GPT API pricing [2026 edition] — model-based rates and the actual cost of one request

GPT API pricing is pay-as-you-go per token. There is no monthly fee. Input and output tokens have separate rates, with output costing more. As of October 4, 2026, Kunavo rates per 1M tokens are gpt-5-6-terra at $0.70 input / $4.20 output, and gpt-5-5 at $2.00 / $12.00 — both below OpenAI's list prices. gpt-6-luna, released on September 22, 2026, costs $0.04 / $0.20 and is the cheapest GPT on Kunavo.

There is one distinction to make first: ChatGPT's monthly plans and the OpenAI API are separate products with separate billing. Paying for ChatGPT does not give you an API key or reduce your API bill. This page covers the latter only: pricing for calls made from code.

Rates verified on October 4, 2026. Kunavo's figures are rendered from this site's catalog, so they match actual charges. OpenAI list prices are based on OpenAI's pricing page.

GPT-5 / GPT-6 Series Rate Table

USD per 1M tokens. “OpenAI list price” is the published rate for the same model. For models currently on promotion, the prices OpenAI actually charges are also listed below.

The latest rates are always shown on the pricing page and each model page. Find the full model list in the model catalog.

GPT-6 is OpenAI's new generation. GPT-6 Astra launched as OpenAI's flagship model on September 3, 2026, followed by GPT-6 Sol and GPT-6 Luna on September 22, 2026. OpenAI describes the two as built on “the advances behind GPT-6 Astra,” positioning Sol for complex coding and agent workflows, and Luna as “the most efficient model for focused, high-volume tasks.” Both have a 1,050,000-token context window and a maximum output of 128,000 tokens.

On Kunavo, GPT-6 Luna is the cheapest GPT at $0.04 / $0.20. GPT-6 Sol costs $0.80 / $4.00, with a lower output rate than GPT-5.6 Terra ($4.20). However, the input rate is lower for GPT-5.6 Terra ($0.70), and for requests where input tokens exceed 2 times the output tokens, GPT-5.6 Terra is cheaper (you can compare them in the table below). Requests with more than 272,000 input tokens are charged at 2× the input rate and 1.5× the output rate for the entire request, as specified by OpenAI, just like other GPT-5.x / GPT-6 models.

What does one request actually cost?

Use this as is to calculate costs with your own numbers.

# Kunavo の GPT 料率（1M トークンあたり USD）: (input, output)
RATES = {
    "gpt-5-6-terra": (0.70, 4.20),
    "gpt-5-5":       (2.00, 12.00),
    "gpt-5-6-sol":   (2.00, 12.00),
    "gpt-6-sol":     (0.80, 4.00),
    "gpt-6-luna":    (0.04, 0.20),
}

def cost(model: str, in_tokens: int, out_tokens: int) -> float:
    i, o = RATES[model]
    # 注意: GPT-5 系の推論モデルは reasoning トークンも出力料率で課金される。
    return in_tokens / 1_000_000 * i + out_tokens / 1_000_000 * o

print(cost("gpt-5-6-terra", 1_000, 300))   # 短い 1 ターン
print(cost("gpt-5-6-terra", 6_000, 500))   # RAG の回答 1 件
print(cost("gpt-5-5", 20_000, 3_000))      # 難しい問題
print(cost("gpt-5-6-sol", 20_000, 3_000))  # GPT-5.6 のフラッグシップ
print(cost("gpt-6-sol", 20_000, 3_000))    # GPT-6 Sol
print(cost("gpt-6-luna", 1_000, 300))      # いちばん安い 1 ターン

Two factors that can skew your bill

The first is reasoning tokens. GPT-5 series reasoning models generate internal reasoning tokens before producing an answer. These tokens do not appear in the response body, but they are billed at the output rate. The table above excludes reasoning tokens, so the actual bill will be higher for more difficult tasks. The output token count in the usage field is the actual number.

The second is retries. Requests that consume tokens and then fail downstream are still billed. For agent workloads, one task from the user's perspective may involve multiple round trips for billing, so this can create the biggest gap. Requests that fail on Kunavo's side (that is, before reaching the upstream provider) are not billed.

Connect your existing OpenAI code

Just change one line, base_url, and replace your key. Both chat.completions and responses work the same way.

from openai import OpenAI

client = OpenAI(
    api_key="sk-kn-...",
    base_url="https://api.kunavo.com/v1",   # 変えるのはこの 1 行だけ
)

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

The same key also works with Claude, image, and video models. See the Quickstart for instructions and the Chat Completions documentation for the endpoint list.

Paying from Japan

Kunavo balance supports cards including JCB (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, Google Pay, and Link at Stripe checkout. It uses a prepaid wallet model, with a minimum top-up of $10 and no balance expiration. There is no charge in months when you do not make calls. Convenience-store payments and PayPay are not supported. Information required for a domestic business is provided in the disclosure under the Specified Commercial Transactions Act, and billing details are compiled in the billing documentation.

Four ways to lower costs

1. Choose a tier based on the size of the task. Do not default to the top-tier model. For high-volume tasks such as classification, extraction, and formatting, start with the least expensive gpt-6-luna that OpenAI positions for those use cases, then move up if needed. As shown in the table above, even for a single RAG answer, costs differ by a factor of about 53 between gpt-6-luna and gpt-5-5.
2. Limit output. Output costs more than input, so max_tokens and output format instructions have a direct effect. For models that use reasoning, this has a double impact.
3. Make use of prompt caching. If you send the same system prompt every time, caching is the biggest lever. See the caching documentation for details.
4. Do not accumulate irrelevant history in the input. Clear the conversation when moving to a different task. Long histories are billed in full on every call.

To compare with other providers' models, see Claude pricing, which uses the same calculation method. For Codex CLI pay-as-you-go pricing only, see Codex pricing.

Frequently asked questions