Series · field notes
A new model lands. Within three days there are twenty projects on GitHub, a hundred posts about them, and almost no way to tell which of it is real. So we read the projects themselves — the code, the measurement files, the limits sections that nobody quotes — and write down what holds up. Every note says what we ran and what we did not.
This series has one entry. It will grow when the next model does, and publishing an entry costs one line on this page — no new navigation, no new card, no new anything.
Read 2026-09-20
Eleven repositories, one shape: the model picks, and the code owns everything else. The famous seven-second browser demo was mostly browser engineering — the median improved 9.450 s to 7.092 s, p = 0.25, and protocol calls fell 1,092 to 101.
Every entry here is a reading of primary sources: repositories, their own measurement files, their own limits sections, and the vendor's documentation. Vendors' claims are labelled as vendors' claims wherever they appear. Where we could not check something — because a harness needs a browser we would not attach to, or a key we would not spend — the entry says so in plain words rather than leaving the gap for a reader to discover.
That is the whole editorial rule, and it is the same one the rest of this site runs on: publish a date you can point at, or publish nothing.
primary sources only — repositories, model cards, vendor documentation We do not republish other people's summaries; every claim is read from the file it came fromWhat we measured ourselves is stated per entry, and what we did not run is stated too The other half of the pair: what it is, what it costs per million tokens, the three question types, and our own raw response.
The other way a model arrives on OpenRouter — anonymous and free, gone in days. Same index-and-entries shape as this one.
Every model we track, with price and context read from the public catalogue, and a dated provenance line under each table.