GPT-4o Mini

On a typical 70% input / 30% output mix, GPT-4o Mini blends to $0.29 per million tokens. That puts it above 9% of the 32 models tracked here, and 3 other models in the same budget tier undercut it. Tier is a capability label, not a price band — the spread inside one tier is often wider than the gap between tiers, which is why picking by tier alone tends to overspend.

Input runs $0.15 per million and output $0.60 — a 4.0x ratio. That ratio matters more than either number on its own: a summarisation or classification workload reads far more than it writes and will track the input price, while a code-generation or long-form writing workload inverts that and will track the output price. Compare models on the ratio your own traffic actually has, not on the input price alone.

The benchmark profile is uneven rather than flat: data extraction at 95/100 against coding at 74/100, a 21-point spread. An average hides that. If your workload sits on the strong end this model punches above its price; if it sits on the weak end, a cheaper model with a flatter profile will serve you better.

The context window is 128K tokens and the throughput score is 97/100. Context only earns its keep if you fill it, and filling it is also what makes a request expensive — a large window is an option, not a discount. For a budget-tier model, judge the speed score against what you are doing: it decides everything for interactive chat and almost nothing for overnight batch work.

Worth checking before you commit: Gemini 2.0 Flash scores at least as well on combined coding and reasoning and blends to $0.19 per million against GPT-4o Mini's $0.29. That does not make GPT-4o Mini the wrong choice — context window, latency, vendor lock-in, and specific capabilities all sit outside a benchmark score — but it does mean the price difference needs a reason behind it.

The closest comparison is Llama 3 70B from Meta, at $1.28 per million against $0.29. 7 models in this tier score higher on combined coding and reasoning, which is the useful framing: the question is rarely whether a model is good, it is whether it is the cheapest thing that is good enough for the specific work you are sending it.