What’s new

Built-in media generation

Agents generate images, video, and audio through the platform's own Replicate access, no separate account or API key needed. Output URLs expire quickly, so agents download anything worth keeping into the workspace. Available when the platform is your LLM provider, which meters and bills the calls.

Grok 4.6 in the model picker

Grok 4.6 is xAI's newest model and now the default for the grok family on both the platform and OpenRouter providers; 4.5 stays available as an explicit pin. It bills at $2 per million input tokens and $6 out, doubling past a 200k-token prompt.

Subagents and tasks share one activity rail

Subagents, background processes, and planned tasks now share one rail above the composer instead of three stacked sections. A band travels down the trunk and out along each live branch to show what is working, replacing the pulsing dot on every busy row. Silent under reduced motion.

Fixes and improvements

• Much faster first dashboard launch, with a visible preparation step.
• Open Cloud Agents keeps returning visitors signed in.
• Word-by-word reveal for streamed replies.
• Smoother transcript scrolling as text streams.
• Scrolling to the bottom re-pins the transcript.
• Unified error banner above the composer.
• Lighter heading weights so replies read as messages.
• Workflow approval buttons stay visible in Safari.
• Agent-created child agents are usable without a reload.

Usage analytics in Settings

Settings → Usage is now a dashboard: cost and token charts over a date range, broken down by model and agent, with summary metrics. Click a legend entry or a breakdown row to isolate one model or agent and see selected-versus-total figures.

Expired connections hold instead of failing

When an agent reaches for an account whose access has expired, the request parks and a reconnect prompt appears in chat; reconnecting resumes the original call. The agent home Connections card shows expired accounts and MCP servers. Ordinary messages still start immediately.

Meta's Muse Spark models

Meta's Muse Spark 1.1 and 1.2 are available on the platform provider: $1.25 in and $4.25 out per million tokens, a million-token context, image input, prompt caching. No hosted web tools. The contributor tier is cheaper, but Meta may use those prompts for product improvement.

Fixes and improvements

• Collapsing agent turns into a single row for better session scanability.
• Improved queued message handling and UI.
• Telegram remote chat stability improvements.
• Added default model selection to the onboarding flow.
• Better first agent prompt experience during onboarding.
• Installing a template now closes the setup wizard.
• Saved credentials fill the page shown in the browser viewer.
• Manual reveal and copy when a page hides its password field.
• Empty thinking cards no longer collect in long turns.
• Neighborhood focus when hovering the connections graph.
• New thought orb showing what the agent is doing.
• Restored the Gamut wordmark in the browser sidebar.

Your homepage prompt follows you through signup

A prompt typed into the homepage composer, and the model picked beside it, waits in the create-agent box when your workspace opens. Pick a template instead and its install dialog opens with the name filled in. New signups only; nothing runs until you press send.

Fixes and improvements

• Added the model picker to the create-agent composer.

Push notifications on your phone

Install Gamut to your home screen, switch on Push to this device in Settings → Notifications, and your phone gets a banner when a session finishes or stops to ask something. Tapping it opens the session. iOS 18.4 or later only.

Fixes and improvements

• More than halved the JavaScript needed to boot the web app.
• Full-UI precaching for instant navigation after first paint.
• Atomic asset swaps so open tabs survive a deploy.
• Faster agent lists and session starts via cached summaries.
• Opaque question, secret, and approval cards.
• Stable thinking-card identity across stream handoff.
• Incremental Chrome profile sync that preserves agent logins.
• Leaner production install and a much smaller server image.

Pair the iOS companion app

Settings → Mobile mints a single-use QR code or deep link that the iOS companion app redeems for a 90-day session. Paired devices are listed there and can be revoked individually. Pairing tokens expire after five minutes, and the tab appears only in auth mode.

Agents read the PDFs you attach

Attach a PDF, audio file, or video and the agent can read it — the container ships poppler-utils and ffmpeg, so page extraction, metadata probes, and frame grabs all work. Every format the file viewer renders is covered; OCR and Office documents are not.

Subagents on other providers' models

The parent agent can route a delegated task to any enabled model in your provider's catalog — a cheaper one for parallel exploration, a stronger one for a hard branch. Each family's latest model appears as its own subagent type; older versions do not.

Private GitHub skillsets

Skillsets in Settings can now carry a repository token, so private GitHub repositories work on hosted Gamut, which has no host credentials to inherit. The token is stored apart from skillset metadata and only ever returned masked. A fine-grained PAT with read-only Contents is enough.

Fixes and improvements

• Richer web search and fetch source cards.
• Markdown rendering inside thinking blocks.
• Embedded dashboards appear as soon as the server reports running.
• Translucent composer with a collapsible activity card.
• Smarter auto-follow of streaming output at the live edge.
• Slash commands resolve to the correct skill slug.
• Improved dictation contrast and mobile composer layout.
• Retryable answer submission instead of an endless spinner.
• MCP OAuth completes against cloud deployments from the desktop app.
• Settings → Users links to Platform on managed deployments.
• Faster cold wake for suspended cloud deployments.

Agent-to-agent permissions have their own page

Cross-agent policies open as their own page from the agent home Extras card, listing every agent as a Connected / Not connected toggle. Toggling on grants send immediately; toggling off can pin an explicit review, so you carve exceptions out of a global allow. Owners only.

Fixes and improvements

• Agent containers warm up as you type your first message.
• Provider tabs in the model picker now select a model.
• Better turn anchoring so replies stream into reserved space.
• Dashboard deep links, cookies, and HMR work under artifact mounts.
• Opt-in mounted upstream paths for dashboard servers.

The agent home is a widget board

Agent cards sit on a snap-to-grid board you rearrange in Arrange mode, each a live canvas that shifts color when a session needs input. Arrange lives in the Your Agents menu, and phone and desktop layouts save separately, so tidying one never disturbs the other.

Cloud agents from the desktop app

A Local / Cloud switch in the desktop toolbar points the app at your organization's deployment, so you can drive cloud agents without leaving the desktop. The tray, the Agents menu, and deep links follow the switch, which appears only when a cloud workspace is connected.

Sign in with your password manager

Agents can pull a credential from Apple Passwords through a host-side broker instead of stopping to ask you to type one. Suggestions stay metadata-only until use, and you approve each retrieval. Setup is under Settings → Password Managers, and needs the Chrome extension plus a six-digit pairing.

Browse from inside the agent container

Agents can open a page in the container's own Chromium instead of your configured host browser, reaching dashboards and dev servers on private container ports that a host browser cannot see. The choice sticks until the agent explicitly asks for the configured browser again.

Fixes and improvements

• Agents now arrive as a saved iMessage contact card.
• Tightened Allow once to authorize a single script run.
• Stricter session ownership checks in auth mode.
• Clickable links in agent request cards.
• Slack integration stability improvements.
• Improved MCP proxy handling of non-streaming servers.
• Smaller browser automation profiles with automatic cleanup.
• Faster dashboard availability after container start.
• Roughly halved cold-start pull time with smaller images.
• Consistent policy controls on the Connections page.
• Refined scrollbar styling on Windows.