Performance Modes — Fast, Medium & Beast Go Deeper

July 13, 2026

The performance slider is no longer just a routing hint. Fast, Medium, and Beast now dial reasoning effort, verbosity, output caps, and how aggressively context is compacted — so the mode you pin actually changes cost and depth.

What you can do

• Pin how hard the model thinks — and how much context it keeps. Fast uses low reasoning effort / low verbosity and caps output sooner; it also compacts context earlier. Medium keeps catalog defaults and the model’s normal soft budgets. Beast raises reasoning (and pro mode where supported) and holds a larger active window with a lighter post-compaction keep.
• Keep a pinned mode authoritative. When you pin Fast / Medium / Beast, that choice wins over per-turn complexity heuristics. On Auto, complexity bands still lead for scaling models; the mode fills knobs those bands don’t set.
• Switch modes mid-thread without thrashing the KV cache. Context budgets ratchet: Beast expands immediately; Fast shrinks only at a safe compaction boundary.

Where this shows up

• You’ve got a quick triage thread and pin Fast — shorter answers, cheaper reasoning, earlier compaction.
• Later you pin Beast for a deep brief and Alfrada keeps more context and thinks harder without you changing models.
• You leave the slider on Auto for everyday work — complexity bands still steer scaling models; the mode only fills in knobs those bands don’t cover.

Try it

• "Pin Fast — short answers only for the next few questions."
• "Switch to Beast and write a full competitive brief with sources for each claim."
• "Leave me on Auto, but tell me which performance mode actually ran after this answer."

Heads up

• Grader / rubric behaviour from the June 28 note still applies — Fast never self-grades; Medium does one review pass on large work; Beast keeps revising. This release adds the deeper model-args and context-budget controls behind those same mode names.
• Knobs are capability-gated — a provider only receives parameters it actually supports (e.g. verbosity on GPT-5.x, string reasoning on some Ollama models).