Only the good stuff, picked for you.
Ranked on what each model is for rather than where it sits on a leaderboard. Fourteen of the twenty carry prices read off the provider own page, with the date we opened it printed beside every number. The rest have no published figures we could read, and say so instead of carrying a guess.
The middle of OpenAI three tiers, and the one we would default to. Output costs a fifth of the model above it and twenty times the model below it, which puts it in the band where you stop deciding per request and just use it.
Anthropic prices this as the middle of its range and describes it as the speed-and-intelligence compromise. Against the model above it in the same family, output costs a fifth as much.
The cheapest way onto Google current generation, and the price is explicitly temporary: it roughly doubles on 1 January 2027. Worth knowing before you build a budget on it.
Off-peak pricing is the interesting part: the same request costs half as much outside the busy window. If your work is a batch job rather than an interactive one, scheduling is a bigger lever here than model choice.
Anthropic positions this for long-running agentic work, which is a different job from answering one question well. Twice the input price of the tier below and twice the output.
The cheap tier: ten cents per million input tokens is low enough that filtering a large pile of text first and only escalating the interesting parts becomes the obviously correct architecture.
Google describes this as built for high-volume agent work and translation. The input price covers text, image, video and audio at one rate, which is unusual and makes it easy to budget.
Cheaper per output token than anything else on this list, and it keeps the same million-token context as the model above it. The trade is documented by the provider itself: no vision, and thinking mode is the default rather than optional.
The top of Anthropic range, described as being for demanding reasoning and long-horizon agent work. At $50 per million output tokens it costs twenty-five times the model three tiers down.
The top tier, and priced like one: fifty dollars per million output tokens, five times the middle tier and a hundred times the cheap one. Worth it only where being wrong is the expensive part.
Anthropic prices this one in two bands by prompt length, which catches people out: ten cents up to a hundred thousand tokens, five times that above it. Do the arithmetic before pointing a long pipeline at it.
Still on the price list a generation later, at twice the input price of the current middle tier and twice the output. The honest use is as the benchmark you price the newer models against.
The most expensive thing OpenAI publishes for text: $12.50 in and $75 out, and it has no long-context row at all, which suggests the window is deliberately narrow. A specialist, not a default.
The Pro tier charges by prompt length as well: two dollars per million tokens up to two hundred thousand, then double. Its long-context output is $18, which is the number to compare against anything advertising a large window.
The family that most often ends up running on hardware you own. Weights are published, sizes range from something a laptop can hold to something that needs a rack, and the licensing is permissive enough that the question is capability rather than permission.
The open family with the widest tooling around it: more fine-tuning recipes, more quantised builds and more inference servers treat it as the default. If you want the shortest path from weights to something running, start here.
A European provider whose selling point is less about the top of the benchmark table and more about where the weights and the data live. Worth shortlisting when residency or procurement rules rule out the American labs.
Built its reputation on long inputs before long inputs were the thing everyone advertised, and it is the Chinese assistant most often compared with the American ones on document work.
Chinese open-weight models with a published API and aggressive pricing. The practical reason to look at it is cost per token in a deployment where the data is staying inside the country anyway.
The one model family wired directly into a social feed, which is its whole argument: it can answer about what is happening right now rather than what was indexed last month. That same wiring is a reason to keep it out of anything sensitive.
The ranking is ours. Anything on this page that can be checked is checked against a named source, and the date we last opened that source is printed with it. If a figure and our verdict disagree, trust the figure and follow the link.
A model is listed only once its provider publishes pricing and documents an API. Every figure carries a source and the date we checked it; the ranking is ours.