Here is the direct answer. As of our July 2026 check, OpenAI's flagship GPT-5.5 lists at 30 dollars per 1M output tokens, the mainstream GPT-5.4 at 15 dollars, and GPT-5.4 mini at 4.50 dollars, with input tokens billed separately at a fraction of those rates. The Batch API halves every one of those numbers for work that can wait up to 24 hours. What that means for your actual monthly bill depends entirely on the shape of your workload, so below the price table you will find worked math for three common apps: a support chatbot, a RAG app, and an agent.
One note on sourcing: these are the prices in our verified dataset, which we re-check weekly against official pricing pages for the live rankings. Numbers in blog posts age; the rankings do not.
OpenAI API prices in 2026
Per 1M output tokens, list versus Batch API:
| Model | List | Batch |
|---|---|---|
| GPT-5.5 | 30 dollars | 15 dollars |
| GPT-5.4 | 15 dollars | 7.50 dollars |
| GPT-4o | 10 dollars | 5 dollars |
| GPT-4.1 | 8 dollars | 4 dollars |
| o3 | 8 dollars | 4 dollars |
| GPT-5.4 mini | 4.50 dollars | 2.25 dollars |
Three structural things to understand before doing any math:
- You pay separately for input tokens (your prompt, history, retrieved documents) and output tokens (the model's reply). Output is the expensive side, typically several times the input rate.
- Reasoning models like o3 burn hidden thinking tokens that bill as output, so the sticker price understates the real cost per request.
- The Batch API is a flat 50 percent discount on everything asynchronous, and prompt caching cuts the input side sharply when requests share a long identical prefix.
Current reseller and batch prices per model live on the per-model pages, for example the GPT-5.5 price table.
Worked example 1: a customer support chatbot
Say you run a support bot doing 1,000 conversations a day, six turns each, with roughly 300 output tokens per turn. That is 180,000 model calls a month producing about 54M output tokens.
- On GPT-5.4 mini at 4.50 dollars per 1M output tokens, the output side is about 243 dollars a month.
- Input volume is much larger, because every turn resends the system prompt and the conversation so far, but input rates are a fraction of output rates. With a long cached system prompt, a realistic all-in bill lands in the 350 to 550 dollar range.
- The same workload on GPT-5.5 at list would put the output side alone at 1,620 dollars. This is why almost nobody should run a support bot on the flagship.
Chat is interactive, so batch does not apply, but prompt caching absolutely does. Structure the prompt static-first and the repeated prefix bills at the discounted cached rate on every turn.
Worked example 2: a RAG app
A retrieval app is input-heavy: every query stuffs retrieved chunks into the prompt. Take 50,000 queries a month with about 4,000 input tokens (system prompt plus chunks) and 500 output tokens each. That is 200M input and 25M output tokens.
- On GPT-5.4, the output side is about 375 dollars a month at list. On GPT-5.4 mini it drops to about 112 dollars.
- The input side matters more here than in chat, because retrieved chunks differ per query and mostly cannot be cached. Budget the input side as a meaningful addition, not a rounding error.
- The embeddings that power retrieval are a separate, much smaller line item, and the ingestion pass is mostly a one-time cost.
The practical RAG pattern in 2026: answer with mini, escalate to GPT-5.4 only when a confidence check fails. Most teams find well-retrieved context matters more than model tier for grounded answers.
Worked example 3: an agent
Agents are where bills surprise people, because one user action fans out into many model calls. Take an agent product doing 100 runs a day, 20 model calls per run, averaging 1,000 output tokens per call. That is 60M output tokens a month.
- All on GPT-5.5 at list: about 1,800 dollars a month on output alone, before the substantial input side, because agents resend growing context on every step.
- Split it, with the planner on GPT-5.5 and the tool-calling steps on GPT-5.4 mini, and the same run count typically lands under 500 dollars.
- If runs are triggered by schedules rather than live users, the Batch API halves whatever remains.
The lesson generalizes: agents need a model routing policy from day one, not as an optimization later.
The levers that cut the bill
In order of effort against savings:
- Batch everything asynchronous. Flat 50 percent off, results within 24 hours, no product changes. Mechanics in prompt caching and batch discounts.
- Cache your prompt prefix. Static content first, variable content last, byte-identical across requests.
- Right-size the model. GPT-5.4 mini at 4.50 dollars versus GPT-5.5 at 30 dollars is a 6.7x gap, and classification, extraction, and routine drafting do not need the flagship.
- Cap output tokens per call type and ask for structured output. Output is the expensive side; do not leave its length to chance.
- Route routine work off OpenAI entirely. Open-weight models on discount hosts serve bulk tasks for cents; the LLM API pricing comparison has the full map.
The complete playbook, with what each lever is worth on a real bill, is in how to cut your OpenAI API bill.
Do not pay cash before claiming credits
New OpenAI accounts and startup programs come with real API credit, and platform programs are worth five figures to qualifying companies. The options and claim steps are in OpenAI API free credits. We track these programs, along with 200 plus other verified startup perks, in the Perkstack catalog; membership pays for itself the first time a claim goes through. You can create an account in a minute.
Bottom line
The OpenAI API costs whatever your token mix makes it cost: 30 dollars per 1M output tokens at the GPT-5.5 flagship, 4.50 dollars on GPT-5.4 mini, and half of either through the Batch API. A support bot on mini runs a few hundred dollars a month, a RAG app similar, and an unrouted agent on the flagship can clear four figures fast. Do the token math for your own shape before committing, check current prices in the rankings since every number here drifts, and claim credits before optimizing.
Related reading: how to cut your OpenAI API bill, LLM API pricing comparison, the cheapest GPT-5.5 endpoint.