Release 0.37.0
Release Date: September 2026
Previous Version: v0.36.0
Build your first workflow from Projects, revisit complete node results, and preview videos before finalizing them. This release brings together new user-facing features since 0.36.0, including the intervening 0.36.1 release. Available models, nodes, and settings depend on your account and workspace access.
Start your first project from an idea
Eligible new workspace owners can describe a workflow directly on Projects, attach reference files, choose a configured template, or start with an empty canvas.
Build creates the project and passes your prompt and references to the assistant, which constructs an editable workflow without running its nodes. Opening Projects by itself does not create a project or open a subscription-plan dialog. The transition into the assistant respects reduced-motion preferences.
Find your way around the workspace
The top bar now shows your workspace, account, or project as a navigation path. Switch workspaces from the workspace segment and open project actions from the project menu.
Settings groups General, Members, API Keys, MCP, Billing, and Node Pricing under Workspace, and Profile, Security, and Hotkeys & UI under You, according to your access. Opening /settings takes you to Profile.
The Assets and AI Assistant panels now sit below the editor top bar. Each has a top-bar toggle and its own close control, so you can keep the tools you need beside the canvas. Navigation also adapts to narrow windows and short screens, including loading and missing-workspace states.
New regular trial workspaces are named Personal. Existing workspace names and campaign workspace names are preserved.
Revisit complete results from earlier runs
Use Previous run and Next run beside a node's price to switch between recorded executions. Atlas restores all outputs from the selected run together—including image arrays and their metadata—while keeping your current inputs unchanged.
The selected result survives a reload and updates downstream readiness. If your inputs have changed since that run, the node tells you that the displayed result is older. After a failed attempt, Restore selected run can recover the retained successful result.
Cached executions appear as separate runs with their own zero-credit charge.
History retains up to 20 runs per node, subject to storage limits. Navigation is available to editors and pauses while a node is running, scheduled, or saving. Shared views and copied projects do not carry over the source project's run history.
The assistant and graph tools also receive clearer result context: which run is selected, its place in retained history, its charge and cache status, and whether it matches the current inputs. This helps the assistant explain an older displayed result separately from the latest execution attempt.
Preview a video, then finalize it
Seedance draft generation and FLUX draft generation and extension now provide a playable Output Video. Connect that preview to ordinary video nodes—for example, to extract a frame—without finalizing first.
Use the separate Finalize Seedance Draft or Finalize FLUX Draft node when you want the final video. Connect the matching Seedance Draft or FLUX Draft metadata output to its finalizer; the playable Output Video follows the ordinary video path. Finalization reuses the draft's original prompt and references.
Finalize FLUX Draft offers 720p and 1080p output, where supported by the draft. FLUX continuation drafts use the same preview-and-finalize pattern.
Finalization is a separate paid operation. Expired or unavailable drafts need to be regenerated, although a previously finalized result may still be returned from cache. Preview and final videos can be used as ordinary API video outputs; draft metadata ports are internal graph connections, not external API file types.
More image and vision-model choices
Seedream 5 Flash is available in image generation, image editing, and Split Image into Layers, wherever enabled for your account.
DeepSeek V4.1 Flash replaces V4 Flash, with saved older selections upgraded automatically. Both DeepSeek V4.1 Flash and GLM 5.3 Flash support image inputs in the model catalog.
Report problems and follow replies inside Atlas
Open Report from the canvas toolbar, Give Feedback from the sidebar, or Send feedback from the help menu. Choose Bug, Dislike, Feature request, Bad result, or Billing, then add a title and details.
Capture your screen, upload screenshots, or paste or drop them into the form. Project link, Console log and Environment controls let you choose the context to include; they start enabled. The console-log option also includes an available session replay. Closing the form or a failed send keeps your draft text for the current browser session.
My reports brings together your reports across workspaces, with status labels and an unread indicator when there is news. Open a report to read the Atlas team's replies, add your own reply, mark it resolved, or delete it. Team answers and fixes can also trigger email notifications; a fixed report shows its release version when one is provided.
