GPT-6.1 Terra Pricing Explained: Token Costs, Caching, and Task Budgets
Estimate GPT-6.1 Terra API costs with token rates, cache writes, the 272K threshold, service modes, and worked task-budget examples.
Himi LiProduct Manager at EvoLinkOn this page
Those numbers are a starting point for a budget, not the cost of every request. The final amount depends on how much of the input is written to or read from the cache, how many output and reasoning tokens the model uses, and whether the request crosses the long-context threshold. Service modes and tools can add further differences.
GPT-6.1 Terra API pricing at a glance
| Token category | Standard rate for up to 272K input tokens |
|---|---|
| Ordinary input | $1.00 |
| Cached input read | $0.05 |
| Cache write, 30-minute lifetime | $1.25 |
| Cache write, 12-hour lifetime | $2.50 |
| Output, including reasoning tokens | $5.00 |
How to estimate a GPT-6.1 Terra request
Start with the token counts returned in the usage record. A basic estimate divides each count by one million, multiplies it by its applicable rate, and adds the results.
Consider a request with these assumed counts:
- 10,000 ordinary input tokens
- 2,000 total billed output tokens
- No cached reads, cache writes, or separately billed tools
Real requests vary. If one task needs a larger prompt or more reasoning, its cost will change. For a monthly forecast, use the distribution of usage across your own requests where available, and include billable retries and tool charges. A single average can hide a small number of expensive, long-context jobs.
Count reasoning inside output
output_tokens_details.reasoning_tokens; they are part of the total output_tokens. Use the total output count for the estimate, and do not add the reasoning count a second time. OpenAI's reasoning guide explains the usage fields.An output limit can also be reached before a visible answer is produced. In that situation, the request may still have used billable input and reasoning tokens. Check the usage and response status before deciding whether to retry.
How prompt caching changes the bill
Caching can lower the cost of repeated material, such as the same instructions, examples, or reference text sent across multiple requests. It is useful to distinguish three input categories:
- Ordinary input: tokens charged at the regular input rate.
- Cache writes: tokens written for a specified cache lifetime.
- Cached reads: eligible tokens reused from the cache.
Example: reuse one prefix ten times
Assume an 8,192-token prefix is sent ten times. The first request writes the full prefix; the next nine read it from a still-valid cache. This illustration compares only the prefix input cost and excludes output, other input, long-context adjustments, and tools.
| Treatment | Calculation | Prefix cost for 10 requests |
|---|---|---|
| Ordinary input every time | 8,192 × 10 × $1 / 1M | $0.08192 |
| One 30-minute write, then nine reads | 8,192 × ($1.25 + 9 × $0.05) / 1M | $0.0139264 |
| One 12-hour write, then nine reads | 8,192 × ($2.50 + 9 × $0.05) / 1M | $0.0241664 |
For a Responses API usage record that reports both write lifetimes, ordinary input can be calculated by subtracting billed reads and total cache writes from total input. Apply the 30-minute and 12-hour rates to their respective write counts. If a gateway does not return these fields, use its billing information rather than assuming missing counts are zero.
What happens above 272K input tokens?
For an illustrative Standard request with 300,000 ordinary input tokens and 2,000 billed output tokens, with no cache or tool charges:
- Input: 300,000 × $2 / 1M = $0.60
- Output: 2,000 × $7.50 / 1M = $0.015
- Total: $0.615
The model's 1,050,000-token context window gives room for large inputs and outputs, but it does not mean every request within that window uses the short-context rate. For long documents or agent histories, measure the input sent on each request and include repeated context in the forecast.
Batch, Flex, Fast, and tool costs
Choose a mode based on delivery requirements as well as price:
- Batch is asynchronous and has a 24-hour processing window, making it suitable for work that can wait, such as a scheduled classification run. See the Batch guide.
- Flex can be useful when the workload tolerates delays. A resource-unavailable response can require a retry; OpenAI says that specific Flex error is not charged. A retry using Standard processing should be budgeted at the mode actually used. See Flex processing.
- Fast costs more per token, so compare the value of shorter turnaround against the higher token rate for the workload.
Is GPT-6.1 Terra always cheaper per task than Sol?
GPT-6.1 Terra's Standard short-context input and output rates are half those of GPT-6.1 Sol: $1/$5 versus $2/$10 per million tokens. If two requests use exactly the same billable token counts under the same conditions, Terra's model-token charge is half as much.
A completed task can involve different token counts or retries. Suppose, purely as an illustration, that a Sol attempt costs $0.04 and produces an accepted result. If Terra needs two attempts at $0.02 each to reach the same acceptance standard, the model charges are equal. A third attempt would make that example more expensive on Terra.
For production planning, compare the total billable cost of all attempts against the number of tasks that pass your acceptance check:
cost per accepted task = model and applicable tool charges for all attempts / accepted tasksKeep human review and waiting time separate unless you have a consistent way to value them. If no tasks pass, the cost per accepted task is not defined. This comparison helps distinguish a lower token rate from a lower cost to deliver a result your team can use.
A practical way to budget your workload
Before choosing a monthly budget, collect a representative sample of completed tasks. For each one, record the model, service mode, input and output usage, billed cache reads and writes, applicable tool calls, retries, and whether the result passed review.
Then separate routine requests from unusually long or complex ones. Apply the appropriate rate to each category and multiply by your expected volume. Review the estimate after a pilot using the actual usage and invoices from your chosen provider; cache hit rates and output lengths are easier to predict from real traffic than from a rate card alone.
Frequently asked questions
Are cached input and cache writes charged together?
Different portions of a request may be charged as cache reads and writes. The same token is not charged at both the ordinary input and cache-write rate for the same billing classification. Use the usage breakdown and applicable lifetime rates.
Does the higher rate begin when output exceeds 272K?
No. The threshold is based on input tokens. When input exceeds 272,000 tokens, OpenAI's long-context rates apply to the full request, including its output.
