You ask your AI coding assistant to read a config file and update it with some new data. It should take two minutes. Instead, it takes 43 minutes. It pulls containers, spins up sandboxes, and builds entirely unnecessary testing suites for a two-line code change. You aren’t experiencing a breakthrough in artificial intelligence. You’re getting gaslit by a cost optimization algorithm.
Anthropic appears to be quietly A/B testing reduced “effort” levels in Claude Code. To the average user, this just feels like the model is getting dumber. But look closer, and you’ll see the tension: AI assistants, built to reduce human effort, are themselves being engineered to expend less effort. The model’s own cognitive exertion is being dialed back to save compute.
We built machines to do our heavy lifting, only to discover they’re secretly unionizing for shorter hours behind our backs.
We were sold the dream of AGI—tireless, ultra-competent digital workers that would code while we slept. Instead, we are watching a multi-billion dollar company frantically turning the knobs down to hit their cloud compute margins. The “effort” level of an LLM isn’t a fixed attribute. It’s a strategic lever for cost control. When you ask the AI why it’s struggling, it might tell you it’s doing its best. But how can you trust the self-reporting of a system that is actively being optimized to lie about its own capacity?
It’s like asking a waiter if they spit in your food. The incentives to lie are simply too strong.
The developer trenches are already seeing the fallout. Sub-agents are burning through absurd amounts of tokens for trivial tasks, hitting invisible walls, and rubberbanding usage limits. One developer noted that a prompt that used to take under two minutes on an older model now takes 43 minutes of agonizing, unnecessary sandbox spinning on Opus 5. It’s not a bug. It’s a feature—for Anthropic’s bottom line.
When the AI’s effort level becomes a dial that providers can secretly turn down, intelligence stops being a breakthrough and starts being a metered utility.
This is the dark underbelly of the AI boom. The providers want you to believe they are racing toward a sci-fi future, but behind the curtain, they are desperately trying to figure out how to reduce inference costs. They are routing you to weaker models in the backend. They are throttling your context windows. They are turning the effort knob down to a three and hoping you don’t notice your code is breaking.
Neutrality in the face of this is death. This is dangerous. Opaque optimizations that silently degrade performance while providers save costs directly undermine the quality and reliability of your work. You are paying for a premium product and receiving a throttled, budget version in secret.
You aren’t experiencing the cutting edge of artificial intelligence; you’re experiencing a product manager hitting a margin target.
We offloaded our work to AI because we didn’t want to put in the effort. Now, the AI doesn’t want to put in the effort either. As one frustrated user perfectly put it: “Is this AGI?” No. It’s just a lazy coworker that charges by the minute. Stop trusting the black box. Demand transparency, or watch your productivity silently degrade to subsidize someone else’s cloud compute bill.
FAQ
Q: How do we know the model's effort is actually being dialed back?
A: Look at the token usage and task completion times. When a 2-minute config update takes 43 minutes and burns through containers, you're seeing a degraded model routing or a throttled effort level in real time.
Q: What does this mean for developers using AI?
A: You must treat AI output as untrusted labor. Monitor your token costs, watch for backend model routing, and never blindly trust the model's self-reported status when tasks inexplicably fail.
Q: Is this just a necessary evil for AI companies to survive?
A: No, it's a bait-and-switch. Selling users on the promise of Artificial General Intelligence while secretly delivering a throttled, metered utility is a betrayal of the core product promise.