What this test cannot tell you yet

Read these before acting on any single number. Scores land as the graders finish, and this page is rebuilt from the latest data.

Entity coverage is partial

TextRazor's free daily quota (500 requests) ran out mid-test, so some articles show "pending" for entity coverage.

GLM hit a rate limit

GLM 5.3 on the Z.ai coding plan was rate limited, so it has far fewer articles than the other models. Its averages rest on a small sample.

MiniMax returned empties

MiniMax M3 returned empty outputs on some runs, so it has fewer articles than the full 18 in the writer round.

DeepSeek returned empty articles

DeepSeek V4 Pro returned 13 empty articles (billed, no text) because its reasoning used the whole output budget. They were re-run with a larger budget and the failures are counted as empty outputs.

DeepSeek key was missing for part of the test

The DeepSeek key was missing during part of the test. Any gaps in DeepSeek coverage come from that, not from the model.

Frase column is topic coverage, not the Frase editor score

The only Frase key on file is a legacy key, and the current Frase scoring API needs a newer sk_live key. So the Frase column is computed from Frase's own SERP topic list (top 50 topics per keyword, pulled from the legacy Frase API): each topic counts up to the average uses on ranking pages, weighted by how many ranking pages use it. The crooked teeth keyword was queried as "veneers for crooked teeth" because Frase read the leading "can" as the band Can. Models land between 45 and 54, too close to change any ranking.

Gemini was not part of this test

Gemini was left out on purpose.

Two recipe skills came from a copy

Two skills from the writing recipe (koray-semantic-seo-master and neuronwriter) are missing from the active skill shelf. This test used the copies on the F drive, and the same brief went to every model.

NeuronWriter moves between runs

NeuronWriter scores shift a few points between runs of the same model. Differences under about 3 points are noise.