Project Peak reports one headline per model: how much of its advertised context window stays reliable, at a usage bar you choose. This page explains what that number means and how to use it. It covers how to read the results, not how the tasks are built.
Usable context is the longest input a model stays reliable to, measured at a fixed difficulty and read at three usage bars: Critical (right 95% of the time, for work you cannot get wrong), Normal (85%), and Exploratory (75%). We put it next to the advertised window, so window efficiency (usable over advertised) is the gap that actually matters.
Usable context is measured at a fixed difficulty tier, so the only thing that moves is context length. Frontier is the hardest reasoning, where the strongest models start to bend. Standard and Light let cheaper or self-hosted models show a real falloff instead of flooring. Switch tiers on the leaderboard to compare at the difficulty your work actually sits at.
Every boundary carries a bootstrap confidence interval. A wide band means we should sample more. A tight one means the estimate has settled. A star next to a bar marks a lower bound we have not resolved yet at the current sampling depth, so read it as a floor, not a finished number.
Two models reliable to the same context can fail differently past it. Decline begins marks where reliability first dips below the model's own plateau. The half-life marks where it has fallen halfway. The slope says whether that is a sharp cliff or a gentle decay. The shape is how you size your margin.
Two models can advertise the same window and stay reliable to very different lengths. Pick the usage bar your work needs, read each model's usable context at that bar, and send the job to one whose usable context clears the length you actually feed it. That keeps you inside the range the model holds, not the one on the spec sheet.
Higher reasoning effort usually buys more usable context but costs more. Sort the leaderboard by Value to see reliable context per dollar, which can surface a cheaper reasoning level than the one that reaches furthest. If a lower level already clears your length at your bar, it is the better buy.
Past its limit, a sharp-cliff model collapses fast while a gentle-decay model fades slowly. For a blocking, auto-merged check, stay well inside the limit of a sharp-cliff model. A gentle-decay model gives you more usable room near its edge. The shape tells you how much headroom a task needs.
Measurements use synthetic, procedurally generated code tasks (no scraped code, so no training contamination), scored exactly and deterministically. We sweep input length at a fixed difficulty and concentrate trials near the boundary, stopping once the confidence interval is tight, so the result is accurate without costing more than it needs to. A difficulty-axis score (how hard a task a model sustains, 0 to 100) is kept as a secondary view. We do not publish enough task internals for anyone to rebuild the generator.