Which is cheapest: GPT-4o, Claude or Gemini? The short answer is that it depends on the task and the volume, not on the list price. One model can cost five times more per token than another and still come out cheaper in production if it solves the task with half the tokens or on the first try. Here are the prices side by side and, above all, how to read them without a nasty surprise on the bill.
API prices per million tokens
APIs charge per token, and almost all of them split the input price (what you send: your prompt, the context, the documents) from the output price (what the model generates). Output usually costs quite a bit more than input.
| Model | Input ($/1M) | Output ($/1M) |
|---|---|---|
| GPT-4o mini | 0.15 | 0.60 |
| Gemini 2.5 Flash | 0.30 | 2.50 |
| GPT-4.1 mini | 0.40 | 1.60 |
| Claude Haiku | 1.00 | 5.00 |
| Gemini 2.5 Pro | 1.25 | 10.00 |
| GPT-4.1 | 2.00 | 8.00 |
| GPT-4o | 2.50 | 10.00 |
| Claude Sonnet | 3.00 | 15.00 |
| Claude Opus | 15.00 | 75.00 |
Indicative prices as of August 2026; always check the official rates, which change often. Between the cheapest and the most expensive model in the table there is a factor of 100 on output. That gap is what makes picking the wrong model the most expensive mistake people make with AI in production.
The list price misleads: look at cost per task
The number that matters is not the one in the table, it is the cost per task: the tokens a real request consumes multiplied by their price, and that times how often you repeat it per month. An expensive model that answers well on the first try can come out cheaper than a cheap one that needs retries, longer prompts or corrections.
That is why the honest comparison is not “which costs less per token” but “which costs less per task solved”. A small, cheap model that solves 95% of your cases and hands the other 5% to a large one usually beats any large model used for everything.
How to choose between the three
For high-volume, low-complexity tasks —classifying, extracting data, summarizing short texts— the small models (GPT-4o mini, Gemini Flash, Claude Haiku) win almost every time: the cost per task is a fraction and the quality is more than enough.
For complex reasoning, code or long texts a large model pays off (Claude Sonnet or Opus, GPT-4.1, Gemini Pro), but only on the tasks that genuinely need it. The strategy that works best is routing: the small model on the front line and the large one only when required.
Watch the structural discounts. Prompt caching (paying once for the context that repeats) and batch processing (up to half price if the task can wait) change the comparison completely. A model with good caching can overtake one that is cheaper on the list.
Check it with your numbers
The table gives you the price; your bill is decided by your volume. Before committing to a model, put your real tokens and your number of tasks into the token cost calculator and compare each one’s monthly cost side by side. The result almost always surprises.
Frequently asked questions
Which is cheapest, GPT-4o, Claude or Gemini? By list price, the small models in each family (GPT-4o mini, Gemini Flash, Claude Haiku) are the cheapest, with GPT-4o mini in the lead. But the cheapest in your case depends on how many tokens each one spends solving your task, not just the price per token.
How much does GPT-4o cost per million tokens? Around $2.50 for input and $10 for output (indicative, August 2026). Its mini version drops to about $0.15 and $0.60.
Is the most expensive model per token the most expensive in production? Not necessarily. What decides the bill is the cost per task —tokens times price, times volume— so an expensive model that solves on the first try and with fewer tokens can come out cheaper than a cheap one that needs retries.
This piece is part of the cluster around cost per task of AI, the guide that anchors everything else.
At e-ficient we measure the cost per task of companies that already run AI in production, including how much each model would cost you at your real volume. The first audit is free if you start a plan with a three-month commitment and returns a diagnosis within 72 hours.