Estimate your monthly OpenAI GPT API spend by token usage and request volume.
| Model | $/1M input | $/1M output | Context |
|---|---|---|---|
| GPT-5.6 | 5.00 | 30.00 | 1M |
| GPT-5.6 Terra | 2.50 | 15.00 | 1M |
| GPT-5.6 Luna | 1.00 | 6.00 | 1M |
| GPT-5.4 mini | 0.75 | 4.50 | 1M |
| GPT-5.4 nano | 0.20 | 1.25 | 1M |
The mini and nano tiers cost an order of magnitude less than flagship on input — route classification, extraction and simple tasks there and reserve flagship for hard reasoning. Cached input and batch APIs can cut costs further for non-urgent or repeated-context workloads.
OpenAI's line-up spans a wide range, and the spread between the flagship and the smallest tier is large enough to change what is economically possible. A workload that is unaffordable on the flagship model is often routine on a mini or nano tier — which makes model selection the first cost decision, not an optimisation to revisit later.
The practical approach is to split by difficulty rather than picking one model for everything. Classification, extraction, routing, tagging and short summarisation rarely need frontier capability. Send those to a small tier, keep the flagship for reasoning that genuinely fails without it, and measure the quality difference on your own data instead of assuming it.
Reasoning tokens. Models that think before answering bill that thinking as output. A short visible answer can carry a much larger invisible cost, so estimates based on the length of the reply alone come out low. Where the API exposes a reasoning-effort setting, it is a direct cost control.
Conversation history. Chat requests resend the entire thread. The tenth message in a session costs far more than the first, even though it looks the same to the user.
Retries. Every retry after a timeout or a rate-limit rejection is billed like any other request. Aggressive retry logic quietly inflates bills during exactly the periods when things are already going wrong.
Cached input charges a reduced rate for prompt prefixes that have been seen recently. Anything that resends a fixed system prompt or a long reference document benefits, and it usually requires only reordering the prompt so the stable part comes first.
Batch submission trades latency for a lower rate on work that can wait — evaluations, backfills, bulk classification. For offline pipelines the latency costs nothing.
GPT-5.4 nano, at $0.2/$1.25 per million tokens. GPT-5.6 is the priciest listed here at $5/$30.
Use the Token Counter, then plug the numbers in above.
Roughly one token per four English characters, but the ratio is worse for code, JSON and non-English text, all of which fragment into more tokens than their length suggests. Measure a representative request with the Token Counter rather than trusting the rule of thumb for anything you plan to bill for.
For well-scoped tasks with clear inputs and outputs, frequently yes — and the cost difference is large enough that it is worth testing rather than assuming. Build a small evaluation set from real examples, run both tiers, and let the result decide. The answer varies by task far more than by model reputation.
Compare across providers: LLM API Cost Calculator · Claude · Gemini