AI editorial benchmark: Sol versus Luna
On this page
On 22 October 2026 we gave GPT-6.1 Sol and GPT-6 Luna the same editorial commission: build a practical article cluster about AI model benchmarking. Each delivered one pillar, ten supporting articles and a complete editorial handoff.
Results and method
| Outcome | GPT-6.1 Sol | GPT-6 Luna |
|---|---|---|
| Frozen quality score | 92.05 / 100 | 85.93 / 100 |
| Production time, rounded | 18 min 18 sec | 9 min 28 sec |
| Total production tokens | 3,087,961 | 3,157,774 |
| Publication placement | Featured blog series | Comparison archive |
Limits: one run per model and a single blinded GPT-6.1 Sol judge, which shares the winning producer’s model designation. Scores are editorial judgements, not statistical proof of model superiority. No independent human validation or billed task cost was measured. The two presentation-order passes were by the same judge.
Featured series · GPT-6.1 Sol
- AI model benchmarking: a practical guide to choosing with evidence — master guide
- Benchmark AI latency and cost per successful task
- Build an evaluation dataset that resists leakage and contamination
- Choose LLM evaluation metrics and write a rubric that works
- Compare AI models fairly across prompts, tools and settings
- Design an LLM evaluation around tasks your users actually do
- Evaluate AI agents by outcomes, tool actions and recovery
- How to read an AI benchmark leaderboard without overreading it
- Human reviewers and LLM judges: build a grading process you can trust
- Is that benchmark improvement real? Repeats and uncertainty
- Keep an LLM evaluation suite useful after the first model choice
Runner-up series · GPT-6 Luna
Read all eleven Luna articles in the comparison archive. Each identifies its generating model and links to the original scored submission.
Download the evidence
- Shared production brief
- Rubric frozen before production
- Sol original submission: articles, briefs and handoff
- Luna original submission: articles, briefs and handoff
- SHA-256 hashes of every original submission file
- Complete review package, with A/B identity mapping
- Per-article scores
- Cluster scores
- 71-entry targeted fact-check ledger
- Production token usage and measurement notes
- Publication edit log
Cached input is included in input; reasoning output is included in output. Neither subset is added again. Total tokens include repeated context. Reviewer usage is separate overhead. Editorial effort ranges in the review are estimates, not observed labour.
Original submissions and publication copies
Scores refer to the frozen submissions, not the edited articles shown on the website. We preserved the original files before reviewing or editing. Publication copies include corrected citations and source summaries, clearer worked methods, website metadata and public links. The edit log records those differences. Original source registers remain historical evidence; the edit log records corrected source mappings for the publication copies.
The blind review’s recommendations and proposed publication plan describe its state at review time. They are retained as evidence and do not themselves authorise or certify a live website release. Downloading the originals allows readers to form their own judgement.