Which model should write, and which should optimize?

6 brand-new Dr. Pia keywords, every model on the identical prompt, brief and facts. Graders were blind to the model. A win means the article beats the #1 ranking page on NeuronWriter and has zero serious fact errors.

Keywords

None had a page yet

Articles written

388

388 scored on NeuronWriter

Models in the data

Across all three rounds

Empty outputs

Billed, no article text

Current leaders, straight from the scores

Round A: Writer

Grok 4.7

87.9 NeuronWriter, 18/18 beat the #1 page

Best blind quality: Claude Sonnet 5.5 (80.9)

Full leaderboard

Round B: Optimizer

Astra

91.0 NeuronWriter, 18/18 beat the #1 page

Best blind quality: Claude Sonnet 5.5 (79.2)

Full leaderboard

Round C: Full recipe

Claude Opus 5.5 then Claude Opus 5.5

85.2 NeuronWriter, 12/12 beat the #1 page

Best blind quality: Claude Opus 5.5 then Claude Opus 5.5 (81.2)

Full leaderboard

Leaders only count models with at least half as many scored articles as the fullest model in that round. NeuronWriter scores move a few points between runs of the same model, so gaps under about 3 points are noise.

Written Oct 18 for the first four models: Claude Opus, Claude Sonnet, GPT and DeepSeek

Recommendation: Sonnet writes, GPT optimizes

Writing alone: Opus scores highest on NeuronWriter, Sonnet reads best to the blind graders.

Optimizing alone: Opus and GPT tie on score, GPT keeps quality higher.

The full recipe (Round C): all four combinations land within 1 point of each other on both score and quality, inside the run-to-run noise. The fact guard and polish steps even out the model differences.

Pick: Sonnet writes, GPT optimizes, then the Sonnet fact guard and polish. It ties for the top score, has the steadiest results (smallest spread), and splits the load across the Claude and ChatGPT subscriptions. DeepSeek is cheap and beats the #1 page on raw score, but it lands last on quality, misses the length target, and returned empty articles, so it is not ready for client work.

The models added since then (Grok, GLM, Kimi, MiniMax and the rest of the roster) are scored in the leaderboards and were not part of this recommendation.

The test setup

Keywords

Keyword	Searches / month	#1 words	#1 NW
can you drink coffee with invisalign	720	539	51
do veneers stain	590	1033	67
zirconia vs porcelain crowns	1900	717	61
why are my teeth transparent	880	1027	42
can you get veneers with crooked teeth	390	3341	64
veneers vs crowns	9900	554	70

Rounds and grading

A. Writer: same brief, only the writer changes. 6 keywords, every model, 3 runs each.

B. Optimizer: same starting draft (Sonnet run 1) for every optimizer. 6 keywords, 3 runs each.

C. Full recipe: the top writers paired with the top optimizers, then the same fact guard and polish. 6 keywords, 2 runs per pairing.

Scored on: NeuronWriter vs the #1 page, Surfer SEO and Surfer AI Search, Frase, length vs the #1 page, house-rule checks, TextRazor entity coverage, and a blind grader (fact errors, quality 0 to 100, readability 1 to 5).

Cost: subscription models ran with no API fees. Only API-billed models show a dollar figure.

Models and how each one ran

Model	Access	Rounds with scores
Claude Opus 5.5	Claude subscription	Round A, B, C
Claude Sonnet 5.5	Claude subscription	Round A, B, C
Claude Fable	Claude subscription	Not scored yet
GPT 6.1 Sol, low and high effort	Codex subscription	Round A, B, C
GPT 6 Astra	Codex subscription	Not scored yet
GPT 5.6 Terra	Codex subscription	Not scored yet
Kimi K3	Kimi Code	Round A
Grok 4.7	grok.com login	Round A, B
GLM 5.3	Z.ai coding plan	Round A
MiniMax M3	MiniMax API key	Round A, B
DeepSeek V4 Pro	OpenRouter	Round A, B

Round C pairings count under both of their models. Scores for models marked not scored yet show up in the leaderboards when they land.