Ten Days in September

Between the fifth and the fifteenth of September, a fair amount changed. Here is the short version, features first, and then the two things that were broken.

Runs that read like results

The run screen used to be a list of steps, which is a fine way to show a machine what it is doing and a poor way to show a person. It is now one progress bar, the current phase in plain words, and — when the template's own recent runs make the number honest — how long is left. When it finishes, the thing you asked for is at the top. Everything else folds away under "Also created", and the steps and raw inputs move under Advanced details, where they are still there when you want them.

You can also ask to be told when a run is done and get a single notice at the end, instead of watching a page.

More models, and a clearer way to choose

The catalogue had drifted behind the providers, so we went through their pages line by line. New reasoning and coding models, a smaller realtime voice model, a transcription model, a new Claude, two new Gemini tiers, a new music model, a new Grok, a new image model, a faster DeepSeek tier and a new video model all joined within the same week. Several end-of-life models retired, and one supposed shutdown date — which would have switched three working models off in October — turned out not to exist at all.

Alongside that, the documentation now explains how a model actually gets chosen. Auto is capability-first: work out what the turn needs, keep the models that can do all of it, narrow by plan and funding, and rank what is left by the published rate for the units the call will really consume. Choose a model by hand and that choice is honoured — which is also why it can be refused for a job it cannot do.

Images, voice, video

Attach several pictures and the assistant now knows which is which, so "change the second one" works and "do each one" runs across all of them. Asking for a card, a poster or an invitation — text inside an image, the thing cheap models fail at — routes to the quality tier, in any language.

Live voice gained a full-duplex engine: the microphone stays open while it talks, so you can interrupt it the way you interrupt a person, and a session started inside a chat arrives knowing what was already said there.

The video studio gained lip-sync. Hand it a clip of a talking face and a voice track, and the mouth follows the words; the output keeps the clip's length and is billed by it.

Inviting people

Creators have a referral link that attributes the account rather than the browser, so it survives the sign-up and follows through to checkout and renewals. The person invited gets a welcome bonus on their first paid month, and sees who invited them.

Two things we got wrong

Nineteen documentation pages had been written, translated and shipped, and were then quietly withdrawn from the site by a maintenance job that mistook them for retired content. Every "Read the docs" link into that set answered 404 — including the one in a post that had just sent people to the site. They are published again, with four new pages, and a check now refuses to let a link and its target drift apart unnoticed.

Separately, the spending totals on the dashboards were inflated — by roughly 2.4× on the worst days — because a background component kept reprocessing work it had already finished. No money was charged that should not have been; only the displayed figures were wrong. They have been recomputed from the billing records, which were right all along.

We would rather tell you about both than hope nobody looks.