The cheapest frontier LLMs

Per-token price is the most quoted and least useful number in this table. What you actually pay is price multiplied by tokens consumed, and modern reasoning models can emit five to twenty times as many tokens as a direct-answer model for the same visible output, because the hidden chain of thought is billed too. A model listed at a third of the price can easily cost more per completed task.

The second distortion is retries. A cheap model that gets a classification right 88% of the time, in a pipeline where an incorrect answer triggers a second attempt or human review, is competing against an expensive model at 97% with the full cost of that review loop attached. Work out the cost of a wrong answer before optimising the cost of a token.

Input and output are also priced differently, usually with output several times more expensive. Retrieval-heavy workloads that stuff long documents into context are dominated by input price, while generation-heavy work is dominated by output. Prompt caching changes the arithmetic again: if your system prompt is long and stable, a provider with aggressive caching can beat a nominally cheaper competitor outright.

Where cheap models genuinely win is high-volume, well-specified, verifiable work — classification, extraction, routing, summarising, tagging, first-pass triage — especially when a validator or schema catches failures automatically. The standard pattern is a cheap model doing the volume with a frontier model handling escalations, which typically lands within a few percent of frontier quality at a fraction of the spend.

Prices here are pulled from live provider listings and change without notice, so treat them as a shortlist tool rather than a quote. Run your real workload against two or three candidates, measure cost per successful outcome, and pick from that.

More on how these numbers are produced in the methodology and what each evaluation measures in the benchmark guide. Spotted a score that disagrees with its source? Tell us.