What shipped, and when
Every change a user can see, dated the day the whole of it became available. Internal work is deliberately absent.
Getting started, model presets, and a usage page
Seven changes shipped together, all aimed at the first hour in the product and at seeing what it is doing.
- A setup checklist on the dashboard that counts real objects — connected apps, skills, agents, automations — and disappears on its own once you are set up.
- A seven-lesson academy, each lesson ending on something you actually do, with a counter in the sidebar.
- Named model presets above the raw model list, plus a !fast and !deep prefix that swaps the model for one message.
- A calendar view of scheduled automations, read in the automation's own timezone.
- A Usage page splitting activity from measured consumption, so neither number pretends to be the other.
- A tone and a set of instructions for the whole account, applied in chat, in automations and in agent runs alike.
Referral rewards and automatic credit top-ups
Recurring rewards on what the people you refer actually spend, and a way to stop running out of credits mid-run.
- Referral tiers from 10% to 20%, recomputed every month from real payments rather than list prices, and paid in credits.
- Automatic top-ups with a monthly ceiling that stops rather than overruns, and warns once per period.
- A fix found on the way: bought credit packs were recorded but granted no usable quota.
Dark mode, working
The dark theme was selectable but never actually applied across the dashboard. It does now.
- Editorial colour tokens are redefined for dark, so the whole dashboard follows the setting instead of a handful of components.
- The warm gold wash of the light theme is preserved rather than flattened to grey.
Pay the invoice that failed
A declined card used to leave the account stuck with no way back from inside the app.
- Subscribers whose payment failed can now settle the outstanding invoice directly, instead of being routed to a checkout that refused to reopen.
Right of withdrawal, in one click
The withdrawal function required of online sellers in France since 19 June 2026, built into the product rather than left to email.
- A withdrawal form reachable without signing in, because someone exercising this right may no longer have an account area.
- Requests are recorded and acknowledged in writing.
Kimi K3, and code execution in a sandbox
A new default model available from Starter, and the ability to run code as part of a task.
- Kimi K3, served from our own deployment, becomes the app-wide default and is available from the Starter plan.
- A sandboxed execute_code tool, so an agent can compute rather than guess.
GPT-5.6 Sol, Terra and Luna
Three new models, and the tool-calling work needed to make them usable.
- The GPT-5.6 family added to the pickers, gated by plan like the rest of the catalogue.
- Tool calls routed through the Responses API so the family works with integrations rather than only with text.
Flows become AI Automations, and agents are configured in chat
A rename that matches what people call them, and one less screen between describing an agent and having one.
- Flows are now AI Automations throughout the product.
- Agents can be configured from the conversation instead of a separate form.
A cloud browser your agents can drive
Agents with the browser skill run against a pay-as-you-go cloud browser, which makes browsing tasks cheaper to leave running.
- Browser sessions moved to Steel, billed per use rather than per seat.
- Screenshots stream back into the conversation as the agent works.
GPT-5.5, Claude Opus 4.8 and Claude Fable 5
Three frontier models added across the paid plans.
- GPT-5.5 available from Starter.
- Claude Opus 4.8 on Business, Claude Fable 5 on Enterprise.
Automations that fail loudly instead of quietly
A scheduled automation that breaks at 3am should tell you, not retry forever.
- An automation that fails three times in a row is paused automatically, and you are emailed.
- Transient integration errors are retried and reconnected instead of counting as failures.
- Timeouts end the run cleanly rather than leaving it half-done.
Marketplace, agent board, skills hub and run history
The release that turned a chat with tools into a workspace you can look inside.
- A marketplace where creators publish skills, free or paid.
- An agent board: a durable task queue agents pick work from.
- A run history recording every step, tool call and result.
- Overnight self-improvement, off by default and reviewable before anything is adopted.