alt Hacker NewsForgot where i saw it discussed; If you observe recent model benchmarks over the past year; the performance is slowly climbing, but if you divide by the token count; the score per token is dropping.
The current trend in state-of-art LLM coding agents is giving more output, thinking longer and checking the results more to catch mistakes. Be it an economics inventive to make users burn through their quota or show increase in usage for shareholders, or a market demand of users liking the ability of models to do independent work without intervention or oversight; the result is what the article seem to call the Vibe Tax.
I myself asked Claude code recently to review a somewhat large PR, to see what it would find. I didn't expect much, but also didn't quite realize how the model would interpret my request; I burned $20 in 3 minutes in API usage, as it ran 2 sub-agents which themselves spun up 5 more each. Most sub-agents were manually checking for things clang-tidy would catch without actually calling clang-tidy. This behavior rose as i changed from sonnet/opus 4.6 to 4.8 and now 5.0.
I don't want to run a agent independently in this way; i ask targeted questions about specific things and review the result. But model development is targeted towards a more hands-off "vibe" workflow, because that's where the money and hype is. As a result, i find the models more frustrating, less trustworthy and more costly to my work. (I've even started using haiku more, since it remains to-the-point without steering away from what i ask)
Replies
Happens with humans too! My senior colleagues check in with me significantly less often and cost significantly more in the meantime!
This is my experience too, but even worse. Opus 5 finished the task, I then asked it to code review it, 61 agents later it came back with a bunch of errors that needed fixing. The first pass had tests, they passed, they were just wrong. I wish more people started reviewing their AI output, because I see a worrying trend of ”we have all these tests the agent wrote so it has to be good”, which is not surprising because understanding tests is not a trivial skill.
It should be expected that more tokens give diminishing returns. Minimally, there's no limit on tokens but there is on quality of output (you can't reach negative bugs, or negative execution time). The graphs I've seen show a curved "frontier" of the tradeoff, and that line has improved over model generations.
That said, the companies are incentivized to sell you tokens, and therefore to have the models use as many tokens as they think you'll let them get away with for a given task / level of performance.