How to read a model announcement
A reviewer's pass over a launch post, in six steps. What to trust, what to skip, and where the real claim is hiding.
What the announcement is actually selling
A model announcement is written by the people who made the model. That is not a flaw — it is the reason the desk reads them at all — but it means the post is a sales document with a benchmark table inside it. The facts are in there. They are just arranged to make the model look good.
The skill is not skepticism for its own sake. It is knowing which parts of the post are claims you can check, which are claims about how the lab measured itself, and which are decoration. The six steps below are the pass a careful editor runs before anything gets filed.
A reviewer's pass, in six steps
Find the primary link first
Before reading any summary, open the lab's own post, paper, or release note. If the announcement has no primary link, that is the story: there is nothing to check yet.
Read the capability claim literally
Strip the adjectives. "Best-in-class reasoning" is marketing; "beats the previous model on AIME 2025 by 3 points" is a claim. Write the claim down in your own words, using only what the post states.
Ask who chose the benchmark
Labs compare against the models that make them look strongest, on the tasks their model is best at. A benchmark list is data, not a verdict. If the comparison omits the obvious rival, that omission is the headline.
Check what the numbers are measured on
Reported scores can come from a tuned variant, a single run, or a temperature you will never use. Look for the words about how it was evaluated. Their absence is a detail, not a secret — note it and move on.
Separate availability from announcement
Waitlist, coming soon, and available today are three different news items. The desk files the one that is real on the day it is filed.
Keep the claim attached to its source
Whatever you carry away, carry the link with it. A claim without its source becomes a rumor within two retellings, and no one will be able to tell you which of the three versions was the real one.
The claim is always in the primary source. Everything else is a summary of a summary.
What to do with the rest
Most of the post — the quotes from executives, the partner logos, the vision paragraphs — is not for you. It is for investors and journalists on deadline. Let it go. The two sentences that survive your pass are the ones worth anything, and they are the two the desk will file.
Run this pass on three announcements and it stops feeling like work. You start reading the benchmark table for what it is and the prose for what it is doing, and the announcements get a lot shorter.