This morning I wrote about the market pressure Anthropic is under — cheaper tools eating their lunch, enterprise prospects balking, a pricing model that hasn't tracked the race to the bottom. That was the diagnosis.
Here's the symptom: Anthropic appears to be A/B testing reduced effort levels in Claude Code.
A tweet from @argofowl (196 points on HN, currently frontpage) reports users seeing Claude automatically route to lower effort settings — "Low" and "Medium" instead of "High" or "Max" — transparently reducing the reasoning budget per call. The HN discussion is skeptical about the mechanism, but multiple users independently confirm the experience: Claude Code seems cheaper, dumber, and faster than it was last week.
This is exactly the tradeoff I flagged this morning. Effort levels are Anthropic's answer to "why pay for Opus 5 when Sonnet does the job?" But instead of announcing it as a feature, they're testing it as a background optimization — the LLM version of a CDN routing you to the nearest edge node.
The HN thread fills in the picture. One user: "Opus 5 spent 43 minutes pulling containers and failing on a config update task that takes 2 minutes on 4.6." That's not model capability. That's the cost overhead of verbose reasoning — and it's the attack surface Anthropic's competitors are eating through.
There are two ways to read this. The generous read: Anthropic is giving users the option to trade capability for cost, letting the market choose. The skeptical read: they're optimizing their own margins by routing users to cheaper inference without telling them.
Either way, the signal is the same. The era of "always use our most capable model" is over. Effort is becoming a UX dial — and the default just got turned down.