it is still there next time you open the canvas.
Doro
The new one is called Doro. Same full set as Gugu Gaga: idle, busy, being picked up, waving, looking around, cheering, slumping, blinking — the whole thing, not a re-skin.
Everything else is unchanged
Drag it anywhere, let it snap to the right edge, position remembered per device, right-click to quiet it down or tuck it away — all identical across both pets.
How it was made
Same pipeline as Gugu Gaga: an AI-generated green-screen video, split into frames, keyed out, packed into a sprite sheet — 226 frames. Made with the very models you can run on your own canvas.
Shared canvases: what you put there stays there
This release is all about shared (team) canvases. Personal canvases are unaffected.
Uploads stay on the canvas
Drop an asset onto a shared canvas and it stays there.
A teammate's recent edits survive your page load
On a shared canvas, edits made in the last few seconds could previously be rolled back when someone else opened the same canvas. Not anymore — the newest edit always wins, and deleted nodes stay deleted.
Retry clears the error message
After a failed generation, hitting retry now clears the red notice on the node instead of leaving it stuck there.
Storyboard parsing works on shared canvases
And not just for the canvas creator — invited members can run storyboard parsing too, and the result lands on the canvas and syncs to everyone live. Writing the result only touches the fields it should: your teammate's freshly edited dialogue, labels, and model settings are left alone.
Undo stays on the canvas you're looking at
Switch to another canvas and press ⌘Z / Ctrl+Z — it no longer drags nodes over from the previous one.
Under the hood, every edit on a shared canvas now goes through the same realtime channel. There is no longer any server path that writes to the database while bypassing realtime sync. That is the real change underneath this release.
Gugu Gaga has moved onto your canvas
That round button in the corner of your canvas is now a penguin.
It watches your jobs for you
Median generation takes 91 seconds, and 70% run longer than a minute — which means you're usually off doing something else.
Now when a job lands, or a node fails, Gugu Gaga speaks up. No more clicking through nodes to check. It also wears a count badge: green for how many are running, red for how many failed.
Put it wherever suits you
Drag it anywhere on the canvas. Let go near the right edge and it tucks itself in; drop it in the middle and it stays in the middle. Its spot is remembered per device.
When the side panel opens it steps out of the way, and when you close the panel it comes back — without losing the spot you chose.
Hover to hear from it
It reports what's actually happening on your canvas right now: how many running, how many failed, or all clear. When nothing's going on, it says other things.
When you'd rather not have it around
Right-click it:
- Quiet mode — it stops speaking up, but stays put and keeps the badge
- Reset position — snaps it back to the corner
- Hide desk pet — tucks it away
Once hidden, bring it back any time from the paw button in the toolbar at the bottom-left of the canvas.
How it was made
One AI-generated green-screen clip, cut into frames, keyed, and packed into a sprite sheet — 10 actions, 222 frames. Made with the same models you have on your canvas.
MCP / CLI: ask the price first, and stop guessing what a model can do
- Ask the price before you spend. New estimate in MCP/CLI:
xinyu_estimate_price/xinyu price --model … --size 2Ktells you what an image will cost in Xins. It runs the very same logic that charges you, so it is the amount you will actually be billed — not a rough guess. - Your AI assistant now knows what each model can really do. The model catalog now states whether a model can edit images and how many reference images it accepts (8–16, depending on the model). Several models that do support editing were not marked as such, so assistants were telling you they couldn't.
- The model must be named — no more silent default. Generating through MCP without specifying a model used to run on a default model and bill you for it. Now it asks you to name one, and nothing is charged.
- Recover a slow job in one call — and never pay twice. New
xinyu_wait_for_job/xinyu job wait <id>blocks until the job finishes. Job lists can be filtered by canvas, so a job is findable even if you lost its id. Timeout messages now say plainly: you already paid, do not generate again. - One more way to look at an image. Added
qwen3.7-plus(images only) — it can read some images other models refuse.
To get it: MCP users just restart the client; CLI users run
npm i -g @xinyuai/cli@latest.- Ask the price before you spend. New estimate in MCP/CLI:
MCP / CLI: the model you pick now always takes effect
- The model you pick in MCP / CLI now always takes effect. Previously, if a parameter was written as
model(the correct name ismodel_id), it was silently ignored — the job ran on the default model and was billed normally. You asked for Seedream 5.0 Pro and could get Nano Banana 2 instead. That can no longer happen. - Common spellings are accepted too.
model/size/quality/scaleare now mapped to the right parameter automatically, so you don't have to check the docs for exact names. - Typos fail loudly. An unrecognised parameter name now returns a clear error naming the right one, instead of spending your Xins on a result you didn't ask for.
