Which AI model is cheapest to run?
The price lists look built to answer that: every vendor publishes a rate per token, and comparing two rates takes ten seconds. The trouble is that cost per token and cost per task can rank the same two models in opposite orders.A taxi with a cheaper rate per kilometer can still be the more expensive ride home if it takes the longer route. The rate on the meter is real, but the fare depends on the route as well, and an AI price list only shows you the meter.
A token is a small chunk of text, the unit AI companies bill by. A short word is usually one token; a longer word might be two or three. I'm oversimplifying, but that's the scale of it. A model bills in both directions: one rate for the text you send, a higher rate for the text it writes back. Multiply tokens used by the price per token and you have the bill. That bill, for one finished job, is the job's cost per task.
The rate is only one of the two numbers on the bill, though. The other, how many tokens the job takes, is decided by the model itself. Ask two models the same question and one may think out loud first, writing a long chain of reasoning before it answers, while the other replies in three sentences. In taxi terms, the model is the GPS the driver follows, and it picks the route. Vendors even count tokens differently, so the same paragraph splits into more billable chunks with one company than with another.
How can the model with the pricier rate come out cheaper?
Say Model A charges a single all-in rate of $1 per million tokens and Model B charges $2, twice the rate. Give both the same job, and say Model A works wordily and takes 3 million tokens while Model B finishes in 1 million. Model A's bill is $3. Model B's is $2. The job cost a third less on the model that looked twice as expensive, because the token gap (three to one) was wider than the price gap (two to one).
The same flip sits in public pricing. At the time of writing, OpenAI's GPT-5.6 Sol lists at $4.00 per million input tokens and $20.00 per million output tokens[1], while Anthropic's Claude Sonnet 5 lists at $2.00 and $10.00[2] for the same two directions. On the price list, Sonnet 5 is the budget option, at half the rate both ways.
The order reverses when someone measures a whole finished job. Artificial Analysis, a company that measures AI models, runs the major ones through the same fixed set of test tasks (a benchmark suite) and adds up what each model spent to complete them[3]. On that suite, Sol finishes a task for $1.04[4] on average, measured before the August rate cut pushed it lower still (Artificial Analysis has not republished the figure at the new rate), and Sonnet 5 finishes the same tasks for $1.72[5]. The model with the higher sticker price did the same work about 40% cheaper.
Is that pair a fluke? The people who run public benchmarks have answered by changing what they measure. Artificial Analysis prices whole completed tasks, and its methodology page gives the reason[3]: "Models that produce longer answers or more reasoning tokens will have a higher cost per task, even at identical per-token prices."
Researchers have also measured how much two models can differ in the tokens they spend on the same work. OckBench, a benchmark built by a team from Georgia Tech, MIT and NVIDIA, found one model spending more than 26 times the tokens of another to reach the same accuracy[6]. That is one extreme pair, and it shows the scale: Sol and Sonnet 5's rates sit within a factor of two of each other, while OckBench measured token use twenty-six times apart.
Charts like these describe something you can experience directly, and I have. Artificial Analysis's tests also produce a quality score for each model, the Intelligence Index. OpenAI's GPT-5.6 Luna, another model in the same family, scores 51 to Sonnet 5's 55, and completes a task for $0.21[7] against Sonnet 5's $1.72, about an eighth of the cost for a model a few points behind — and the $0.21 predates OpenAI's 80% cut to Luna's rates in late July, so the gap is wider now. Run a pair like that on your own jobs and the gap moves from the chart to your monthly bill.
These measurements get published as charts, usually with the quality score on one axis and measured cost per completed task on the other. Start from a model you are considering. If another one on the chart scores higher on quality and also costs less per completed task, cross yours off, unless something practical keeps it in, like speed or where you are allowed to run it. Is that simple? Maybe.
That habit settles most ordinary cases. The numbers still come from someone else's tasks, and a different mix of work can move the figures and even the ranking. If your work looks nothing like a benchmark suite, run the comparison yourself on the jobs you actually care about. The chart gets you to a shortlist. After that, the only numbers that count come from your own jobs.
Which model wins on your tasks?
[1] OpenAI, "API Pricing", accessed August 2026, https://developers.openai.com/api/docs/pricing
[2] Anthropic, "Pricing" (Claude Developer Platform docs), accessed August 2026, https://platform.claude.com/docs/en/about-claude/pricing
[3] Artificial Analysis, "Methodology", accessed August 2026, https://artificialanalysis.ai/methodology
[4] Artificial Analysis, "Four frontier launches in eight days: six labs now field a model above 50 on the Artificial Analysis Intelligence Index", July 2026, https://artificialanalysis.ai/articles/four-frontier-launches-in-eight-days-six-labs-now-field-a-model-above-50-on-the-artificial-analysis-intelligence-index
[5] Artificial Analysis, "Claude Sonnet 5" model page, accessed August 2026, https://artificialanalysis.ai/models/claude-sonnet-5
[6] Georgia Tech / MIT / NVIDIA, "OckBench: Measuring Token Efficiency in Large Language Models", arXiv:2511.05722, https://arxiv.org/abs/2511.05722
[7] Artificial Analysis, "GPT-5.6 benchmarks across Intelligence, Speed and Cost", July 2026, https://artificialanalysis.ai/articles/gpt-5-6-has-landed