OpenAI Gateway API Pricing: Real Cost Examples

The price in one table

If you searched for openai gateway api pricing, here is the answer for this service up front. It bills per token, in two lines, from a balance you fill in advance. There is no monthly fee, no seat price and no minimum commitment beyond the first top-up.

Item	Rate
Input tokens (your prompt)	$0.15 per 1M
Output tokens (the reply)	$0.50 per 1M
Failed or refused requests	Not charged
Credit expiry	Never
Subscription	None

One thing to be clear about: this is a hosted endpoint for a single open-weight text model, reached with the model ID uncensored. It is not the official API of OpenAI or any other vendor, so the rates above are what this service charges and say nothing about what other providers charge. For anyone else's numbers, check their current pricing page.

How a bill is calculated

Every request is metered in two parts. The prompt you send, including the system message and the whole conversation history, counts as input. The text the model writes back counts as output. The response includes a usage object with both numbers, so you never have to guess.

The formula is short:

cost = (input tokens ÷ 1,000,000 × 0.15) + (output tokens ÷ 1,000,000 × 0.50)

Output is a little over three times as expensive as input, per token. That shapes how you save money. Cutting a long prompt helps, but capping rambling replies often helps more per token removed.

The context window is 100,000 tokens, prompt and completion together, and max_tokens is optional. If you leave it out, a single reply can in principle use the remainder of the window. For a chatty model on an open-ended prompt, a sensible max_tokens is the cheapest insurance you can buy.

Worked examples

Each example below shows the arithmetic, so you can swap in your own numbers.

One chat turn

A support-style exchange with a 2,000-token prompt and a 500-token reply:

• Input: 2,000 ÷ 1,000,000 × $0.15 = $0.0003
• Output: 500 ÷ 1,000,000 × $0.50 = $0.00025
• Total: $0.00055 per turn

Divide $20 by $0.00055 and you get roughly 36,363 such turns. The $0.50 trial credit covers about 909 of them.

A batch job

You want to summarise 5,000 documents. Each one is 3,000 tokens long and the summary is 400 tokens.

• Input: 5,000 × 3,000 = 15,000,000 tokens, so 15 × $0.15 = $2.25
• Output: 5,000 × 400 = 2,000,000 tokens, so 2 × $0.50 = $1.00
• Total: $3.25

A small app for a month

Say your product handles 1,000 requests a day for 30 days. Each request has 1,500 tokens in and 300 out.

• Requests: 1,000 × 30 = 30,000
• Input: 30,000 × 1,500 = 45,000,000 tokens, so 45 × $0.15 = $6.75
• Output: 30,000 × 300 = 9,000,000 tokens, so 9 × $0.50 = $4.50
• Total: $11.25 for the month

A prompt that fills the window

The worst case for one call is a prompt near the 100,000-token limit. With 100,000 tokens in and a 1,000-token reply: 0.1 × $0.15 = $0.015 for input, plus 0.001 × $0.50 = $0.0005 for output, which is $0.0155. Even a full window costs about a cent and a half.

Why long conversations cost more than they look

Chat APIs are stateless. To keep a conversation going, your app sends the full history with each new message. So the input for turn 30 includes turns 1 through 29 again, and the bill grows faster than the number of messages suggests.

Take a roleplay or companion session where the history has grown to 20,000 tokens, and the user sends 100 more messages at that size. Each reply is 300 tokens.

• Input: 100 × 20,000 = 2,000,000 tokens, so 2 × $0.15 = $0.30
• Output: 100 × 300 = 30,000 tokens, so 0.03 × $0.50 = $0.015
• Total: $0.315

Notice that input dominates here, even though output is pricier per token. Three habits keep that in check: summarise old turns into a short memory note, drop messages that no longer matter, and keep the system prompt tight, since it is resent every time.

Top-ups, bonuses and payment methods

You add credit in any whole amount from $20 to $500. Larger top-ups earn a bonus on the amount you add:

Top-up	Bonus	Credit you receive
$20	None	$20.00
$50	+5%	$52.50
$100	+10%	$110.00
$200	+15%	$230.00
$500	+15%	$575.00

You can pay with crypto (USDT on TRC20, or USDC on Base). Credit does not expire, so a larger top-up for the bonus does not put you on a clock. The pricing page shows the current tiers.

Is there a free tier?

Not an unlimited one, and it is better to say so plainly. What exists is a one-time trial: after you sign up with Google, or with an email and password, you can claim $0.50 of credit with one click on the API key page. It needs no card, lasts 7 days and is limited to one per person.

At the rates above, $0.50 is enough to run an honest test. Using the chat turn from earlier, it covers around 909 requests, which is plenty to check streaming, tool calling, JSON mode and how the model handles your actual prompts. Trial keys have lower caps than full keys: 2 concurrent requests and 60 per minute, compared with 8 and 300.

To keep going after the trial, top up from $20. Nothing renews automatically, because there is no subscription to lapse.

Ways to lower the bill

Because the price is linear in tokens, every saving is easy to reason about. These are the changes that tend to pay off, roughly in order of effort.

1. Trim the system prompt. A 1,200-token system prompt sent with 10,000 requests is 12,000,000 input tokens, which is $1.80 by itself. Halving it saves $0.90 on that run.
2. Cap the reply length. If the answer should be one paragraph, say so in the prompt and set max_tokens. Output is the dearer side of the rate.
3. Window the history. Send the last N messages plus a short summary instead of the full log.
4. Ask for structure. A JSON reply with three fields is shorter than a prose explanation of the same data.
5. Do not retry blindly. A failed request is not charged, but a successful request that you repeat for no reason is.

It also helps to estimate before you build. Multiply your expected daily requests by the average tokens in and out, convert each to millions, and apply the two rates. If the result is a few dollars a month, as in the small-app example above, spend your time on quality instead of on squeezing tokens. If it is hundreds of dollars, the trimming steps above deserve a sprint.

Tracking and controlling spend

Because you prepay, the worst outcome of a bug is an empty balance, not a surprise invoice. Even so, it helps to watch the meter. Each response carries an X-Balance-USD header with your remaining credit. This Python snippet reads it and also works out the cost of the call from the usage numbers.

Copyfrom openai import OpenAI

client = OpenAI(base_url="https://api.openaigatewayapi.com/v1", api_key="paste-your-key-here")

raw = client.chat.completions.with_raw_response.create(
    model="uncensored",
    messages=[{"role": "user", "content": "Give me three names for a build-status bot."}],
    max_tokens=150,
)
resp = raw.parse()

cost = resp.usage.prompt_tokens / 1_000_000 * 0.15 + resp.usage.completion_tokens / 1_000_000 * 0.50
print("this call: $%.6f" % cost)
print("balance left:", raw.headers.get("X-Balance-USD"))

A few controls worth adding to any production integration:

• Set max_tokens on every call where a long answer would be a bug.
• Log the usage numbers next to the X-Request-Id header, so a cost spike can be traced to a request.
• Alert when X-Balance-USD drops under a threshold you choose.
• Stay inside the caps: 8 concurrent requests and 300 per minute on a full key. A 429 response with code rate_limited or concurrency is not billed, but a retry storm wastes your time.

For the request format and every parameter, see the documentation. If you are comparing this against running your own proxy layer, the write-up on LLM gateway costs and complexity covers the other side of that decision.

More guides

• OpenAI API for Developers: Use a Compatible Endpoint
• OpenAI Gateway API Explained in Plain Terms
• LLM Gateway: Cost vs. Complexity Analysis