Give each agent loop its own turn and budget ceiling in code, and set a hard spend limit on the organization and on every project. Once a project or organization reaches its cap, OpenAI returns a 429 error, and recorded spend can run slightly over while enforcement catches up. Your usage tier sets the monthly usage limit and the per-model rate limits, so check it before a launch multiplies traffic.
When you need this
- Agents run unattended, in pipelines or overnight jobs.
- Several teams or products share one OpenAI organization.
- A launch or a new feature will multiply traffic within days.
Evaluation criteria
Usage tier
Does your tier's monthly usage limit cover the month you are planning?
Look forBuild at US$5 in credit purchases with a US$500 monthly usage limit, Launch at US$100 with US$5,000, and Grow at US$500 with US$200,000. The tier upgrades automatically as credit purchases reach each threshold.
Hard spend limits
Is a hard spend limit set for the organization and for each project?
Look forSet monthly hard limits under Organization limits and Project settings, with spend alerts below them. Organization limits return 429 with organization_spend_limit_exceeded. Project limits return 429 with project_spend_limit_exceeded. Recorded spend can slightly exceed the limit while enforcement propagates.
Rate limits per model
Do your peak requests and tokens per minute fit the tier's limits for each model?
Look forOn Build, 5,000 requests and 1,000,000 tokens per minute for GPT-6 Astra, Sol, and Terra, and 2,000,000 tokens for Luna. On Grow, 15,000 requests and 40,000,000 tokens per minute. One builder's game hit a 30,000 tokens-per-minute limit on its first busy day.
Ramp rate
Will traffic grow faster than the ramp-rate limit allows?
Look forIncreases of no more than 50% every 15 minutes once traffic passes 1 million tokens per minute. A slow_down error can arrive inside your rate limits, because it measures how fast traffic grew. OpenAI serves Fast mode requests that ramp too fast at standard speed and price.
Separate pools
Is offline work kept out of your synchronous limits?
Look forKeep Batch jobs in their separate pool so evaluations and backfills leave live-traffic limits available. Queued prompt tokens count per model. You can create up to 2,000 batches per hour.
Cost per completed task
Do you budget per completed task, or only per million tokens?
Look forCount reasoning tokens and long prompts in cost per completed task. Above 272,000 input tokens, the full request costs 2x input and 1.5x output. One builder cut a monthly bill from about US$400 to US$120 by routing to a cheaper model first.
Questions to ask vendors
- Which projects need their own hard spend limit, and who is alerted before it is reached?
- What peak tokens per minute do we expect at launch, and how long will the ramp take within the 50% per 15 minutes guidance?
- When would Scale Tier or Reserved Tier pay off against pay-as-you-go traffic?
Red flags
- Spend alerts set with no hard limit, so a runaway agent keeps spending.
- Spend reviewed monthly from the invoice when it should be reviewed daily from usage data.
- Batch and caching discounts assumed in the budget but never measured.
- Agent loops with no turn or budget ceiling of their own.
Products in this piece
Sources
- Official documentation
- 8
- Public source
- 3
- OpenAI API docs: Rate limits
- OpenAI API docs: Spend limits
- OpenAI API docs: Admin APIs
- OpenAI API docs: Batch API
- OpenAI API docs: Fast mode
- OpenAI API docs: Pricing
- OpenAI API docs: GPT-6.1 Terra
- OpenAI API docs: GPT-6 Luna
- Hacker News comment on hitting a 30,000 tokens-per-minute limit (Apr 27, 2025)
- Hacker News comment on routing to cheaper models, US$400 to US$120 a month (Nov 16, 2025)
- Hacker News comment on cost per task for GPT-5.5 and GPT-5.5 Pro (Jul 6, 2026)