Models ranked
Blended price uses your input/output ratio. Value score = intelligence ÷ blended price, normalised so the leader is 100. Intelligence is a reference index, not an authoritative benchmark.
| Model | In / Out $/1M | Blended $/1M | IQ* | Value | Your $/mo |
|---|
Headline price hides the real decision
Comparing LLMs on their input price per million tokens is how most cost tables work, and it's misleading in two ways. First, output tokens usually cost three to five times as much as input, so a model that looks cheap on input can be expensive for a workload that writes long answers. Second, price says nothing about capability — the point of the model. A tool that costs a tenth as much but fails a third of your tasks isn't cheap; it's expensive once you count the retries and the human cleanup. Price-to-performance folds both problems into one number: a capability score divided by the blended price of your actual token mix, so you compare value, not just cost.
That reframing changes decisions. For high-volume, forgiving work — classification, tagging, first-draft extraction — the best-value pick is often a small or open model that's nearly as good as frontier for a fraction of the price. For work where a mistake is costly — production code, legal or medical drafting, agent actions — the value math rightly favours a stronger model, because its higher score outweighs the premium. The smart pattern most teams land on is to route easy requests to the value leader and reserve a frontier model for the hard ones. Once you've picked, size the real bill on the LLM token cost calculator and cut it further with prompt caching.
How to use it
1. Enter the typical input and output tokens for one request — this sets your real cost mix.
2. Optionally add monthly requests to see each model's dollar cost for your workload.
3. Choose whether to rank by value, raw price or intelligence, and read the table. The value leader is the model giving the most capability per dollar for your token shape.
Common mistakes
Judging on input price alone. Output-heavy work is priced by the output rate — set your real ratio. Chasing the absolute cheapest. If it fails your task, the rework costs more than the tokens saved. Over-paying for a frontier model everywhere. Most requests don't need it; reserve it for the hard slice. Trusting one score. The intelligence index is a starting point — the only benchmark that counts is your own evaluation set on your own prompts.
FAQ
What token ratio should I use?
Chatbots and Q&A are often close to even; extraction and classification are input-heavy (long context, short answer); writing and code generation are output-heavy. Use your logs if you have them; otherwise start near 3:1 input:output and adjust.
Where does the intelligence score come from?
It's an approximate composite of where each model generally sits across public reasoning, coding and knowledge benchmarks in 2026. It's meant to rank tiers directionally, not to certify one model as smarter by a point.
Why isn't the cheapest model always the value leader?
Because value is capability divided by price. A model that's 40% cheaper but 40% weaker has similar value; one that's slightly pricier but meaningfully more capable can lead. The ranking captures that trade-off for your mix.
Estimate only. Intelligence index is a directional reference; validate on your own tasks before choosing.