ChatGPT API Costs 2026 — How OpenAI API Billing Works, GPT-6 Pricing, and Actual Amounts

The ChatGPT API is the OpenAI API: it's billed per token and has no monthly fee. As of October 5, 2026, OpenAI's latest flagship, GPT-6 Astra, is officially priced at $10.00 input / $50.00 output per 1M tokens; on Kunavo it's $4.00 / $20.00. The everyday workhorse GPT-5.6 Terra costs $0.70 / $4.20 (officially $2.00 / $12.00); the new GPT-6 Sol and GPT-6 Luna, released on September 22, cost $0.80 / $4.00 and $0.04 / $0.20, respectively. Output costs more than input, and reasoning tokens for reasoning models are also billed at the output rate.

The point that causes the most confusion: ChatGPT subscriptions and the OpenAI API are billed separately. A ChatGPT Plus subscription doesn't include an API key or API credits. To call the API from code, you need to apply for a separate key and pay separately per token. This page covers only the latter—calling the API from code.

Rates verified on October 5, 2026. OpenAI's official API prices are listed at openai.com/api/pricing, and ChatGPT plan prices at chatgpt.com/pricing (plan prices change often, so this page doesn't repeat them). Kunavo's figures are drawn from this site's model catalog and match the amounts actually billed by the API.

ChatGPT subscription vs. OpenAI API—two separate bills

When is a subscription a better deal? Every $20 buys about 1,600 turns on GPT-5.6 Sol or about 4,600 turns on GPT-5.6 Terra (2,000 input / 700 output tokens per turn). If you mainly chat heavily in ChatGPT and use more than that per month while staying within the plan's limits, a monthly plan is a better deal. If you're calling from code or your usage varies widely, paying per token costs less.

GPT API pricing table—per 1M tokens

GPT-6 Astra is OpenAI's latest flagship, announced on September 3, 2026. Its output rate on Kunavo is about 1.7 times that of GPT-5.6 Sol, so reserve it for problems cheaper models can't handle instead of using it as the default. GPT-5.5 and GPT-5.6 Sol have the same per-token price (cache writes are calculated differently; see below). One more thing to know: when called through Kunavo, the GPT series currently accepts only text input, so use Claude for tasks that require image input.

GPT-6 Sol and GPT-6 Luna are two models OpenAI released on September 22, 2026. OpenAI says both build on GPT-6 Astra's technical advances; each has a 1,050,000-token context window and a 128,000-token maximum output. OpenAI positions GPT-6 Sol for complex coding and agentic workflows; on Kunavo, it costs $0.80 / $4.00. GPT-6 Luna is described by OpenAI as its most efficient model for focused, high-volume tasks; on Kunavo, it costs $0.04 / $0.20 and is the cheapest GPT in this table. New projects don't have to start with GPT-5.6 Sol: GPT-6 Sol on Kunavo costs $0.80 / $4.00, with both input and output rates lower than GPT-5.6 Sol's $2.00 / $12.00.

What does one request actually cost?

The rate table is less useful than the cost of an actual request. Everything below is calculated from the Kunavo rates in the table above:

GPT-6 Sol or GPT-5.6 Terra?GPT-6 Sol's output rate is lower than Terra's, but its input rate is higher. That means Terra is cheaper for requests with lots of input and little output—such as the RAG response and 100,000-item batch above. The higher the share of output, the more economical GPT-6 Sol becomes.

