AI Agent StoreShort-context rates apply through 272,000 prompt tokens.
Global processing rates.
SAME MODEL · DIFFERENT SERVICE TIER
How much is
faster output worth?
GPT-6.1 Sol gained Ultrafast mode on October 8. It costs six times the Standard token rates; Fast costs twice as much. Keep the same model and token assumptions here, then measure the time saved in your own workflow.
OpenAI’s own prices checked 2026-10-15. Service tiers change how the model is served. Thinking effort is a separate setting; paying for a faster tier does not establish a higher quality score.
Estimate the token bill
Tokens are the pieces of text the model processes. Include charged thinking tokens in output. Caching means reusing previously stored input at a different rate: enter the percentage of your total input that qualifies, or zero if none does. Cached input is counted once. This estimate excludes cache writes, tools, retries, taxes and other fees.
Short-context rates apply through 272,000 prompt tokens.
Global processing rates.
Short-context rates apply through 272,000 prompt tokens.
Global processing rates.
Compare prices and available regions
Standard
US dollars per million tokens, before regional processing charges.
- Input
≤272K / >272K prompt - $2 / $4
- Cache read
≤272K / >272K prompt - $0.1 / $0.2
- Cache write
≤272K / >272K prompt - $2.5 / $5
- Output + thinking
≤272K / >272K prompt - $10 / $15
Global, US and EU processing documented.
Fast
US dollars per million tokens, before regional processing charges.
- Input
≤272K / >272K prompt - $4 / $8
- Cache read
≤272K / >272K prompt - $0.2 / $0.4
- Cache write
≤272K / >272K prompt - $5 / $10
- Output + thinking
≤272K / >272K prompt - $20 / $30
Global, US and EU processing documented.
Ultrafast
US dollars per million tokens, before regional processing charges.
- Input
≤272K / >272K prompt - $12 / $24
- Cache read
≤272K / >272K prompt - $0.6 / $1.2
- Cache write
≤272K / >272K prompt - $15 / $30
- Output + thinking
≤272K / >272K prompt - $60 / $90
Global, US and EU processing documented.
Measure the benefit before paying for it
Use the same examples, tools and reasoning effort. Record time to first useful output, completion time, failures, billed tokens and total cost. Faster output tokens do not establish faster planning, fewer retries or better answers.
Ultrafast is broadly available for Sol and Astra, subject to separate rate limits and residency eligibility. Native requests use the Responses API with service_tier set to ultrafast. OpenAI recommends a persistent WebSocket connection for frequent tool calls; network overhead can reduce the gain. Gateway exposes its own serving contract, which must be checked separately.
Batch processing is a separate option for work that can wait. The two new Sol batch listings are excluded from this comparison of requests that return an immediate response. Pro mode can use more tokens, so equal token prices do not imply an equal task bill.
Release notice ↗ · Native pricing ↗ · Ultrafast access and API guide ↗
Compare model upgrades and routed offers ↗