By Thorsten Meyer | October 13, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Public leaderboards tell you which model is smartest. They don’t tell you which model should write your articles. So I measured it on my own job, and the answer was cheaper and more surprising than I expected.
Every week brings a new model and a new leaderboard. I write about those leaderboards, and I find them useful. But a general intelligence score cannot answer the question I actually have to answer: which model should write the news articles my sites publish every day?
So between September 8 and October 13, 2026, I tested 71 model configurations on exactly that task, with my real production prompt and real news sources. They included API models, subscription models, free-tier models on OpenRouter and open-weight models running on my own hardware. Every article was written twice, once for the English edition and once for the German edition, and graded by two AI judges.
The short version:
- GPT-6 Astra writes the best articles: 4.00 out of 5 in both languages, with perfect source fidelity.
- GPT-6 Luna writes them in production: 3.67 in English, 3.68 in German, for about a hundredth of Astra’s price.
- Turning up reasoning effort made nothing better. It only made articles slower and more expensive.
- The real danger was not the main model but the fallbacks. Two cheap backup models in my failover chain invented quotes. I removed both.
How I tested
I kept the test as close to production as I could:
- One task: turn a real news source into a finished article, using the exact prompt my pipeline uses.
- Two sources: a rich one (6,340 characters, a product filing) and a thin one (619 characters, a short political news item). Most of my news starts from short sources, so the thin one is the real test.
- Two editions: English (global) and German (for Germany, Austria and Switzerland).
- Two judges: GPT-6.1 Sol (reasoning high) and Claude Opus 5.5. Both see the original source. They score from 1 to 5 on source fidelity, fabrication, epistemic clarity, news value, structure, polish and, for German, language quality.
- A robustness run: the strongest configurations then wrote 10 more articles from real sources, mostly thin ones as in production. The results come with 95% bootstrap intervals.
- Validity: an article only counts if my pipeline can actually use it, meaning valid output after the same repair step production uses.
One rule of thumb I set for myself: with two test articles, gaps under about 0.3 points are noise. I treat the robustness run as the more reliable ranking.
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.
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.
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
As an affiliate, we earn on qualifying purchases.
Finding 1: The best model is not the one I run
GPT-6 Astra is the clear winner. It scored 4.00 in English and 4.00 in German with source fidelity 5.00 in the standard test, and 3.93 across 12 articles in the robustness run. Its interval (3.83–4.02) does not even overlap with GPT-6 Luna’s (3.59–3.76). This is a real difference, not noise.
It also costs about $0.12 per article, roughly 100× as much as GPT-6 Luna at about $0.0012. Astra takes around 50 seconds per article; Luna takes 17.
I decided on October 7 that Astra is too expensive for news. Luna is not the best writer in my test, but it is consistently good. It scores 3.67 in the robustness run, with source fidelity of 4.71 out of 5, and every article it wrote was valid. In production over the past week it cost about $0.0013 per article at a median of 17 seconds.
For a high-volume task, the right model is the cheapest one that clears your quality bar reliably, not the one at the top of the chart.
AI content generation tools for journalism
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Finding 2: More reasoning bought nothing
I ran GPT-6 Luna at all six reasoning-effort levels:
| Effort | Score (EN / DE) | Cost per article | Time |
|---|---|---|---|
| none (production) | 3.67 / 3.68 | 0.12¢ | 17 s |
| low | 3.67 / 3.64 | 0.13¢ | 18 s |
| medium | 3.58 / 3.68 | 0.16¢ | 26 s |
| high | 3.75 / 3.79 | 0.20¢ | 30 s |
| xhigh | 3.71 / 3.46 | 0.26¢ | 50 s |
| max | 3.58 / 3.57 | 1.43¢ | 211 s |
Every difference is within the noise. Max effort costs about 12× as much and takes 12× as long as none, for no gain. High is the only level with a slight edge, and in the robustness run that edge (3.73 vs. 3.67) sits inside the intervals. Production stays on “none”.
Writing a news article from a source is not a reasoning problem. It is a discipline problem: stick to the source, structure it well, don’t embellish. Extra thinking tokens don’t buy discipline.
AI language model for news articles
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Finding 3: The best coder is not the best writer
This one surprised me. On the new FrontierCode benchmark, Claude Opus 5.5 is the best coding model available. As a news writer in my test, it scored 3.46 in English and 3.39 in German, with source fidelity of only 3.75. In the robustness run it scored 3.52, below GPT-6 Luna, even though Opus was one of the two judges.
The rest of the Claude family looks like this (all at effort medium unless noted):
- Claude Sonnet 5.5: 3.67 / 3.61, effectively tied with Luna, but at an API-equivalent cost of about $0.036 per article, roughly 30× Luna
- Claude Haiku 5.5: 3.46 / 3.43 (high: 3.54 / 3.61), a big jump from Haiku 4.5 at 2.33 / 2.07
- Claude Fable 5.1: 3.21 / 3.29, at about $0.24 per article
Opus is how I build software, as I described in my October 2026 AI stack. It is not how I write news. Rankings don’t transfer between tasks. That is the main reason to run your own benchmark.
Infographic 02 · What quality costs
More money and more reasoning barely move the score
The spread in price per article is about 300×; the spread in quality among the top models is a few tenths of a point. Turning reasoning up did not help. Thin sources and cheap fallbacks are where articles go wrong.
Article score vs. cost per article
Standard test (2 sources), English edition. Up and to the left is better. Hover a dot for its values.
Claude and Codex runs are shown at their API-equivalent cost. Local and free-tier models (≈ $0) are not plotted. With two test articles, score gaps under ~0.3 are noise.
GPT-6 Luna: reasoning effort ladder
Same model, six effort levels. Score English / German, cost and time per English article.
| Effort | Score | Cost | Time |
|---|---|---|---|
| none ★ | 3.67 / 3.68 | 0.12¢ | 17 s |
| low | 3.67 / 3.64 | 0.13¢ | 18 s |
| medium | 3.58 / 3.68 | 0.16¢ | 26 s |
| high | 3.75 / 3.79 | 0.20¢ | 30 s |
| xhigh | 3.71 / 3.46 | 0.26¢ | 50 s |
| max | 3.58 / 3.57 | 1.43¢ | 211 s |
“max” costs about 12× as much and takes 12× as long as “none”, for no gain. Production stays on “none”.
When the source is thin
English score on a rich source (6,340 characters) vs. a thin one (619 characters). Most news starts from a short source.
| Model | Rich | Thin | Change |
|---|---|---|---|
| DeepSeek V4.1 Flash | 3.33 | 2.0 | -1.33 |
| Gemma 4 26B (local) | 2.92 | 1.92 | -1.00 |
| GPT-5.4 mini | 3.5 | 2.75 | -0.75 |
| Claude Opus 5.5 | 3.67 | 3.25 | -0.42 |
| GPT-6 Luna ★ | 3.75 | 3.58 | -0.17 |
| GPT-6 Astra | 4.08 | 3.92 | -0.16 |
A model that drops on the thin source fills gaps with invented detail.
The fallback that invented quotes
Source fidelity of GPT-4.1 nano, the second choice in the failover chain until 7 Oct. Overall 1.62 / 1.68. The judges flagged invented quotes from “an industry analyst” and an anonymous company spokesperson.
The fallback that broke the format
Valid articles from gpt-oss-120b, run locally as third choice. It wrote quotes outside the JSON string and invented named quotes. Removed from news on 7 Oct.
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. Cost: OpenAI list price or the amount OpenRouter charged; subscription runs at API-equivalent.
Powered by Thorsten Meyer AI
AI writing assistant for news publishers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Finding 4: Your fallbacks are where quality dies
Until October 7, my news route had two backup models behind the main one. The test showed they were the weakest in the field:
- GPT-4.1 nano (second choice) scored 1.62 / 1.68 with source fidelity 1.00, the lowest possible. The judges flagged invented quotes from “an industry analyst” and an anonymous company spokesperson.
- gpt-oss-120b (third choice, running locally) produced 0 valid articles out of 4. It wrote quotation marks that broke the output format and invented named quotes, including one attributed to a “Deputy Defense Minister”.
Both were removed from the news route on October 7. GPT-6 Luna now runs first, with GPT-6 Luna as its own fallback through a second access path.
This is the lesson I would most like other operators to take away. Everyone benchmarks their primary model. Almost nobody benchmarks the model that takes over when the primary fails. A failover chain is only as good as its weakest link, and that link writes real articles under your name.
The generational progress is also striking. OpenAI’s small models went from GPT-4.1 nano at 1.62 to GPT-5.4 nano at 3.25 to GPT-6 Luna at 3.67. A fallback that was acceptable a year ago can be a liability today.
