AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: 71 AI Model Configurations Tested: Choosing One For My News Pipeline on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

  • Fast, free delivery for dorm and study essentials
  • Prime Video and Amazon Music included
  • Member-only deals
Try Prime for Young Adults Free trial for eligible 18–24 year olds
As an affiliate, we earn on qualifying purchases.

TL;DR

ThorstenMeyerAI.com reports testing 71 AI model configurations on news-writing tasks between Sept. 8 and Oct. 13, 2026. GPT-6 Astra received the highest scores, but GPT-6 Luna was chosen for production because it performed consistently at a much lower reported cost; two backup models were removed after producing fabricated quotes or unusable output.

ThorstenMeyerAI.com reported October 13 that it tested 71 AI model configurations on news-writing tasks and selected GPT-6 Luna for production, although GPT-6 Astra scored higher. The test’s reported cost and quality trade-off matters to publishers using AI at volume, while its findings remain specific to the site’s prompts, sources and evaluation method.

The test ran from September 8 to October 13, 2026, using the publication’s production prompt and two real news sources: a 6,340-character product filing and a 619-character political news item. Each source was turned into an English and German article. The configurations covered API and subscription models, free-tier models accessed through OpenRouter, and open-weight models running on the publisher’s hardware.

Two AI judges, GPT-6.1 Sol and Claude Opus 5.5, assessed the original source alongside the generated article. They scored source fidelity, fabrication, clarity about what was known, news value, structure and polish; German articles also received a language-quality score. The publisher said a separate robustness run used the strongest configurations to generate 10 additional articles, mostly from short sources. Results were reported with 95% bootstrap intervals, and an output counted only if it remained usable after the pipeline’s normal repair step.

GPT-6 Astra scored 4.00 out of 5 in both languages in the standard test, with a source-fidelity score of 5.00. Across 12 articles in the robustness run, its score was 3.93. GPT-6 Luna scored 3.67 in English and 3.68 in German in the standard test; its robustness-run score was 3.67, with source fidelity of 4.71 and valid output for every article. Meyer reported costs of about $0.12 per Astra article versus $0.0012 per Luna article, with average generation times of roughly 50 and 17 seconds, respectively.

At a glance
reportWhen: Published October 13, 2026; testing ran…
The developmentThorsten Meyer published results from a test of 71 AI configurations for a news pipeline, naming GPT-6 Luna as the production choice despite GPT-6 Astra’s higher scores.
Thorsten Meyer AI
Own benchmark · 71 configurations

Infographic 01 · Who writes the best news article?

The best writer is not the one in production

71 model configurations, one real job: turn a news source into an article for an English and a German edition. GPT-6 Astra writes the best articles. GPT-6 Luna writes them for about a hundredth of the price.

4.00GPT-6 Astra, English and German, with perfect source fidelity (5.00). About $0.12 per article, roughly 100× GPT-6 Luna.
3.67GPT-6 Luna at reasoning effort “none”, the model in production. $0.0012 per article, 17 seconds.
1.00Source fidelity of GPT-4.1 nano, the old second-choice fallback. It invented quotes and was removed on 7 Oct.

Robustness run: 12 real news sources, English edition

Dot = mean score (1–5), bar = 95% bootstrap interval. German edition mean below each score. Overlapping bars are not a real difference.

