OpenAI API Pricing (October 2026): GPT-6 and GPT-5.x Token Costs

The OpenAI API charges per million tokens (MTok), with separate prices for input, cached input and output. As of October 17, 2026, the main models are GPT-6 Astra ($10 / $50), GPT-6.1 Sol ($2 / $10) and GPT-6 Luna ($0.10 / $0.50). The price you actually pay also depends on how you send requests: Standard, Batch, Flex, Fast or Ultrafast, and whether your prompt is over 272K tokens. This page explains each of these, with worked examples using the rates on OpenAI's official pricing page.

The API is billed separately from ChatGPT. If you want a monthly subscription for chatting or coding rather than per-token billing, see ChatGPT pricing and OpenAI Codex pricing.

How OpenAI API billing works

Every request is metered in tokens. A token is a chunk of text, roughly three-quarters of an English word on average. You pay for:

• Input tokens: everything you send, including the system prompt, chat history, tool definitions and any retrieved documents.
• Cached input tokens: input that matches a prefix OpenAI already processed recently. These are billed at a much lower rate.
• Output tokens: everything the model generates, including hidden reasoning tokens. OpenAI's reasoning guide says reasoning tokens are billed as output tokens and count toward `max_output_tokens`.
• Tool calls: built-in tools such as web search, file search and code containers have their own per-call or per-session fees on top of tokens.

New API accounts use prepaid billing. According to OpenAI's help center, the minimum credit purchase is $5 (the default is $10), auto-reload is switched on by default during setup, and purchased credits expire after one year. When your balance hits zero, requests start failing with a `credit_balance_exhausted` error. Credits are non-refundable except where the law, a contract or an approved exception applies.

Current model prices (Standard tier)

These are the Standard rates per 1M tokens for prompts up to 272K input tokens.

The newest models are not always the most expensive. GPT-6.1 Sol costs half as much as GPT-5.6 Sol on input and output, and less than GPT-5.5 or GPT-5.4. Unless you have tested an older model and found it works better for your task, the GPT-6 family is usually the better value. Older models such as gpt-4.1, gpt-4o, o3 and o4-mini are still on the price list too.

Processing tiers: Standard, Batch, Flex, Fast and Ultrafast

The same model can cost very different amounts depending on how quickly you need the answer.

Some example rates: GPT-6.1 Sol is $1 / $5 on Batch and Flex, $4 / $20 on Fast, and $12 / $60 on Ultrafast. GPT-6 Astra is $20 / $100 on Fast and $60 / $300 on Ultrafast.

The Batch API guide describes a 50% discount compared with synchronous APIs. A batch can hold up to 50,000 requests in an input file of up to 200 MB, and it completes within 24 hours. Requests that are still unfinished when the window closes are cancelled, but you still pay for the ones that completed. Output files are deleted 30 days after the batch completes.

With Flex you get Batch prices on ordinary Responses or Chat Completions calls, but you trade away reliability. If there isn't enough capacity, you get a 429 "Resource Unavailable" error, and the Flex guide says you are not charged when that happens. You can then retry later or fall back to Standard by setting service_tier to "auto".

Fast mode is the new name for Priority processing, changed on July 30, 2026. Both service_tier values still work. A project can make Fast the default in its settings, so check that setting if your bill is unexpectedly about twice what you estimated.

Long-context pricing (over 272K input tokens)

Prompts of up to 272K input tokens count as short context. Above that, a higher long-context rate applies:

The pricing page defines the two tiers by input size but doesn't say explicitly whether the higher rate applies to the whole request or only to tokens above 272K. The examples below assume the whole request is billed at the long-context rate, which is the more conservative estimate.

Prompt caching: the biggest discount most apps can get

Caching is on by default for supported models. When the start of your prompt (the system prompt, tool definitions, a long document) matches a recent request, those tokens are billed at the cached rate. The prompt caching guide says the minimum cacheable prefix is 1,024 tokens for GPT-5.6 and later models. A cached prefix stays reusable for 30 minutes after its last write or reuse.

• Cache reads cost 0.1x the input price on GPT-5.6 and later models, and 0.05x on GPT-6.1 Sol. That is why GPT-6.1 Sol's cached input is $0.10 while GPT-6 Sol's is $0.20.
• Cache writes on GPT-5.6 and later cost 1.25x the uncached input price. This is not an extra fee on top: each input token is billed at exactly one of the three rates (uncached, cached or write).
• Earlier models do not charge extra for cache writes.
• Put content that stays the same (instructions, examples, tool schemas) at the start of the prompt and content that changes (the user's question) at the end. Only matching prefixes are cached.

Worked cost examples

All examples use the Standard short-context prices above unless stated otherwise, and assume no tool calls. Real bills depend on your actual token counts, which you can read from the usage object of each response.

1. A support chatbot: 100,000 requests a month

Each request sends 2,000 input tokens and gets 500 output tokens back. Over a month that is 100,000 × 2,000 = 200M input tokens and 100,000 × 500 = 50M output tokens.

The jump from Luna to Astra is 100x. Many teams route simple questions to Luna and only send hard cases to Sol or Astra.

