Whizi release notes: August 2026

20 August 2026

• Import your ChatGPT and Claude history. A sidebar entry takes an export file from either service and turns it into ordinary Whizi chats. The file is read in your browser and only the transcript is sent, so a multi-megabyte export never has to be uploaded to be discarded. ChatGPT stores a conversation as a tree of every branch it has ever had, so the import walks back from the node that was last on screen and reconstructs the conversation as you actually left it, rather than replaying discarded drafts and regenerated answers.

• A reference page for every model in the catalogue. The /models family gives each metered model its own page: the list rate per million tokens, what one standard answer costs, the context window, the credits it spends in Whizi, and the plan that unlocks it. The pages regenerate from the price data, so they cannot drift from what you are actually charged.

• Docs rebuilt around the questions people ask. The documentation is now a reference surface rather than a feature tour. Every price and quota sits in one table, the full credit scale is published with all fourteen rungs, and supported file types, billing, cancellation and data handling each have a page that answers the question directly.

• AI glossary and context window comparison. Two new reference pages: a glossary defining the terms that show up in model documentation, and a table comparing context windows across the catalogue so you can see what actually fits in one conversation.

19 August 2026

• Every reply is credited to the model that wrote it. Each assistant message now records its own model instead of borrowing whatever the conversation was set to. Before this, switching models partway through relabelled every earlier reply as if the new model had written it, and stepping back through versions credited the wrong model for answers it never produced. The name, avatar and chip on a reply are now the model that actually generated it.

• The AI Model Cost Index. A published dataset of what a standard answer costs across the whole catalogue, with the spread between the cheapest and most expensive model stated outright. It is free to cite and refreshes whenever list prices move.

• Deep research models retired from the catalogue. sonar-deep-research billed an entire multi-search run per message rather than per turn, so a single message came back anywhere between $0.73 and $2.05 against a rung priced for roughly what one ordinary message costs. The spread between a cheap run and an expensive one was wider than the gap between credit rungs, so no rung could price it honestly and the model was removed rather than repriced. Conversations already pinned to it keep answering.

• Pricing page opens on monthly billing. The plan comparison now defaults to the monthly rate, so the first number you see is the one you pay if you do not commit to a year.

18 August 2026

• Replies survive a closed tab. Generations now run independently of the request that started them, so a reply keeps going if you close the tab or lose connection mid-answer. Reopening the chat replays what you missed and re-attaches to the live stream. Stop still works on a resumed reply, and the partial answer that gets saved is the one the server actually produced.

• Reasoning shows above the answer. On models that expose a reasoning trace, the trace now renders above the reply in both the live stream and the saved conversation. It used to sit below, which looked correct while the answer was still empty and wrong after a refresh, when the finished answer pushed the reasoning to the bottom.

16 August 2026

• Pin, star, rename and delete chats. Each chat in the sidebar has a menu: rename it, pin it to a section at the top, star it in one of five colours, or delete it behind a confirmation. Pins and stars persist between sessions, and titles are held to a single line so rows keep their height. Available in all 25 languages.