Anthropic appears to be A/B testing reduced effort levels in Claude Code

Based on 15 sampled comments from this thread. Analyzed 2026-08-23.

The thread is less interested in the displayed effort number than in whether Claude Code's higher reasoning modes have become visibly worse in real work. Several developers describe simple edits expanding into long, costly agent detours: container setup, broad repository inspection, and test construction that add nothing when the requested outcome is a single configuration-file change. Others say medium effort now beats high or extra-high effort because the latter is more likely to wander, overthink, or produce confident but wrong work. Some have responded by downgrading expensive subscriptions and using competing models to repair Claude's mistakes. An Anthropic team member says the current experiment only changes how an internal numeric effort value is mapped, not the selected effort or model quality, and points users toward feedback reports for reproducible regressions. That clarification won some appreciation, but it did not settle distrust. Commenters connect perceived quality swings to opaque routing, load management, usage limits, and token-based billing, arguing that providers can quietly alter the product while customers cannot independently audit what they received. The recurring practical conclusion is stark: AI coding agents are useful only when they stay scoped, predictable, and cheaper than supervising their unnecessary reasoning.

The sharp dispute is whether the apparent degradation is merely a harmless effort-scale remapping or evidence that Claude is being quietly optimized away from reliable coding performance. Anthropic's case is that evaluations show no model-quality change and reports of clear regressions should be investigated. Skeptics counter that repeated real-world failures, time-dependent quality swings, and excessive token burn are exactly what a clean internal evaluation can miss.

• Praise: Anthropic's direct explanation of the effort-number experiment and request for concrete feedback is a welcome path to separate interface confusion from genuine model regressions.
• Praise: Medium reasoning effort reportedly gives some users cleaner, more dependable coding results, suggesting that less deliberation can prevent agentic scope creep.
• Praise: The growing availability of alternative proprietary and open-source coding models gives developers an escape hatch when one provider's agent becomes unreliable or expensive.
• Concern: High-effort Claude Code reportedly turns trivial one-file edits into sprawling sandbox, container, and whole-repository exercises, making the agent slower to supervise than the task itself.
• Concern: Several users say top reasoning modes now miss elementary logic and backend changes while producing elaborate prose, eroding trust in a premium coding assistant.
• Concern: Unannounced A/B tests on paying users fuel suspicion that model behavior can change without consent, visibility, or a meaningful opt-out.
• Concern: Token billing feels dangerously opaque when providers control both the hidden reasoning budget and the model routing, leaving developers unable to forecast workflow costs.
• Concern: Perceived quality fluctuations by time and load lead commenters to suspect silent capacity management, even though the provider says the tested mapping does not affect performance.