OpenAI API Usage Tiers Shrink From Five to Three

Your sprint plan probably has a line that says "API capacity: fine." On 6 October, OpenAI changed the ladder that line depends on. The API changelog says it "simplified API usage tiers from five to three: Build, Launch, and Grow."

OpenAI API usage tiers now set your rate limits and monthly ceiling from three credit-purchase thresholds: $5, $100 and $500. The changelog entry doesn't say how the two removed tiers map onto the new ones, so don't assume your old tier number translates. Check your own account instead (steps below).

The numbers: thresholds, monthly limits and rate limits

The rate limits guide lists a Free tier plus three paid tiers. Free is not a paid tier. It requires an allowed geography.

| Tier | Total credit purchases | Monthly usage limit | | --- | --- | --- | | Free | None (allowed geography) | $100 | | Build | $5 | $500 | | Launch | $100 | $5,000 | | Grow | $500 | $200,000 |

Upgrades are automatic. Your organization's tier goes up as its total credit purchases reach each threshold.

Rate limits come in requests per minute (RPM) and tokens per minute (TPM), per model family:

| Tier | Model | RPM | TPM | | --- | --- | --- | --- | | Build | Astra, Sol, Terra | 5,000 | 1,000,000 | | Build | Luna | 5,000 | 2,000,000 | | Launch | Astra, Sol, Terra | 10,000 | 4,000,000 | | Launch | Luna | 10,000 | 10,000,000 | | Grow | Astra, Sol, Terra | 15,000 | 40,000,000 | | Grow | Luna | 30,000 | 180,000,000 |

Astra, Sol and Terra carry identical numbers, and Luna has its own row. The guide doesn't spell out that they draw from one pool. The Limits page does: models listed under a shared limit all count against the same rate limit.

Why this matters for sprint capacity

Tokens run out before requests do. On Build, 1,000,000 TPM divided by 5,000 RPM is 200 tokens per request. Any request longer than that hits the token limit first. The crossover is 400 tokens on Launch and about 2,700 on Grow for Astra, Sol and Terra. A chat or RAG call is almost always past that, so plan on TPM.

The jumps are uneven. Launch to Grow takes Astra, Sol and Terra from 4M to 40M TPM, and Luna from 10M to 180M. That is 10x and 18x for $400 more in total credit purchases. Build to Launch is 4x and 5x.

The monthly usage limit is a second ceiling. It sits next to the rate limits, and the guide notes it is separate from the spend limits you set for an organization or project. A workload can fit the TPM and still hit the monthly cap.

Which tier a pilot or production workload lands in

A pilot with $5 of credits purchased is on Build: $500 a month, 5,000 RPM, 1M TPM on Astra, Sol and Terra. That is plenty to prove a workflow works, and thin for a load test.

Production depends on your forecast. Estimate monthly spend and peak tokens per minute, then find the smallest tier that clears both. If you're between tiers, the purchase threshold is a budget line, so put it in the sprint plan before the launch sprint.

What to do on Monday morning

1. Open Settings > Organization > Limits and read the Rate limits section. Write down your tier, each model's RPM and TPM, and which models sit in a shared limit group.
2. Turn your traffic estimate into a number. This script takes the guide's figures and an average request size:

# Figures for Astra, Sol and Terra from OpenAI's rate limits guide.
# Replace them with the numbers on your own Limits page.
TIERS = {
    'build': {'rpm': 5_000, 'tpm': 1_000_000},
    'launch': {'rpm': 10_000, 'tpm': 4_000_000},
    'grow': {'rpm': 15_000, 'tpm': 40_000_000},
}

def sustainable_rpm(tier: str, avg_tokens_per_request: int) -> int:
    limits = TIERS[tier]
    return min(limits['rpm'], limits['tpm'] // avg_tokens_per_request)

for name in TIERS:
    print(name, sustainable_rpm(name, 3_000))
# build 333, launch 1333, grow 13333

At 3,000 tokens per request, Build sustains about 333 requests a minute, not 5,000. Use your own average and check how the guide counts tokens toward TPM.

3. Compare your forecast monthly spend with the monthly usage limit for your tier.
4. If you need more, use Upgrade tier in the Usage Tiers section of the same page. Upgrades follow total credit purchases, so buy ahead of a load test, not during one.

Gotchas

• Long context models, such as GPT-5.5 in the guide's example, have a separate rate limit for long context requests, visible in the developer console.
• Standard rate limits differ from Ultrafast rate limits.
• Spend limits you configure don't change your monthly usage limit.

Three tiers make capacity planning easier to explain, but a launch that stalls on a tier limit is still a production incident, so read your Limits page before the sprint starts.