OpenAI Usage Tiers Drop to Three, and the Decisions API Opens: What Agencies Should Check

OpenAI usage tiers went from five paid levels to three on October 6, 2026: Build, Launch and Grow. The top tier, Grow, now opens at $500 in total credit purchases, down from $1,000. The same day, OpenAI opened the Decisions API in public beta.

Both changes matter to agencies that run API accounts for clients. The new OpenAI usage tiers set how much each client account can spend in a month and how fast it can send requests. The Decisions API is a cheaper way to run the yes, no and pick-one jobs CMS sites send to models. This post covers the limits, what to check per client, how Anthropic and Google compare, and when the Decisions API beats a chat request. Facts are as of October 10, 2026.

The short version: existing paid organizations were moved to the new OpenAI usage tiers automatically, and no action is required. Still, check each client’s tier and monthly usage limit, because a small account on Build stops at $500 a month. The Decisions API is a good fit for routing, flagging and scoring on GPT-6 Luna, at $0.10 per million input tokens and no output charge. It is a beta, so keep a fallback.

What changed on October 6

OpenAI’s API changelog has two entries for October 6. The first says OpenAI "Simplified API usage tiers from five to three: Build, Launch, and Grow." It adds that organizations upgrade automatically as total credit purchases reach each tier’s minimum.

The second says OpenAI "Released the Decisions API in beta with gpt-6-luna." At DevDay on September 29, this API was in limited preview. Our OpenAI DevDay 2026 recap told readers not to build client work on it yet. It is now a public beta.

An announcement in the Announcements section of OpenAI’s developer forum, posted October 6, says paid organizations were moved to their new tier automatically. It says no action is needed. It also says the $500 Grow threshold is down from $1,000 for the old top tier.

The new OpenAI usage tiers and their limits

OpenAI’s rate limits guide lists the new OpenAI usage tiers. It says "The three paid usage tiers are Build, Launch, and Grow." A free tier sits below them.

The guide lists no waiting period. Under the old system, according to AlphaSignal’s coverage, the five tiers opened at $5, $50, $100, $250 and $1,000, and the upper four also required 7 to 30 days since the first payment. We could not confirm the old table on an OpenAI page, so treat those figures as reported.

The guide gives example standard limits in requests per minute (RPM) and tokens per minute (TPM).

The guide says these "differ from Ultrafast rate limits." Each organization’s real limits per model are on its limits page.

What happens to existing client accounts

OpenAI says paid organizations moved to the new OpenAI usage tiers automatically. It has not published a table that maps old Tiers 1 to 5 onto the new names. Neither the changelog nor the forum post mentions downgrades. Check each account yourself.

Two details matter for agencies. Rate limits are set at the organization and project level, not per user or per key, so client sites in one organization share a pool. And the monthly usage limit is separate from any spend limit you set.

The monthly usage limit is the number to watch. A client on Build stops at $500 a month in API usage, whatever its traffic does. Reaching $100 in total credit purchases moves an organization to Launch and a $5,000 cap. Reaching $500 moves it to Grow.

While you are in each organization’s settings, check key lifetimes too. Our post on OpenAI API key expiration, published earlier today, covers that step.

What agencies should check on each client account

Check OpenAI usage tiers for each client organization once, then record the results in your account notes.

1. Which tier the client is in. Open the organization’s limits page and record the tier name and monthly usage limit.
2. Last month’s spend against that cap. If a client used $400 last month on Build, one busy week can hit the $500 cap.
3. Total credit purchases. A client at $80 in purchases is $20 away from Launch. That is often cheaper than an outage.
4. Peak requests and tokens per minute. Compare your busiest minute with the RPM and TPM for the model you call.
5. Shared projects. If several sites share one organization, their traffic adds up against one limit.
6. Retry handling. Make sure your code reads the `x-ratelimit-remaining-requests` and `Retry-After` headers and backs off.

For spend by project and model, our post on the April 2027 OpenAI API shutoff covers the usage reports.

Rate-limit headroom for chatbots and site search

Under the new OpenAI usage tiers, most nonprofit and association sites will not hit 5,000 requests a minute. Tokens per minute is the tighter limit. Each chatbot turn sends the system prompt, the chat history and any retrieved page text. That can reach a few thousand tokens per request.