Two things can make the actual amount higher than the table shows. First, reasoning tokens: reasoning models generate thinking tokens before replying. These don't appear in the reply, but are billed at the output rate, so the difference grows with the difficulty of the task. Second, retries: failed requests aren't billed, but if a request succeeds and you ask again because the format is wrong, each request costs money. Check usage in the response to see the actual amount, then calculate your own figures:

# OpenAI API 的費用，用實際 token 數算。
# Kunavo 費率，每 1M token USD：(input, output)
RATES = {
    "gpt-6-astra":   (4.00, 20.00),
    "gpt-6-sol":     (0.80, 4.00),
    "gpt-6-luna":    (0.04, 0.20),
    "gpt-5-6-sol":   (2.00, 12.00),
    "gpt-5-6-terra": (0.70, 4.20),
}

def cost(model, inp, out):
    i, o = RATES[model]
    # reasoning token 不會出現在回覆裡，但按「輸出」費率計費
    return inp / 1e6 * i + out / 1e6 * o

print(cost("gpt-5-6-terra", 6_000, 500))       # RAG 回答一次
print(cost("gpt-6-astra", 20_000, 3_000))      # 難題一次
print(cost("gpt-6-luna", 500, 20) * 100_000)   # 10 萬筆批次

Long prompts and caching: two rules that change the rate

Requests exceeding 272K tokens. GPT-5.5, GPT-5.6 Sol and Terra, and GPT-6 Astra, Sol, and Luna—in other words, every model in the table—are billed at 2× input / 1.5× output for the entire request when its prompt exceeds 272K tokens. This is OpenAI's published long-context rule, and Kunavo applies it the same way, so the discount relative to the official price stays the same on either side of the threshold. For long documents that can be split, try to keep each part below 272K.

Caching. Cache reads cost 10% of the input rate; cache writes cost GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol, and GPT-5.6 Terra: 1.25 times the input rate and GPT-5.5 at the standard input rate. For workflows that send the same long system prompt every time, this is the biggest lever. See the caching documentation for how it works.

Switching from existing OpenAI code

Kunavo is an OpenAI-compatible endpoint, so you don't need to change SDKs: update base_url and the key, and both chat.completions and responses continue to work as before.

from openai import OpenAI

client = OpenAI(
    api_key="sk-kn-...",
    base_url="https://api.kunavo.com/v1",  # 只改這一行
)

resp = client.chat.completions.create(
    model="gpt-5-6-terra",
    messages=[{"role": "user", "content": "你好"}],
)
print(resp.choices[0].message.content)

Full steps are in the quickstart; endpoint specifications are in the Chat Completions documentation (English).

To be clear: when it makes sense to go directly to OpenAI

The gateway uses shared capacity, with no dedicated quota or contractually guaranteed SLA. For a scale that requires guaranteed quota, a specified rate limit, or an SLA, contract directly with OpenAI. If you only chat in ChatGPT and don't write code, you also don't need the API—as calculated above, a monthly plan is usually a better deal for heavy chat use.

Paying from Taiwan

Kunavo balances use Stripe and support cards including JCB (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, and Google Pay. It is prepaid, with a minimum top-up of $10; balances do not expire, and failed requests are not charged. The more you top up at once, the more bonus you receive: $100 credits $110, $1,000 credits $1,200, and $5,000 credits $6,250.

Taiwan has no local payment methods—JKoPay and LINE Pay aren't on the supported list. For team use, each key can have a monthly spending limit (returns 402 when reached) and an IP allowlist configured in /app/keys; costs can be split by person. See the billing documentation for details.

How to lower your GPT API costs

1. Choose the model for the task size. For simple, high-volume tasks such as classification, extraction, and routing, try GPT-6 Luna first. Use Terra or GPT-6 Sol for everyday work, and move up to Astra only for tasks they can't handle—the table above shows that the same task can cost up to 100 times as much.
2. Keep output short. Use max_tokens and output format constraints to limit length; lower reasoning effort for tasks where you can.
3. Use caching effectively. Put the fixed system prompt first and avoid changing its content between calls; cache reads cost only 10% of the input rate.
4. Keep prompts below 272K tokens. Above that, the whole request is billed at 2× input / 1.5× output.

To estimate a full month's costs, use the OpenAI API pricing calculator (English). If rate limits, rather than cost, are the issue, see OpenAI API rate limits (English). For the same kind of comparison with other providers, see Claude costs and subscriptions. If you use these GPT models with Codex, see Codex costs for how subscriptions and API keys are billed.

Frequently asked questions