router/Optimization techniques

Make every optimization a deliberate choice.

Inspect the transformations, see the request before and after, and understand exactly which behavior each technique changes.

Know what is live before you evaluate it.#

See every technique work.#

Pick one. Compare the original request with its transformation.

Illustrative example · output reduction

Concise Output Instruction

Adds an instruction to answer directly, without a preamble or repeating the question.

What is the answer to 6 × 7?

A shorter answer can omit detail. Evaluate the result.

Workspace settings
Understand the tradeoff
Pruning can remove context. Concise instructions can change an answer. Model routing changes the serving model. Preview the change and evaluate it on your workload before enabling it.
Response caching is explicit
Exact caching stores eligible response text per user for reuse within 15 minutes, up to 100 entries of 256 KiB each. It never stores prompt text. Only temperature-zero, non-streaming plain-text requests without tools qualify. Disabling the account setting clears existing entries; an active experiment explicitly enabling caching can still create new entries. Expired entries are cleaned during writes.

Research radar#

Ideas under evaluation, not running on your production traffic.

Batch API RoutingResearch preview
GEPA Prompt OptimizationResearch preview
LLMLingua-2 Prompt CompressionResearch preview
OpenAI Flex Service TierResearch preview
Provence RAG Context PruningResearch preview
Provider Prompt CachingResearch preview
Semantic CacheResearch preview
Sketch-of-Thought ReasoningResearch preview

Availability is determined by the workspace controls and the implementation serving your request. A research listing is not an active optimization.