What will GPT-OSS 120B cost?

How to read the comparison

Start with the number of API calls, rather than the number of people using your app. One conversation or agent task can make several calls. Include any earlier messages or tool results sent again in the input-token estimate. For longer workflows, use the chat and agent token estimator first.

The table compares published token charges for the same named model on four providers, one of them our own Standard API. Matching model names do not establish identical deployments, output lengths, quality, latency, or service limits. This is a cost estimate, not a performance benchmark or a provider ranking.

What the cache assumption means

“Cached input” is the percentage of total input tokens you assume will actually be billed at a cached-input rate. It is not the percentage of requests that share a prompt, and the calculator does not predict whether a cache hit will occur.

The same assumption is applied to providers with a verified cached-input rate so that you can inspect the pricing difference. Actual cached shares can differ between providers. Check each provider’s matching rules and eligibility before using this scenario as a budget. When a cached rate is unverified, the estimate uses the normal input rate and labels that limitation.

The calculation

For each provider:

calls = requests per day × days
input = calls × average input tokens
output = calls × average billed output tokens
cached input = input × assumed cached-input share

cost = (uncached input × input rate
      + cached input × cached-input rate
      + output × output rate) / 1,000,000

Rates are in USD per million tokens. Calculation uses unrounded values; only the displayed currency is rounded. Reset restores the example workload. Its values are planning assumptions, not observed customer usage.

What to include in your token counts

Use the provider’s billed output usage, including reasoning tokens when they are charged as output. Include repeated calls and retries in request volume. The estimator cannot infer these from the text visible to an end user.

This comparison covers input, cached-input, and output token charges. It excludes taxes, credits, negotiated discounts, tool execution, storage, networking, and any separately charged cache-write or retention fees. Batch, priority, reserved capacity, and subscriptions need their own rate cards.

From an estimate to a budget

Run a representative workload with the provider you are considering, then replace the example values with measured token usage and the cached share in your billing records. Check success rate and response time alongside cost. A low price per token does not by itself tell you the cost of completing a task successfully.

The Standard API serves gpt-oss-120b at the rate in the table. Log in to run the same workload against it.