Route the task
Use the smallest model that can meet the quality bar with a clear verification loop.
Estimate your monthly cost.
Enter monthly token volume. Cached input applies only to the input share you mark as cacheable. Batch applies a 50% estimate to the eligible total.
The rate card.
These values are the working rates from OpenAI's July 9, 2026 launch information supplied to this project. Recheck official pricing before public use.
| Tier | Input / 1M | Cached input estimate | Output / 1M | Best fit |
|---|---|---|---|---|
| Sol | $5.00 | $0.50 | $30.00 | Frontier judgment and final review |
| Terra | $2.50 | $0.25 | $15.00 | Balanced daily work |
| Luna | $1.00 | $0.10 | $6.00 | High-volume, checkable tasks |
Cost discipline
Three levers matter more than prompt tricks.
Reuse stable context
Prompt caching can reduce the price of repeated input when the provider recognizes eligible cached content.
Batch what can wait
Use Batch for jobs that do not need an immediate response, subject to current provider rules.
Example: 10,000 support summaries.
Suppose each job sends 2,000 input tokens and returns 300 output tokens. That is 20M input and 3M output tokens before caching or Batch.
$38 baseline
20 × $1 input + 3 × $6 output.
$95 baseline
20 × $2.50 input + 3 × $15 output.
$190 baseline
20 × $5 input + 3 × $30 output.
This example is arithmetic from the displayed rates, not a promise of quality or savings. Test a representative sample before routing production work.
Budget checkpoint
Token price is not total cost.
Include retries, tool calls, failed jobs, storage, orchestration, human review, and the cost of errors. A cheaper model that needs repeated rework can cost more.
Continue the cluster