This article was created with the help of AI. The test, the posts and the numbers come from my own news pipeline. On October 8 I ran ten of my news posts through two Claude models. Opus 5.5 already wrote them for me every day. Fable 5.1 is the bigger model, and its price per token is 2.5 times highe
r. A blind reviewer preferred Opus in six of the ten posts, each time by a slight margin. Fable took twice as long and cost 3.2 times as much per post. I run two AI news feeds, Claude News and Promtime.
Claude takes an English source and rewrites it into a post by each channel's rules. Claude News is for developers: Claude releases, the API, pricing. Promtime covers AI in general, in plainer language. Every story goes out in two languages and in several formats, from a long Telegram post to a short one for social networks, so the rewrites add up to a few dozen calls a day.
Benchmarks mostly measure math and code. My task is narrower: retell a news story accurately, keep it lively and stay strictly inside the format. I wanted to see the difference on my own posts before paying for it. Claude Code built the test harness.
I set the task and checked what came out. I took the five latest published posts from each feed, only those with a single source between 600 and 20,000 characters long. Both models got exactly what the live pipeline sends: the channel rules, a description of the style, examples of good posts and a hint for the ending. The same code assembled the request.
Both were called directly through the API with the same settings, one run per post. I didn't judge them myself. A separate Claude session got the texts labelled A and B, shuffled for every post, so it never knew which model wrote what. It checked numbers and names against the source and scored each text from 1 to 5 on seven criteria: facts, headline, structure, language, ending, the short social post and the channel rules.
The reviewer didn't get the last word either. Five of the distortions it reported were checked against the sources separately, and all five held up. About twenty more numbers and names from both models were spot-checked, with no mismatches. Opus won six posts, Fable four.
The reviewer called all ten decisions a slight edge. Neither model invented a single fact. Opus made eight minor distortions, Fable seven. By criterion, Fable was a little more accurate on facts, 4.4 against 4.3.
Opus was better at endings, 4.5 against 4.2, and at the short social post, 4.2 against 3.8. The total hides a split. In the developer feed the reviewer picked Fable in four posts out of five. In the general feed it picked Opus in all five.
Five posts per feed is too few to call it, but it points one way: dense, precise text wins when the reader is technical. Ten posts with one run each is enough to see that there's no big gap. It isn't enough to rank two models this close. The judge was a Claude session too, so its taste isn't neutral ground.
And it's one task: retelling news from a given source under strict rules. Code, long reasoning and agent work are another matter. It digs deeper. One source was a page of measurements for the new Haiku 5.5.
Both models put the main number in the headline: the first token of an answer arrives after 415 seconds. Opus ended its post on the bright side: the wait at least comes cheap. Fable noticed something the price list doesn't show. The model is very verbose.
Its run through the benchmark index took 440 million output tokens against a median of 100 million, so by the cost of one task it ranks only 46th of 182. That was the real story. It also took Fable 130 seconds and nearly five times the cost of the Opus post, which took 34. It is looser with limits.
In a story about Meta and Microsoft, its short social post came out at 248 characters against a hard ceiling of 240. Opus wrote 164. That was the only length violation in the whole run. It was also the first thing the reviewer named when it gave the post to Opus, even though it found Fable's text slightly more precise.
It generalizes more boldly, and that's where it went wrong. In a story about GPT-6, Fable wrote that Sol goes to paid plans and Luna to free ones. In the source, Luna goes to the Free and Go tiers, and Go is paid. The reviewer flagged it in the headline, the ending and the social post, and still rated Fable's headline as the stronger one.
Opus took that post by a slight margin. By the price list, Fable costs 2.5 times as much as Opus for input and output tokens alike. Per post it came out 3.2 times as expensive by the median and 3.7 times on average. The difference is in the output.
Both models read the same prompt, but Fable spends twice as many tokens on reasoning and the answer: 6.2 thousand on average against 3 thousand. Output tokens are the expensive ones, and that's also where the time goes: 63 seconds against 29 by the median, 130 at worst. And the prompt is big. Every rewrite carries 16-24 thousand input tokens: the channel rules, the style description, the examples of good posts.
The source itself is the smaller part. On average, Fable's input alone cost more than a whole Opus call, answer included. Before writing a word, Fable had already cost more than a finished Opus post. Rewriting stayed on Opus 5.5, and it's still there a month later.
Neither model made anything up, and both slipped on small details about equally often. For this task the rules and examples in the prompt seem to matter more than the class of the model. I'll come back to Fable if I start a feed where getting to the bottom of a story matters more than fitting the format. The Haiku 5.5 post shows it earns its price there.
Both feeds from this test also publish in English, on Promtime.
Key Takeaways #
- •A blind reviewer preferred Opus in six of the ten posts, each time by a slight margin.
- •This story was reported by Dev.to , covering developments in thedev space.
- •AI advancements continue to reshape industries — read the full article on Dev.to for complete coverage.
📖 Continue reading the full article: