OpenAI API Pricing: Per-Model Cost Per Token

Compare the cost per token to the price you pay to ship the right-quality answer. These aren't necessarily the same.

Model	Input	Output	Batch in	Batch out	Cache read
GPT-4.1	$2.00	$8.00	$1.00	$4.00	$0.50
GPT-4.1 Mini	$0.40	$1.60	$0.20	$0.80	$0.10
GPT-4.1 Nano	$0.10	$0.40	$0.05	$0.20	$0.03
GPT-4o	$2.50	$10.00	$1.25	$5.00	$1.25
GPT-4o Mini	$0.15	$0.60	$0.07	$0.30	$0.07
GPT-5	$1.25	$10.00	$0.62	$5.00	$0.12
GPT-5 Mini	$0.25	$2.00	$0.12	$1.00	$0.03
GPT-5.1	$1.25	$10.00	$0.62	$5.00	$0.12
GPT-6 Astra	$10.00	$50.00	$5.00	$25.00	$1.00
o3	$2.00	$8.00	$1.00	$4.00	$0.50
o3 Mini	$1.10	$4.40	$0.55	$2.20	$0.55
o4 Mini	$1.10	$4.40	$0.55	$2.20	$0.28

All prices in USD per 1M tokens. Batch runs asynchronously at a reduced rate; cache read applies to reused input tokens. A dash means the provider does not publish that rate.

The spread from the cheapest model to the flagship is more than 20-fold, which makes model choice look like a pure cost lever. It’s not, here’s why.

The sticker price is not the real rate

Batch processing halves the standard rate, and cached input (a reused system prompt or document) costs a fraction of it. The headline number is where the estimate starts, not where the bill lands.

GPT-4o input: one headline rate, three real rates ($/1M tokens)

When every model is good enough, cost is the differentiator

All five models answered a sample support email task,[^task] and tied at 90 out of 100. With quality flat, GPT-4o cost 32 times more than Nano for the identical result. On an easy task, paying more buys nothing.

Same task, same quality (90/100), very different cost

Paying more does not always buy better quality

On a harder task with one right answer,[^task] the field splits. GPT-4o, GPT-5.1 and o4 Mini score 100; the two cheapest drop to 50 and 25. But mid-priced GPT-5.1 matches the flagship for less, so a higher price does not guarantee a better answer. Which model wins depends on the task, not its token price.

Harder task: quality and cost do not move together

The best model today is not the best model forever

Providers update these models constantly, so the model that clears your bar today can slip below it next week while a rival climbs past it. Measuring once picks the right model now; measuring on a schedule keeps quality high and cost down with no quiet regressions or cost creep.

The best model today is not the best model forever (illustrative)

Build the hygiene in from day one

If you are picking an OpenAI model before you have written the app, this is the moment to set up good model-selection and pipeline hygiene, while it costs nothing to change. The teams who stay cheap and reliable do four things from the start.

Define your quality bar first. Write down what a good answer looks like for your task, so “good enough” is a measured line, not a gut call.
Pick on evidence, not reputation. Score the candidate models on your own task and take the cheapest one that clears the bar, rather than defaulting to the flagship because the job feels important.
Measure the whole pipeline, not just the model. Your prompt, your context, and your parameters move quality as much as the model does; a diagnostic that names the weak layer saves you from blaming the wrong thing.
Re-check on a schedule. Providers change models and prices without telling you, so a standing benchmark keeps your choice current instead of quietly drifting.

Baked in at the start, this is a few minutes of setup. Bolted on after a surprise bill or a quality complaint, it is a cleanup job. The pricing table is where the decision starts; this is the habit that keeps it right.

What you get on Pro

Score any OpenAI model on your own task
Cost, quality, and latency side by side
Pipeline diagnostics that identify actionable improvements
Scheduled benchmarks alert you when models regress or drift
API access to integrate measurements into your applications

Drive it from your agent, over MCP

Let your agent run the comparisons and read the results for you
Work from where you already are: your IDE, chat, or console
Over 40 tools your agent discovers automatically, no integration code
Agentic recipes for model selection, drift checks, and optimization loops

$129/mo $10 for your first week

Then $129/mo. Cancel anytime.

The table lists the OpenAI models we support for evaluation, not OpenAI's full catalogue. Rates come from our live registry and are updated frequently. The scored runs are real: five models, two tasks, October 2026. Your own task will score differently, against your prompts and data.