Price, context, and benchmarks side by side

A comparison only exists where the underlying numbers differ. Figures we could not confirm from the provider are shown as Not verified rather than estimated, because a wrong price is more damaging than a missing one.

Benchmark results carry their source and its publication or capture date. That date is not necessarily the date a benchmark was run; check the linked source for explicit measurement details.

Frontier tiers earn their price on hard reasoning and long agentic tasks. For classification, extraction and summarization — most production volume — a budget tier is usually indistinguishable in output and several times cheaper.

Output tokens cost three to five times input across almost every provider here, so generation length drives the bill far more than prompt size.