In-app reporting is available when enabled for the environment. Sending a report requires workspace membership and a verified email address. Contact support in the account menu still opens email, and feedback entry points fall back to email when in-app reporting is unavailable.
Keep a separate setup in each browser
Atlas remembers the last workspace used by each account in that browser. It also keeps that browser's open state and width for the Assets and AI Assistant panels, plus mouse or trackpad input mode, zoom step and directional panning speed.
These choices survive sign-out and take precedence over account defaults, so your laptop and desktop can keep different canvas setups. Panel and canvas preferences update across tabs in the same browser. A browser without its own choice starts from the account's saved value.
Adjust canvas input settings under Settings → You → Hotkeys & UI. Open, close or resize the editor panels to save their layout preferences.
Schedule several selected nodes together
Select runnable nodes on the canvas and choose Schedule selected to queue them with their prerequisites. Branches share work already in progress and reuse valid prerequisite results. Explicitly selected inactive targets run again, and another branch can join while a shared prerequisite is running.
Use Cancel queued to cancel selected waiting nodes and skip their queued dependents. Upstream work and other branches continue.
Run only the nodes you selected
Choose Run selected only when you want to execute a selection using available inputs. Atlas starts no unselected nodes and waits for selected dependencies or inputs already being produced by active work. The assistant can use the same selected-only scheduling behavior.
Blocked selections and their selected dependents are skipped while independent selections continue. Skip messages identify missing required fields.
Paste an image directly onto the canvas
Press Ctrl+V or Cmd+V on the canvas to upload a clipboard image and create an Input Image node at the pointer. Standard paste uses the browser's native paste event, including in Firefox without an extra asynchronous clipboard-read prompt.
Copied graph nodes keep priority over clipboard images. Images take priority over accompanying text, and plain text creates an Input Text node. Pasting inside text editors or the assistant keeps its usual behavior; read-only canvases do not create nodes.
Choose a node, then its connection field
Drag an input or output connector onto empty canvas to search compatible nodes. Search now shows one row per node. Select a node to choose among its compatible inputs or outputs; a single compatible field creates and connects immediately.
Both drag directions keep the existing compatibility and access checks. Back restores your search and group. Single-field choices skip the field-selection step.
ElevenLabs v4 speech and dialogue
Text to Speech (ElevenLabs) now offers Eleven v4 and Eleven v4 Turbo alongside the existing models.
Text to Dialogue adds v4 for two-speaker A:/B: scripts; Turbo is single-speaker. Eleven v3 remains available and stays the default.
Shape v4 delivery with Stability and Similarity, choose a best-effort Seed, and set Text Normalization to Automatic, On or Off. Previous Text and Next Text provide continuity without being spoken. These controls are available on both v4 speech variants and v4 dialogue.
Connect Voice Selection's Selected Voice output to TTS Connected Voice to override its manual voice dropdown.
Atlas accepts up to 10,000 characters for v4 speech and 2,000 spoken characters for v4 dialogue. Use both A: and B: turns for dialogue and choose a voice for each speaker. Existing v3 limits are retained.
Prepare audio tags for the selected ElevenLabs model
Add Audio Tags now includes an ElevenLabs Model selector for v3, v4 and v4 Turbo. Match it to the downstream TTS or Dialogue node. Provide a delivery direction and optional Voice A for narration, or Voice A and Voice B with A:/B: turns for dialogue.
The node adds descriptive delivery tags while preserving the example's words and speaker turns. It checks the selected model's limit, including added tags, and rejects incompatible speaker structure or over-limit generated text instead of silently truncating it.
Browse the refreshed ElevenLabs voice catalog
The refreshed catalog organizes voices by category, role and delivery, with full voice names, M/F marks and searchable traits. The current catalog contains 1,577 voices, including new character, game and film voices and verified replacements; availability follows your account and workspace access.
A group-path indicator below the search box keeps your current location visible when you scroll past group headers.
New default voices are Zoe and Darian. Saved retired default and unavailable community voice selections are remapped to verified replacements, while saved models, explicit settings, generated audio, layout and connections are preserved.
Last updated