To get it: MCP users just restart the client (Claude Desktop / Cursor / Windsurf / Cline); CLI users run
npm i -g @xinyuai/cli@latest.- The model you pick in MCP / CLI now always takes effect. Previously, if a parameter was written as
Canvases open faster — the canvas appears first, images follow
- The canvas shows first, images follow. Opening a canvas no longer waits for every thumbnail to finish downloading — the canvas appears right away and images fill in in the background. The more nodes you have and the slower your connection, the bigger the difference.
- Canvas media is cached long-term. Images on your canvas are now kept in your browser's cache, so reopening the same canvas barely re-downloads anything.
- On a flaky connection you won't be left staring at a loading bar anymore.
Canvas Agent: new model picker, Kimi K3, cheaper Opus 5
Canvas Agent: redesigned model picker
The flat list stopped scaling once the lineup grew. It's now one panel:
- Search — type
claude,chatgpt, oranthropicand it matches - Featured — one pick each for free, cheapest, balanced, newest, strongest
- Grouped by brand — Gemini / DeepSeek / Claude / ChatGPT / Kimi
The selected row now has a highlight, and long model names are no longer clipped.
New brain: Kimi K3
Moonshot's flagship — 1M context, always-on reasoning with the thinking visible. Good for long-chain tasks.
Claude Opus 5 got cheaper
We moved it to an integration that supports prompt caching. Context you resend every turn (canvas snapshot, tool definitions, mounted skills) is now billed at the cached rate, so the same conversation costs noticeably fewer credits.
The list price didn't change — the savings go straight to your bill.
GPT-5.6 pricing increase
Luna / Terra / Sol are up about 14%. Our upstream's limited-time discount no longer covers cost at the old price.
They remain below typical pricing for comparable models, and we'll lower them again as soon as the discount returns.
The Canvas Agent is currently in closed beta; the agent-related updates above are visible to beta users only.
- Search — type
Clearer top-ups: membership bonus stated up front, with personalized savings hints
Two small changes, both about getting more from your top-ups:
- The membership top-up bonus is now clearly stated. PRO members get +10% on every top-up and Ultimate members +20% — this has always been credited, but it was previously a footnote on the top-up page. It's now a prominent callout.
- The top-up panel now gives personalized advice. If your top-up volume over the last 30 days would make a membership worthwhile, the panel shows exactly how many extra Xin Points a year you'd get for the same spend. If it wouldn't pay off, nothing is shown.
We also unified the credit unit name across the app: it's Xin Points everywhere now.
No more spinners between pages — Dashboard, Templates, Assets & Notifications open instantly
Starting today, the four pages you visit most — Dashboard, Templates, Assets and Notifications — no longer show a spinner every time you switch back to them:
- Switch away and back — content appears instantly. Page data is cached locally and rendered immediately; fresh data loads silently in the background and updates on screen when ready.
- The notification list remembers where you were. Pages you scrolled through stay loaded when you come back — no more starting over from page one.
- Favoriting assets is instant. Tap the heart and it takes effect immediately, no round-trip wait; if the request fails, it reverts automatically.
The farther you are from our servers, the bigger the difference — each page switch used to cost a full network round-trip (typically 0.5–1.5s from mainland China). That wait is gone.
One deliberate exception: pages involving money (balance, billing) are never cached — we'd rather make you wait a moment than show a stale number.
One more small update: for canvas Agent users (currently in closed beta), the quick-start shortcuts on the Agent panel are now Storyboard / Character sheet / Script / Read canvas — each backed by a real skill, ready to run in one tap.
Teach the Agent — and see what stuck
You can now teach the Agent — and see what stuck.
- See what it remembered, and remove any of it. A new "Remembered habits" entry in the Agent menu lists every standing preference you have taught it, each removable. They apply automatically on every turn, so you never repeat yourself — they last across conversations in this canvas, and apply only to you; other members' Agents are unaffected.
- Teaching it says so, right there. Say "remember: always lock the face with this portrait" and the run shows "Remembered · always lock the face with this portrait". If it already knew, it says so — no more guessing whether it took.
- The composer and the empty state tell you the feature exists. Just say "remember…".
- No more wall of cryptic red crosses. Internal steps like "load the spec first, then retry" now fold away, with the successful step marked "retried N×". Real failures still show — in a plain sentence instead of an error code.
The canvas Agent is still in closed beta — accounts on the list will see this after the update.
