How models are ranked, and what they cost

A lab can be first on a coding test and ordinary on a fraud test. The honest sentence names the test, the date, and who measured it. “Best model” without those three is an advertisement.

Meta’s October 12, 2026 post compares Muse Wonder and Muse Wonder Full with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. Meta says the pair lead on agent, memory, and coding tests. Those comparisons are Meta’s, and some of the other labs’ scores in that card are Meta’s own runs. A fuller safety report was promised, not published with the post.

The rank that changes a team’s work is a held-out set of the actual job. Coinbase did not ask which model was famous. It asked which system caught Onramp fraud, on four detection metrics, at what delay. On that private test, a post-trained 9B model beat Claude Opus 4.5. The percentages are Coinbase’s. The test is not public, so another company cannot rerun it. Another company can copy the method.

Vendors bill input and output separately, per million tokens. Input is what you send, including the instructions and the retrieved notes. Output is what the model writes back, including the reasoning some models spend before the answer. A loop multiplies both, because every extra step is another call.

The table is the standard list price on each company’s own API, for a prompt inside that vendor’s short-context band, read on October 13, 2026. Cached input is cheaper. Batch is often half. A very long prompt can reprice the whole request. A cloud reseller can charge a different number. Use the linked page before you budget.