OpenAI Ultrafast mode now available on AI Gateway
By Steven Van ·
GPT-6 Astra requests served at Ultrafast cost six times the standard per-token rate and support US and global processing.
Developers building interactive applications or iterating on code can request faster GPT-6 Astra output through Vercel's AI Gateway using OpenAI's Ultrafast service tier.
Requesting Ultrafast
For openai/gpt-6-astra, set openai.serviceTier to ultrafast in AI SDK 6 or later, or set service_tier on Chat Completions, Responses, and Responses WebSocket requests. The unified gateway.serviceTier option does not accept ultrafast.
OpenAI recommends WebSocket mode to keep overhead between turns low. The Ultrafast documentation includes persistent-connection examples for Responses and AI SDK. Responses WebSocket requests must send service_tier on every response.create frame. The Responses example reuses the previous response's context; the AI SDK example reuses the connection with independent prompts.
Regions and billing
Ultrafast supports US and global processing. Requests pinned to other regions, including the EU, run at the standard tier. Standard processing also remains the default when no service tier is specified.
Requests served at Ultrafast cost 6× the standard per-token rate. The requested tier is a best-effort routing hint, so it may not be granted. Requests that fall back are billed at the rate for the tier actually served, which can be checked in the response or request logs.