The canvas Agent shows its plan before it starts
When you give the canvas Agent a task, it now lays out the steps it intends to take in the sidebar, then works through them in front of you.
- The plan comes first. No more guessing what a stream of tool calls is doing — each step is named, and ticks off as it completes.
- Deleting or overwriting waits for your OK. If the plan contains a step that deletes a node or overwrites finished work, the Agent stops and asks before touching anything.
- Cross out steps before you confirm. Don't want a particular step? Cross it out. The server refuses to run it — this isn't a hint to the Agent, it's enforced.
- Changes of mind are stated. If the Agent rethinks its approach mid-task, the list is marked "Plan adjusted" rather than quietly becoming something else.
- Longer tasks keep more of their context. Multi-step work picked up across sessions carries the earlier steps with it.
The canvas Agent is still in closed beta — accounts on the list will see this after the update.
Claude Opus 5 is here — Anthropic's newest flagship
Claude Opus 5 — Anthropic's newest flagship model — is now available.
- Where — pick it in the canvas text node's model list, or set it as your agent's brain.
- Price — same as Claude Opus 4.8. No premium.
- Good at — long-form writing, complex reasoning, tasks that need careful step-by-step work.
Opus 4.8 stays available too — pick whichever fits.
Your canvas can have a background now — image or video
You can now give your canvas its own background, so different projects are recognisable at a glance.
-
Pick straight from the canvas — the panel lists every image and video on your current canvas; one click sets it as the background. Filter by All / Images / Videos.
-
Video backgrounds work too — muted and looping. It pauses automatically while you pan or zoom, so it never competes with the canvas for performance.
-
Tune it — opacity and blur sliders, kept subtle by default so your nodes stay readable. The dot grid is adjustable too — turn it all the way down for a clean surface.
-
Set once, applies to every canvas — it lives in your own browser, so collaborators aren't affected: everyone can have their own background on the same canvas. You'll need to set it again on a different device or browser.
Look for the gear icon at the bottom-left of the canvas.
-
Steadier generation: no stuck buttons, no duplicate jobs on weak networks
This release focuses on stability during generation — all of these show up in everyday use:
-
The send button now reflects the real job state. It becomes clickable again as soon as the job finishes, and panning or zooming the canvas mid-generation no longer affects it. Jobs still running after a page refresh, and generations started by a collaborator, now correctly show as in progress too.
-
Repeated clicks on a weak network no longer create duplicate jobs. If the page feels unresponsive and you click send a few times, only one job is created.
-
A dropped connection is no longer treated as a failed generation. Losing the connection only means progress is temporarily invisible — the job keeps running on the server. The node is no longer marked failed, and any existing result is kept.
-
Moving the canvas no longer interrupts text generation. Panning or zooming while a prompt node was generating used to cut it off. It no longer does.
-
Cancel in the Storyboard panel really cancels. It now stops the job on the server and stops billing for it.
-
The multi-select "Arrange" menu is now translated. It no longer falls back to Chinese in the English UI, and newly created "Playlist" / "Merged Video" nodes follow your interface language.
-
Canvases open far faster, especially on limited connections
- Opening a canvas now uses dramatically less data. Previously, canvases with many nodes (especially images and audio) quietly pulled a large amount of original files in the background. Now only what's actually displayed gets loaded — data usage can drop to a fraction of what it was, so canvases open faster and lighter.
- The improvement is most noticeable on mobile data, limited connections, or cross-border access — opening a large canvas could previously cost hundreds of MB; now it takes just a few.
- No settings needed; already active.
Tidy layout now follows what you see; results visible without opening the panel
- Tidy layout now orders nodes the way you see them on canvas. Previously, tidying a batch of large images (3K portraits, say) could split visually-aligned images into different rows and leave things messier than before. Row detection now scales with image height, so it works for large and small nodes alike.
- The agent launcher shows a result badge: how many are generating, how many failed — visible with the panel closed, and it survives a page refresh. Before, collapsing the panel to look at the canvas meant no completion signal at all.
Agent features are in closed beta and rolling out gradually. Tidy layout is live for everyone.
Sharper image reading, and nodes land where you are working
- The agent now reads images with a stronger vision model: more detail from the same picture (materials, secondary light sources, small objects in frame) and less of what was never there. Video analysis is upgraded too.
- Nodes the agent generates from text now land where you are currently looking, instead of at the far edge of the canvas.
- Selecting an image no longer triggers a stray "2×2 or 16:9 character sheet?" dialog — you are only asked when you actually want a character sheet.
- The agent can now read full node ids, so "node not found" no longer happens for a node sitting right there on the canvas.
