July 31, 2026
Route? Rude. Reboot the drama
Everyone is building LLM routers, we deprecated ours
AI shortcut craze hits a wall as commenters cheer, roast, and call the whole thing messy
TLDR: A company shut down its AI model-picker after saying the savings weren’t worth the confusion, inconsistency, and extra work. Commenters mostly yelled “called it,” while others roasted the writing, questioned the whole idea, and turned the thread into a mini tech soap opera.
In a deliciously spicy plot twist, a company that built its own AI traffic cop is now saying: actually, never mind. After testing its system across 7,000 users, it decided the idea of automatically picking a cheaper or smarter chatbot for each request just wasn’t worth the chaos. Their big complaint? A request can look simple at first and turn into a monster later, and the money saved can vanish once you factor in all the extra unpredictability, maintenance, and quality issues.
And oh, the comments came in swinging. One of the loudest reactions was basically, “Yep, routers suck.” Another crowd favorite argued that if this kind of choosing needs to happen at all, it should happen on the user’s device, not behind the scenes. Skeptics also piled on to say the whole idea may be built on a fuzzy premise: if you can’t clearly define what counts as “simple” or “complex,” what exactly are you predicting in the first place?
But the funniest energy came from the drive-by roast squad. One commenter said their faith that a real human edited the post actually went up because of one clunky sentence—a brutal little compliment disguised as an insult. Then, out of nowhere, the thread swerved into classic internet randomness with an “anti-American?” detour that had everyone blinking. So yes, the company killed its router—but in the court of public opinion, the real headline was the comment section: equal parts victory lap, nerd fight, and meme factory.
Key Points
- •Manifest launched its LLM router in March, deprecated it in June, and planned to shut it down on September 1 after four months of use.
- •The company says its router classified requests into four complexity tiers: simple, standard, complex, and reasoning.
- •Based on usage across 7,000 cloud users, Manifest reports mixed results and says task complexity cannot be reliably determined from the prompt alone.
- •The article argues that caching, especially prefix caching, often reduces costs more effectively than routing because cache reads are said to be 75% to 90% cheaper than uncached inputs.
- •Manifest says routing can reduce behavioral consistency and increase maintenance complexity in agentic workflows, including evals, prompts, and observability.