How much does the OpenAI API cost?

OpenAI's API is metered per model, per million tokens, with separate input and output rates like every other major provider. What makes it harder to estimate than most is the sheer number of live models: several GPT generations, a parallel o-series of reasoning models, mini and nano variants of each, plus embedding models priced on an entirely different scale. Picking the wrong one by habit — defaulting to whatever you used last year — is the most common way teams overpay.

GPT tiers vs the o-series

The GPT-numbered models are OpenAI's general-purpose line: full, mini and nano sizes trading capability for cost, aimed at chat, drafting and agentic tool use. The o-series is priced and positioned for harder reasoning tasks — multi-step math, complex planning — and historically carries a premium output rate to match the extra compute spent thinking before it answers. If your task doesn't need that depth, a GPT-tier mini model is very likely cheaper and fast enough.

Embeddings are priced on a different scale entirely

Embedding models have no output tokens — you pay only for input — and the per-token rate sits far below any chat model, because you're paying for a vector, not a generated response. If your pipeline is doing retrieval-augmented generation, the embedding cost of indexing your corpus is usually trivial next to the generation cost of answering with it; don't spend estimation time there.

Legacy tiers are still live, and still billed

OpenAI keeps older GPT generations reachable well after a newer one ships, marked legacy rather than pulled outright, which means it's easy to keep paying an older tier's rate out of habit long after a cheaper or better-priced current model would do the same job. If a model in your codebase hasn't been revisited since it was the newest option, that's worth a five-minute check against the current lineup — legacy status is a strong hint the model has since been superseded on price, capability, or both.

Cached input and batch: the two discounts worth building around

OpenAI automatically discounts input tokens that match a recently-seen prefix — repeated system prompts, tool schemas, or long context reused across a session — at a materially lower rate than fresh input, with no code change required on your end beyond keeping that prefix stable. The Batch API is the second lever: submit a job asynchronously and get results within a defined window at a flat discount versus the synchronous API, on both input and output tokens.

• If your app re-sends the same system prompt or tool definitions on every call, structure requests so that block stays identical — that's what makes it cache-eligible.
• If a job doesn't need a response in the same request cycle (batch scoring, offline summarization, eval runs), route it through the Batch API rather than the synchronous endpoint.
• Nano and mini variants exist specifically for high-volume, latency-tolerant tasks — classification, tagging, short extraction — where the full-size model's extra reasoning is wasted spend.

See the live pricing table above for exact current rates — OpenAI has repriced its lineup more than once in the last year, so treat any number you remember as provisional.

A rough monthly estimate

Multiply average input tokens per request by your input rate, average output tokens by your output rate, sum them, then multiply by requests per day and roughly 30 for a month. That single-request cost is the number worth optimizing before you scale traffic — a 10x cheaper model choice at the prototype stage is worth far more than a 10x traffic optimization after launch. Meterlark's cost calculator runs this for every current model side by side, including cached and batch rates, so you can compare a model swap before you commit to one.

One more habit worth building in: recheck the estimate once real usage exists, not just at the planning stage. Actual prompt lengths, actual completion lengths and actual request volume rarely match the numbers used to plan a launch, and OpenAI's own pricing moves on its own schedule regardless of what you assumed when you shipped — see why AI API prices keep changing for how often that's happened across the market.