  1. GPT-6 AstraOpenAI3.93DE 3.86
  2. GPT-6.1 Sol · judgeOpenAI3.75DE 3.74
  3. GPT-6 Luna (high)OpenAI3.73DE 3.64
  4. GPT-6 Luna (none) ★OpenAI · in production3.67DE 3.62
  5. Claude Sonnet 5.5Anthropic3.65DE 3.59
  6. Muse WonderMeta3.58DE 3.41
  7. Claude Opus 5.5 · judgeAnthropic3.52DE 3.35
  8. Claude Haiku 5.5 (high)Anthropic3.47DE 3.50
  9. Grok 4.7xAI3.47DE 3.35
  10. Nex-N2 miniNex · local3.35DE 2.99
  11. Ling 3.1 FlashinclusionAI · free tier3.30DE 2.91
  12. Step 5 PreviewStepFun3.26DE 2.96
  13. Claude Fable 5.1Anthropic3.22DE 3.26
  14. MiMo-V2.6-ProXiaomi3.22DE 2.97
  15. GPT-5.4 nanoOpenAI3.09DE 2.82
  16. DeepSeek V4.1 FlashDeepSeek2.96DE 3.05
  17. Qwen3.8-Flash-NextAlibaba · local2.33DE 2.14
  18. Gemma 4 26BGoogle · local2.30DE 2.25
  19. Nemotron 3 UltraNVIDIA · free tier2.24DE 2.00
  20. gpt-oss-120bOpenAI · local1.63DE 1.54

Two judges, two caveats: GPT-6.1 Sol and Claude Opus 5.5 grade every article and were also tested as writers, so they partly grade themselves. Claude models ran with effort medium unless noted.

Source: Thorsten Meyer AI model-scout benchmark, 8 Sep – 13 Oct 2026. Task: write a news article from a real source with the production prompt, English and German editions. Judges: GPT-6.1 Sol (reasoning high) and Claude Opus 5.5, both with the source in context; score 1–5, mean of both judges. Robustness run: 2 test articles plus 10 more real sources, mostly short like production; mean with 95% bootstrap interval.
Powered by Thorsten Meyer AI

Cost Put Luna Into Production

The reported results underline a practical distinction for publishers: the highest-scoring model may not be the one that makes sense for routine production. Meyer said he chose GPT-6 Luna because it delivered consistent, usable articles at a much lower reported price than GPT-6 Astra. For high-volume publishing, small per-article costs can accumulate, while a model’s reliability on short source material can matter as much as its best score.

The test also highlights a risk beyond the primary model. Meyer reported that two low-cost fallback models produced invented quotations or failed to produce valid articles. A failover system can preserve output when a preferred model is unavailable, but poor fallback performance may introduce errors precisely when the main route cannot be used. The report says those two models were removed from the news route on October 7.

These findings are not a general ranking of AI systems. They describe performance on one publisher’s workflow and evaluation criteria. Other outlets may use different prompts, source types, languages, editing processes and standards for acceptable copy, so the report supports task-specific testing rather than a universal claim about which model writes best.

Hisense 43-Inch E6SR Series Hi-QLED 4K UHD Smart Roku TV (2026 Model)

Hisense 43-Inch E6SR Series Hi-QLED 4K UHD Smart Roku TV (2026 Model)

Ninja NC501 CREAMi XL Deluxe 11-in-1 Ice Cream and Frozen Treat Maker

Ninja NC501 CREAMi XL Deluxe 11-in-1 Ice Cream and Frozen Treat Maker

Ninja Flip Toaster Oven and Air Fryer, 8-in-1, Stainless Steel, SP151

Ninja Flip Toaster Oven and Air Fryer, 8-in-1, Stainless Steel, SP151

As an affiliate, we earn on qualifying purchases.

A Pipeline-Specific Benchmark

Meyer framed the test as a response to the limits of general-purpose leaderboards. Such rankings may compare broad capabilities, but the report argues they do not directly show which model can produce publishable news from a particular source and prompt. The evaluation therefore used the same prompt as the production pipeline and included both a long source and a short one, with the shorter political item intended to reflect the publisher’s typical starting material.

The test also compared different model settings. GPT-6 Luna was run at six reasoning-effort levels. The report says the production setting, “none,” scored 3.67 in English and 3.68 in German at about 0.12 cents per article and 17 seconds. At “max,” it scored 3.58 and 3.57, while reported cost rose to 1.43 cents and generation time to 211 seconds. Meyer concluded that the score differences did not exceed the test’s stated noise threshold.

