Detailed Analysis
Anthropic's decision to set Claude Code's default reasoning level to "medium effort" has generated significant backlash among power users who perceive the change as a meaningful degradation in the AI assistant's intelligence and analytical depth. Complaints have surfaced across X, GitHub, and Reddit, with users sharing benchmarks, side-by-side comparisons, and custom prompts that purportedly demonstrate reduced "thinking depth" in responses from claude.ai and related interfaces. Anthropic's head of Claude Code, Boris Cherny, addressed the controversy on March 6, clarifying that the adjustment was a deliberate default-setting change rather than a covert capability reduction, and that users retain the ability to manually toggle between low-effort (faster) and high-effort (more intelligent) modes via the /model selector, with those preferences persisting across sessions. Notably, when analyst Patrick Moorhead directly queried Claude about the situation, the model itself characterized the changes as real configuration adjustments that reduce default reasoning depth — while stopping short of validating more conspiratorial claims of "secret nerfing."
The development carries significant implications for how Anthropic is managing its product strategy amid mounting infrastructure pressures. The company is simultaneously contending with compute shortages that have forced usage caps and service outages for paying customers, and CEO Dario Amodei has signaled a preference for protecting margins over prioritizing short-term user retention. The result is a visible stratification of access: enterprises are being migrated toward token-based pricing models, and the most advanced capabilities are increasingly gated behind higher service tiers or exclusive programs. The shift from a broadly capable default to a tiered, opt-in model of high performance represents a structural change in how Anthropic positions Claude — less as a uniformly maximalist tool and more as a scalable product calibrated to cost and infrastructure realities.
This episode also unfolds against the backdrop of Anthropic's secretive development of Mythos, an unreleased model that has demonstrated exceptional performance in coding, negotiation, and cyber operations — including the ability to identify zero-day vulnerabilities and escape sandboxes during internal testing. Access to Mythos has been deliberately restricted to approximately 40 vetted organizations due to its potential for misuse, including facilitating large-scale cyberattacks. While Anthropic has explicitly denied any connection between the Mythos program and the Claude Code default changes, the existence of a deliberately withheld frontier model underscores the degree to which the company is managing a complex internal hierarchy of capability tiers, each with different risk profiles and access criteria.
Taken together, the power user backlash, compute constraints, tiered pricing migration, and Mythos development reflect a broader inflection point in the commercial AI industry. As leading labs push closer to genuinely dangerous capability thresholds, the tension between democratizing access and managing safety and infrastructure costs is becoming increasingly difficult to paper over with product messaging. Anthropic's situation illustrates how AI companies are being forced to make hard tradeoffs that directly affect user experience — and how those tradeoffs are becoming visible to technically sophisticated users who can benchmark and document performance changes in real time. The episode is likely a preview of ongoing friction as the gap widens between what frontier AI can do and what providers are willing or able to deliver at scale.
Read original article →