Four workloads the serving path is tuned for.
TOOLS · SESSIONS
The cross-turn cache removes 31% of prefillDETAILS →
DOCUMENTS · 1M WINDOW
1M tokens live on serverless, no surchargeDETAILS →
OFFLINE · THROUGHPUT
$0.07/M blended on the gpt-oss-120b classDETAILS →
STREAMING · LATENCY
Streamed byte for byte, measured at the socketDETAILS →