Other results in the report suggest that performance on one task does not necessarily carry over to another. Claude Opus 5.5, which Meyer described as the top coding model on a separate benchmark, scored 3.46 in English and 3.39 in German for news writing. Claude Sonnet 5.5 scored close to Luna in the initial test, but Meyer estimated its article cost at about $0.036, around 30 times Luna’s reported cost.

“The right model is the cheapest one that clears your quality bar reliably, not the one at the top of the chart.”

— Thorsten Meyer, ThorstenMeyerAI.com

Limits of the Reported Results

The findings come from a single publisher’s evaluation, not an independent or industry-wide benchmark. The supplied report does not provide the full set of article outputs, judge-level scores or detailed scoring records, so readers cannot independently assess how each judgment was reached. It also does not establish how the models would perform on other beats, longer reporting, breaking-news updates or additional languages.

The initial comparison used only two source texts, while the robustness run added 10 articles for the strongest configurations. Meyer says gaps under about 0.3 points in the two-article test should be treated as noise, making the larger run more informative. Even so, the article does not give the full robustness-run results for every one of the 71 configurations. The report’s cost figures are estimates tied to its use and setup; actual costs may vary with pricing, token use and pipeline design.

The source material says GPT-6 Luna now runs first after the two weak fallbacks were removed, but its final sentence is incomplete where it describes the backup arrangement. The report does not identify a replacement fallback, explain how the pipeline behaves if Luna is unavailable, or provide longer-term production results.

Production Use After the Test

Meyer’s report says GPT-6 Luna remains the production model after the October 7 removal of GPT-4.1 nano and gpt-oss-120b from the news route. The account gives no further date for a new benchmark, model change or published follow-up. The next useful evidence would be performance over a larger stream of real production articles, including whether source fidelity and valid-output rates hold across varied stories.

For readers considering similar systems, the report points toward testing the complete workflow rather than relying on a model leaderboard alone: compare outputs against their sources, check for fabricated quotations, measure editing and repair needs, and record cost and latency under real usage. Whether Luna’s reported balance of price and quality carries over to other publishers remains untested in this account.

Key Questions

Which model scored highest in the test?

GPT-6 Astra scored 4.00 out of 5 in both English and German in the standard test. It also scored 3.93 across 12 articles in the robustness run, according to the report.

Why did Meyer choose GPT-6 Luna for production?

Meyer said Luna performed consistently and cost about $0.0012 per article, compared with about $0.12 for Astra. He reported a Luna robustness score of 3.67 and valid output for every article in that run.

Did higher reasoning effort improve Luna’s articles?

The report says score differences across Luna’s six reasoning-effort settings were within the test’s noise range. The “max” setting took about 211 seconds and cost about 1.43 cents per article, compared with 17 seconds and 0.12 cents at “none.”

What happened with the backup models?

Meyer reported that GPT-4.1 nano generated invented quotations and scored 1.62 in English and 1.68 in German. The local gpt-oss-120b setup produced no valid articles in four attempts and also generated named quotes the source did not support. Both were removed from the news route on October 7.

Can these results identify the best model for every newsroom?

No. The report tests one publisher’s prompt, two source texts, two languages and a defined evaluation process. Other newsrooms would need to test models against their own sources, editorial standards, costs and production requirements.

Source: ThorstenMeyerAI.com

You May Also Like

When a Content Network Starts Publishing to Itself

Major content networks are now increasingly publishing to their own properties, creating interconnected ecosystems that boost engagement and control.

Zdf Livestream

ZDF has launched a dedicated livestream to broadcast the upcoming solar eclipse, providing real-time coverage for viewers across Germany and beyond.

Did Claude’s Coding Model Change For Game Reverse Engineering?

A Kotaku headline suggests Claude’s game reverse-engineering ability may have been reduced, but the available material provides no evidence or Anthropic response.

When a Content Network Starts Publishing to Itself

Discover how content networks self-publish, the risks involved, and practical fixes. Learn how to prevent your network from strangling itself silently.