At 3,000 tokens per request, Build’s 1M TPM on Sol allows about 330 requests a minute. On Luna, at 2M TPM, it allows about 660. That covers a normal day, but a registration opening or news spike can push past it.

Site search that sends queries to a model works the same way. Retrieved passages raise the token count fast. Check the limits for every model the feature calls, including embeddings.

For low-risk tools, Build is usually enough per minute. The monthly $500 cap is the more likely problem. For a chatbot that a client depends on during events, reaching $100 in total credit purchases moves it to Launch, with four times the TPM on Sol.

What the Decisions API does

The Decisions API guide describes one endpoint, POST /v1/decisions. It runs only on gpt-6-luna. You send an input and a list of questions. Each question has a type.

• Predicate. Checks a condition and returns a probability from 0 to 1 that it is true.
• Choice. Picks one of your listed options and returns a probability for each, plus a `confidence` field.
• Score. Rates the input against levels you define, such as low, medium and high, with a `confidence` field.

Input can be plain text or text with images, sent inline as base64 data. The guide says "Hosted HTTP or HTTPS image URLs and file_id inputs aren’t supported by this endpoint." Several independent questions can go in one request. The response comes back as an answers array, and a refusal answer type can appear.

The guide says input costs "$0.10 per 1M tokens," with no charges for output tokens, cache reads or cache writes. Regional processing premiums and long-context multipliers still apply. OpenAI says the API answers about 10 times faster than the Responses API. That is OpenAI’s figure, so test it on your own traffic.

The guide says the API "is in public beta, and we expect to GA in the coming weeks." It supports Zero Data Retention and HIPAA use for eligible customers, with data residency in the US and Europe.

When the Decisions API beats a chat request

The Decisions API wins when the answer is a fixed choice and you want a probability with it. Many CMS jobs fit.

• Routing webform submissions. A choice question sends each contact form to membership, events, billing or “other.”
• Flagging comments and listings. A predicate question checks whether a forum post or job listing breaks the client’s posting rules.
• Scoring submissions. A score question rates conference proposals or member resources against a short rubric before staff review.
• Checking uploaded photos. A predicate question checks whether a directory photo shows one person’s face.

The probabilities are the main gain. You can auto-approve above 0.9, send 0.5 to 0.9 to a person, and reject below that. With a chat request, you get a label and have to guess how sure the model was.

The cost gain is smaller than it looks. GPT-6 Luna on the Responses API costs $0.10 per million input tokens and $0.50 output. A 20-token label adds about a fifth to the cost of a 500-token form. Reasoning tokens also bill as output, and our April shutoff post notes that GPT-6 models default to medium effort. Set effort low and the gap stays small. Our GPT-6 Sol and Luna post covers Luna pricing in more detail.

Stay with the Responses API when you need your own JSON shape or tool calls. The guide says to use Structured Outputs or function calling for those jobs. It also says answers that depend on an earlier answer need separate requests.

Decisions API or the free Moderation endpoint

For general harm checks, start with OpenAI’s free endpoint. The moderation guide says "The moderation endpoint is free to use." It checks text and images for harassment, hate, self-harm, sexual content, violence and illicit activity.

Use the Decisions API for rules that belong to one client. "Is this post a sales pitch?" and "Does this listing include a salary range?" are not harm categories. They are house rules, and they fit a predicate question.

A practical setup runs both: Moderation for harmful content at no cost, Decisions for the client’s own rules, and a human queue for anything in the middle.

Treat the beta like any new service. Log each answer, set a timeout, and fall back to a person or to your old chat request if it fails. For a comparison point at the same price, see our post on Claude Haiku 5.5, which matches Luna’s list price.

How Anthropic and Google set their tiers

OpenAI usage tiers are one of three systems agencies deal with. The names and rules differ, so check each provider’s own page.

Anthropic’s rate limits page does not list a dollar threshold for each tier. New organizations may start in an Evaluation tier with lower limits. "Build" is Anthropic’s second tier and OpenAI’s first, so do not compare by name.

Google’s Gemini API rate limits page counts Google Cloud spend on the linked billing account. Its limits apply per project, not per API key, and exact numbers appear in AI Studio.

Frequently Asked Questions