Finding 5: Thin sources expose fabrication
When a source is short, a weak model fills the gaps with invented detail. The English scores on the rich source versus the thin one make this visible:
| Model | Rich source | Thin source | Change |
|---|---|---|---|
| DeepSeek V4.1 Flash | 3.33 | 2.00 | −1.33 |
| Gemma 4 26B (local) | 2.92 | 1.92 | −1.00 |
| GPT-5.4 mini | 3.50 | 2.75 | −0.75 |
| Claude Opus 5.5 | 3.67 | 3.25 | −0.42 |
| GPT-6 Luna | 3.75 | 3.58 | −0.17 |
| GPT-6 Astra | 4.08 | 3.92 | −0.16 |
The best models barely notice a thin source. The weaker ones lose a point or more. If your test only uses rich, well-written sources, you will overrate exactly the models that fail in production.
Finding 6: Local, open and free are not there yet for this job
I wanted a local model to win. None did.
- My latest batch of four local candidates was rejected. Aleph Alpha’s Kolibri-1 wrote the German edition in English, with language quality 1.0, invented a quote, and with reasoning took up to half an hour per article. Xing4.0 produced invalid output in 5 of 8 attempts at about 3 tokens per second. Kimi-K3 REAP and Bonsai-2 were rejected without a run: one model card says outright that the model has no factual knowledge, and the other is a compression of a model family that had already topped out at 3.17.
- The best local model, Qwen3.8-Flash-Next, scored 3.46 with thinking enabled but needed about 312 seconds per article. Without thinking it scored 2.88. In the 12-article run it fell to 2.33 with source fidelity 2.0, so I took it out of news on October 13.
- The best open-weight model in the robustness run, Nex-N2 mini, scored 3.35, but only half of its extra articles were valid.
- Free tiers: of 14 free models on OpenRouter, the best was Ling 3.1 Flash at 3.25 / 3.36, at about two minutes per article and with only 3 of 4 articles valid. Almost all of them fabricated.
Open-weight models have closed much of the gap in coding. For careful, source-faithful news writing in two languages, they have not yet.
The newcomers
- Muse Wonder (tested October 13): 3.62 / 3.36, 3.58 in the robustness run, at about $0.07 per article, roughly 60× Luna. A good writer at a premium price, which matches what Artificial Analysis found about its cost.
- Mistral Large 4 (tested a day after release): 2.75 / 2.50 with reasoning, 2.21 / 2.75 without, below GPT-5.4 nano. It invented a quote, attributed to Russian officials, saying Putin had backed the proposal. That fits what I wrote about its gap to the frontier and why I would not run agents on it.
- Grok 4.7: 3.38 / 3.43, at about $0.08 per article and two minutes per article.
A caveat about the judges
Using AI judges is practical, but it has a bias problem I can measure. Claude Opus 5.5 is consistently more generous than GPT-6.1 Sol. The median gap is about +0.5 points for non-Claude writers and about +0.9 for Claude writers. That pattern fits a self-preference effect. With two test articles per configuration it is not proven, but it is why I use two judges from different labs, report their mean, and lean on the 12-article run for decisions.
Both judges were also tested as writers. GPT-6.1 Sol did well (3.83 / 3.79). Opus did not (3.46 / 3.39). So the bias, if it exists, didn’t simply crown the judges.
What I would tell anyone running AI content at scale
1. Benchmark your own task. General leaderboards ranked Opus first. My task ranked it below a model that costs a hundredth as much. 2. Benchmark your fallbacks, not just your primary model. Mine were inventing quotes. 3. Test on thin sources. That is where fabrication shows. 4. Start at the lowest reasoning effort and prove that more helps before paying for it. 5. Compare cost per article, not price per token. The real spread in my test was about 300×. 6. Use two judges from different labs, and treat small gaps as noise.
Verdict
The best writer in my test is GPT-6 Astra, and I don’t use it. GPT-6 Luna at no reasoning effort is good enough, consistent and about a hundred times cheaper, and that is what production needs. The bigger win of these weeks was not choosing a better main model. It was finding and removing the two fallbacks that were quietly making things up.
Sources
- Thorsten Meyer AI model-scout benchmark, 8 September – 13 October 2026: 71 configurations; standard test with 2 real sources in English and German; robustness run with 12 sources; judges GPT-6.1 Sol and Claude Opus 5.5.
- Pricing: OpenAI list prices (as of October 7, 2026) and amounts actually charged by OpenRouter. Subscription models are shown at their API-equivalent cost.
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
