Every Service Tier, Standard to Ultrafast

OpenAI publishes a separate price table for each service tier — Standard, Batch, Flex, Fast and, since September 29, 2026, Ultrafast — and no model appears in every one of them. They're the same model priced for different delivery guarantees, and the most useful thing to know is which guarantee you're actually buying.

Standard, and the two discounted tiers

Standard is the synchronous reference price. Batch and Flex are both discounted 50% off it, and the two tables are identical for every text model that offers both. What differs is delivery: Batch accepts a file of requests, completes within a fixed 24-hour window, draws on its own rate-limit pool and has no output-token limit; Flex is an ordinary synchronous request, billed at Batch rates, in exchange for slower responses and occasional resource unavailability. Pick between them on the guarantee, never the price.

The two premium tiers

Fast, renamed from Priority, buys up to 2.5x faster responses in the API. It costs a fixed multiple of the Standard rate — the same multiple across the GPT-5.6 and GPT-6 families, a higher one on GPT-5.5 — and the Fast-mode page computes it per model. It shares the model's Standard rate limit, and several models — the Pro models, GPT-5.4 Nano — don't offer it at all.

Ultrafast is the top tier, offered for GPT-6 Astra and, since October 8, 2026, GPT-6.1 Sol — at a price that is a large multiple of each model's Standard rate, with token limits of its own. OpenAI's API guide gives no speed multiple for it; Codex's figure for Astra is up to 8x the token-generation speed of Standard. It's built for agentic loops where a person is waiting on many quick tool calls.

Checking the tier you actually got

The tier you asked for isn't always the tier that ran. The response's service_tier field reports what processed each request: Fast requests downgraded during a steep traffic ramp come back as default, billed at standard rates, and on GPT-5.6 and earlier models Fast requests report priority. Group usage by service tier in the dashboard before trusting any cost report that assumes the requested tier is the billed one.

Picking a tier

Start from the deadline, not the price. Work that can finish by tomorrow belongs on Batch. Synchronous work where nobody is waiting belongs on Standard or Flex. Fast earns its premium only on a critical path; Ultrafast only when token generation speed itself is the bottleneck. Inside a ChatGPT plan the arithmetic differs again — speed modes burn included usage faster than purchased credits.

The tables above show each tier's definition and every priced model's input rate on each tier it publishes.