Spend fewer tokens without changing your API.

Appends a terse-output system instruction so the model answers directly without preamble or restating the question.

On long conversations (more than 24 non-system turns) keeps the system message(s), the first user message, and the most recent turns, dropping the stale middle of the history.

Drops consecutive messages with an identical role and content, and removes exact-duplicate system messages that appear more than once in the conversation.

Serves deterministic repeat requests (temperature ≤ 0, non-streaming) straight from a per-user response cache, skipping the upstream call entirely.

Detects fenced ```json blocks and inline JSON objects over 120 characters inside message text, then parses and re-serializes them minified.

When a request specifies no max_tokens / max_completion_tokens, sets a default output cap (2048 tokens) to prevent runaway generations.

Routes short, simple prompts (under ~500 estimated input tokens, no code fences, no tool calls) from a frontier model to a cheaper sibling via a configured route map.

Collapses runs of spaces and tabs, strips trailing whitespace, and squeezes 3+ blank lines down to one blank line in every text segment of the request.