- Ask the agent to tidy the canvas without grouping the nodes first.
The agent features above are in closed beta and rolling out gradually.
Tidy layout: pick a grid, line everything up
- Select several nodes, open Tidy layout in the toolbar, and pick the rows and columns (say 3 × 3). Grids that fill exactly and grids that leave a gap are both marked up front, so you know the result before you click. Undo works.
- Align and Tidy layout now open on hover — no click needed.
- More horizontal room between nodes, so dragging a connection no longer catches the neighbouring node's port.
- The Agent can tidy loose nodes directly, without grouping them first.
- Nodes the Agent creates now carry a title that says what they are ("City square panorama · central fountain") instead of a generic "Image". Ask for "the city shot with the fountain" later and it can find it.
The Agent items above are in closed beta and rolling out gradually. Everything else is live for all users.
Prompt caching is on for chats, and image pricing gets finer
- Prompt caching is on for conversations: repeated context inside a conversation bills at the cached rate, one tenth of base, per the model providers' own pricing tiers. In practice that's about 12% less on Claude chats and up to 57% less on GPT-5.6. Pricing still follows each model's official rates. Nothing to configure.
- SeeDream images: 14 credits at 2K, 10 at 1K, with wide ratios like 21:9 billed at the 1K rate. Extra reference images are billed per image, with the first few free.
- Cancel a queued generation and the credits return to the wallet they came from. A task with no response finishes and refunds within three hours.
- Resume an interrupted Agent task and only the newly completed steps are billed.
- The canvas ledger filters by date and generation type. Shared canvases settle against the canvas wallet, and the members page shows each person's net spend.
Items mentioning the Agent are currently in closed beta and rolling out gradually. Everything else is available to all users.
Canvas Agent is live
- Full write-up in News: “Canvas Agent: describe it once, let it finish the job.”
Canvas Agent is now open to everyone — open any canvas and it's there. No request needed.
Edits stay local, output stays sharp
- Ask the Agent for a local change and it changes that one thing. Wardrobe, pose, and background carry over from the original.
- Aspect ratio follows the source file's actual pixels, so a portrait source returns portrait output.
- SeeDream ships lossless PNG, so detail holds when you zoom. SeeDream 5.0 Pro has no prompt length cap, and camera controls are available.
- GPT Image 2 Lite and Nano Banana accept reference-image uploads for image-to-image.
- Transient interruptions during image generation retry on their own — no need to resubmit.
- Image nodes show the current quality tier while collapsed, and Wan's thinking mode is a toggle on the panel.
Items mentioning the Agent are currently in closed beta and rolling out gradually. Everything else is available to all users.
Three new model families for the Agent
- The Agent model picker adds Claude Fable 5, Claude Sonnet 5, three GPT-5.6 tiers (Luna / Terra / Sol), and DeepSeek V4 Flash and Pro. Each entry notes what it's good for.
- GPT-5.6 offers low and high thinking effort; high runs roughly 1.7× the reasoning of low.
- SeeDream 5.0 Pro joins the image models. Kling O3 joins video, and new Agent video nodes default to universal reference — locking the character without locking the composition.
- Seed Audio 1.0 joins audio.
Items mentioning the Agent are currently in closed beta and rolling out gradually. Everything else is available to all users.
Figma-style shortcuts and a node list
- Canvas shortcuts follow Figma: V select, Z zoom, M annotate, S screenshot, and L opens a node list drawer you can filter by type and status, then jump straight to a node. The shortcut panel groups keys by tool, canvas, and node.
- Reference assets reorder by drag, previews carry a title and number, and removing one renumbers the rest.
- Text and Markdown nodes are plain documents, and the cursor stays where you typed.
- Thirty scrollable areas across the canvas gained scrollbars, so long content scrolls.
- Generating nodes show live progress, and failed nodes are marked on the canvas itself.
- On Windows Edge, Enter commits the candidate in Chinese input instead of sending the message.
Generations from the CLI and external assistants land on the canvas
- Generations started from the CLI, Claude, or Cursor land on the canvas — nine kinds in all. Uploading a local file drops it in as a node too.
- Reading a large canvas starts with an overview, then search and detail on demand: a 199-node canvas summarizes in about 12,000 characters, and full prompts page in.
- Search canvas assets by name and reuse what's already there instead of re-uploading.
- Placed nodes carry @image-N / @audio-N / @video-N reference chips, same as nodes you create by hand.
- Generation progress polls, assets come back as complete links you can open, and results preview inline in the conversation.
- An unrecognized model name returns an error with the valid options. Current version is 0.1.19.