AI Model Context & Character Limits

Every model has its own reference page inside the app

Inside TPEE the API Character Limits panel lists all 182 models, grouped by the company that makes each one. From this build, a double-click on any model's tile opens that model's own reference page in a window you can resize, maximise and scroll in both directions. It collects in one place what would otherwise be a dozen browser tabs: the model's name and the company behind it, the day it was released, what it costs — including the point at which a long prompt starts costing more — a plain description, what the model can do and what it cannot, the problems people actually run into, what it depends on, what you can give it, what it gives back, how it is put together, what is publicly known about the data it was trained on, how that data was prepared, what it is meant for, where its limits are, and the links to the maker's own documentation so every figure can be checked at its source.

Where a maker does not publish something — and for a large hosted model the training data usually is not published — the page says so in plain words rather than filling the gap with a guess. From this build every model's tile opens its page — all 182 of them. Each page states at the top which kind it is. A sourced page is written from the maker's own documentation, with every figure traceable to a link printed at the foot of the page, and Claude Haiku 5.5 (1,000,000-token context, $0.10 and $0.50 per 1M tokens up to a 100K prompt, released 07 October 2026) is the first of those. A catalogue page is composed from the figures TPEE itself already holds — the model's context window, its price, the day it was released, and the point at which a long prompt starts costing more than a short one — and it names the maker's own documentation as the authority for everything else, rather than paraphrasing it. The panel also carries a box reading Double Click Model Tile for Full Information, because a feature nobody is told about is a feature nobody uses. The app stays 100% offline: the links are printed for you to open in your own browser, and the app never opens a connection to anything.

Choosing a model to work with got easier in the same build. A button reading Model Selection Helper sits on its own line directly under the heading of the panel above — it moved there from the right-hand side of the panel in this build, and it was renamed to match the window it opens — and it opens a window of that same name — a window you can resize, which scrolls both ways and narrows the whole catalogue on every figure the app holds: part of a model's name, the supplier column it sits in, the maker, the size of the context window (at least 8,000, 32,000, 128,000, 200,000, 256,000 or a million tokens), a ceiling on the input price and a ceiling on the output price, the models that cost nothing at all, the year a model was released, whether a release date has been published, and which kind of reference page a model has. The result can be sorted seven ways, copied to the clipboard as plain text, and any row can be selected — which makes that model the one the rest of the app is working with — or sent straight to its reference page. The panel itself now scrolls sideways as well as up and down, so a third column of suppliers is reached by dragging rather than being wrapped away. The helper reaches nothing: it filters the catalogue compiled into the app, and no figure on it comes from anywhere else. Alongside the move, every one of the 182 model tiles was checked again against the makers' own published documentation, and forty-six figures across thirty-four models were corrected where the maker's own page disagreed — several context windows that had been rounded, a handful of prices and a handful of release dates. No model was added, removed or renamed, so the catalogue still holds 182 models from 23 suppliers.