2. The same chatbot with caching

Say 1,500 of the 2,000 input tokens are a fixed system prompt and knowledge snippet that hits the cache on GPT-6.1 Sol. Uncached input is 100,000 × 500 = 50M tokens × $2 = $100. Cached input is 100,000 × 1,500 = 150M tokens × $0.10 = $15. Output stays at $500. The total is $100 + $15 + $500 = $615, down from $900, a saving of about 32%. Cache writes (billed at 1.25x, so $2.50 per MTok) add a little each time the cache has to be refreshed, which is why steady traffic benefits most.

3. Classifying 1 million documents overnight with Batch

Each document is 800 input tokens and the label is 50 output tokens, on GPT-6 Luna. That is 800M input and 50M output tokens. On Standard: 800 × $0.10 = $80 plus 50 × $0.50 = $25, for $105. On Batch: 800 × $0.05 = $40 plus 50 × $0.25 = $12.50, for $52.50. One batch holds up to 50,000 requests, so this job needs at least 20 batch files.

4. One long-context request

You send a 400,000-token codebase to GPT-6.1 Sol and get a 5,000-token review. At the long-context rate: 0.4M × $4 = $1.60 input plus 0.005M × $15 = $0.075 output, for about $1.68. The same token counts at the short-context rate would be 0.4 × $2 = $0.80 plus 0.005 × $10 = $0.05, or $0.85. Trimming the prompt below 272K tokens (for example by sending only the relevant files) roughly halves the cost of each call.

5. A reasoning-heavy agent task

An agent step on GPT-6 Astra sends 20,000 input tokens and returns 1,000 visible tokens, but used 9,000 reasoning tokens to get there. Output is billed on 10,000 tokens, not 1,000: 0.02 × $10 = $0.20 input plus 0.01 × $50 = $0.50 output, for $0.70 per step. If you only counted the visible answer, you would have estimated $0.25. Lowering the reasoning effort setting is often the quickest way to cut agent costs.

6. Add-ons: web search and data residency

Web search costs $10 per 1,000 calls plus the search content tokens at the model's rate. 10,000 searches a month is 10 × $10 = $100 before tokens. If you use a regional processing (data residency) endpoint for a model released on or after March 5, 2026, add 10%: the $900 chatbot in example 1 becomes $900 × 1.10 = $990.

Tools, images, audio and embeddings

OpenAI is also winding down its fine-tuning platform. Existing fine-tuned models stay available for inference until their base models are deprecated.

Usage tiers and spending limits

Your rate limits and monthly usage cap go up automatically as your total credit purchases grow. The rate limits page currently lists these tiers:

The usage limit is a ceiling on spending, not an allowance of free usage. To control costs, set an auto-reload recharge limit in billing settings. OpenAI notes that a recharge limit only caps automatic purchases, not your use of credits you already hold.

Which model should you pick?

• Prototyping or high-volume simple tasks: start with GPT-6 Luna. At $0.10 / $0.50, a million short requests usually cost tens of dollars.
• Most production apps and coding agents: GPT-6.1 Sol. It has the best cached-input discount in the lineup ($0.10), which matters for agents that resend long context.
• Hardest problems where quality matters more than cost: GPT-6 Astra, ideally only for the steps that need it.
• Overnight jobs, evaluations, data enrichment: any of these on Batch or Flex for about half price.
• Legacy GPT-5.x models: only if you have tested them and they beat GPT-6 on your task. Most cost more than GPT-6.1 Sol.

How it compares with Claude and Gemini

Per-token prices are only a rough guide, because models use different tokenizers and different amounts of reasoning on the same task. As a starting point, compare GPT-6.1 Sol and GPT-6 Luna with the mid-tier and small models on our Claude API pricing and Gemini API pricing pages. Google's Gemini API also has a free tier for some Flash models, covered in Gemini API free tier, while OpenAI has no free token allowance.

Ways to pay less

1. Structure prompts so the fixed part comes first and gets cached.
2. Move anything that can wait up to 24 hours to Batch, and use Flex for evaluations.
3. Keep prompts under 272K input tokens where possible.
4. Cap output with max_output_tokens and use lower reasoning effort for routine steps.
5. Check that your project isn't set to Fast mode by default unless you need it.
6. If you are a VC-backed startup, check OpenAI for Startups for possible API credits, and see our guide to getting free AI credits for other programs.

Turn off auto-reload or set a recharge limit before you share an API key or run an agent in a loop. Leaked keys and runaway loops are the most common way a small test turns into a large bill.

Recent pricing changes

• July 30, 2026: Priority processing renamed Fast mode. The "priority" service_tier value still works.
• September 29, 2026 (DevDay): GPT-6.1 Sol launched in the API at $2 / $10, with $0.10 cached input. Astra Ultrafast became available in the Responses API.
• Through at least November 21, 2026: GPT-5.6 Sol is on promotional pricing, so expect its rate to change after that.
• Regional processing endpoints carry a 10% uplift for models released on or after March 5, 2026.

For background on the company behind the API, see who owns OpenAI and ChatGPT. If you only need a free key for experiments, our list of free AI API keys covers providers with real free tiers.