Findings

Measured claims, with receipts.

Each finding states one claim, shows the numbers behind it, and explains how we measured it.

Published findings

What our measurements showed

PublishedAugust 5, 2026

GPT-5.6 Luna matched Sol on our hardest question class

In our per-class evals for a production analytics workload, GPT-5.6 Luna matched GPT-5.6 Sol's accuracy on the hardest question class. It did so at a fraction of the price and about half the latency.

Read the finding

Follow along

New findings, when we have them

We publish when a measurement says something worth repeating, so the feed is quiet by design. Subscribe by RSS, or ask for them by email and a person will add you to the list.