AI Model Selection Rebuilt Around Real Speed and Cost

Version: 2.11.0, Date: August 30, 2026

AI Model selection is the focus of this release. The picker has been rebuilt across all ten places it appears, with search, sorting, a speed filter and a Top 5 Picks shortcut, and the ordering behind it now scores models on values that are actually published rather than on a placeholder that sat at the same number for most of the catalog. Alongside that, several models that were failing outright now run reliably.

• Speed is now something you can see and filter on. Every model picker shows a speed badge and lets you filter by it.
• Rankings run on real data. The default ordering scores cost, speed and context window instead of a placeholder quality figure.
• Long reasoning runs no longer time out. AI Models that think for a minute or more now complete instead of failing partway.
• Colour tones on prompt cards. Pick a palette for a card and shuffle within it.

AI Model Selection

• New: Every AI model now shows a speed badge, and you can filter the list by speed. Speed is expressed as end to end response time, from under five seconds to over sixty, so it reflects how long you wait for a result rather than a raw throughput number.
• New: The picker gained search, a sort menu covering AI model, cost, provider and speed, a Top 5 Picks shortcut and a dense list view for scanning a lot of AI models quickly.
• New: A Cheap and Fast preset scores price and speed evenly rather than sorting by one and then the other. AI Models whose provider does not publish a response time are left out of it rather than being ranked as though they had one.

• Improved: Clicking a selected AI model to deselect it now returns you to the exact view you were browsing, so you do not lose your filters and scroll position.
• Improved: The default ordering now scores cost, speed and context window, all values that come from the providers themselves. It previously leaned heavily on a quality figure that was never supplied by any vendor and sat at the same value for most of the catalog, which meant the order you saw was partly arbitrary.
• Improved: More of the AI models a provider offers now reach the catalog. A filter was quietly dropping a large share of OpenAI’s list before anyone could review it, so AI models that were available were never appearing.

• Fixed: The fastest response ordering was ranking on context window and never looking at speed at all. It now does what its name says.
• Fixed: Reset restored a different ordering than the one the page loads with, so pressing it did not actually return you to the starting view.
• Fixed: Several AI models were priced incorrectly in the catalog. Audio and transcription models charged by the minute, AI models offered at no cost, and AI models priced per character rather than per token were all being read wrongly or skipped entirely.

• Removed: The Output Tokens and Quality dropdowns are gone. Quality was not based on provider data, and both have been replaced by the speed filter and the new sort menu.

Prompt Studio

• New: Color tones for prompt cards. Choose from ten palettes, including bright, pastel, neon, muted, dark, jewel, gray, earth, monotone and randomize, and Shuffle re rolls the colors within the tone you picked.

• Improved: The Main Idea and AI Model blocks now draw from the same color allocation as everything else on the card, so every color on a card is unique. Existing cards keep the bright tone and look exactly as they did.
• Improved: Prompt names can now be up to 80 characters and descriptions up to 200, raised from 30 and 70. Auto write still keeps its description within 70 characters so the gallery listing stays tidy.
• Improved: Topic set headings on a card are centred and bolder, and subtopic titles are bold and a size smaller, so the grouping is easier to follow at a glance.

• Fixed: Shuffle only changed the hue while the shade pattern stayed fixed, so on the gray and monotone palettes it appeared to do nothing at all.
• Fixed: The card header briefly flashed the wrong icon while loading, borrowed from the analyzer pages. It now shows a neutral placeholder that holds its shape until the real icon arrives.

Across the App

• Fixed: Newer reasoning AI models were failing outright with an error saying they were not chat models, because requests were being sent to the wrong place based on a fixed list of model names. Requests now route based on what the provider actually publishes for each AI model, so a newly released AI model works without waiting for an update.
• Fixed: AI Models that think for a minute or more were being cut off after about forty five seconds of silence and failing. Long runs now continue in the background and are polled until they finish, up to nine minutes, and cancel cleanly if they exceed that so nothing keeps billing.
• Fixed: Features that ask a AI model for structured output were failing on the newer models because of how the request was formed. Those now work.