How much is
faster output worth?

Should you try Terra or pay for a faster Sol?

Terra’s Standard input and output rates are half Sol’s. With 10,000 input and 2,000 output tokens per request, 1,000 requests cost $20 on Terra Standard, $40 on Terra Fast or Sol Standard, and $80 on Sol Fast. These examples exclude caching, tools and regional charges.

Equal bills do not mean equal results. Try Terra on screenshots, image matching or writing; keep the same examples when comparing Sol. Record acceptable answers and retries as well as waiting time. Terra supports turning thinking off; Sol does not. Terra has Standard and Fast options in this calculator, with no verified Ultrafast offer.

When is a longer cache worth paying for?

A cache stores the repeated beginning of your input, such as instructions and reference documents. The default 30-minute minimum costs 1.25 times the ordinary input rate to write. The new 12-hour minimum costs 2.5 times that rate. These are total write prices, not extra charges added on top. Reads cost the same with either lifetime.

For example, at Sol’s Standard short-context rates, writing a 10,000-token prefix for 12 hours costs $0.05. Two later full cache hits cost $0.001 each: $0.052 in total, compared with $0.06 for three uncached passes. This comparison assumes the prefix remains identical, qualifies for caching and is fully reused. It excludes new questions, output and other charges. If the requests fit within 30 minutes, the cheaper default write may be sufficient.

Choose the lifetime around how often the same material will actually be reused. A 12-hour minimum is not a deletion deadline. The Responses API’s newer cache controls differ from older models’ retention settings. OpenAI’s cache controls and billing examples ↗

OpenAI’s government-compliance (FedRAMP) endpoints also carry a 10% price increase from November 1, 2026. They are not a region option in this calculator; confirm the eligible model and service tier for your account.

Could sending less data reduce waiting?

Claude’s new message-thread beta lets an application send only the new turn instead of uploading the whole conversation again. Anthropic says this can reduce the wait before an answer starts, especially in long agent conversations. It does not change token prices, and prompt-cache discounts still need their own cache settings.

Keep a copy of the conversation: a thread expires after five minutes without a request, so the application may need to resend it. Thread contents are deleted within an hour of last use; this feature does not qualify for zero data retention. Check supported platforms and setup ↗

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 ↗ · Terra model details ↗