Anthropic, a leading AI company, is facing scrutiny after a premium Claude subscriber revealed that the company quietly downgraded the AI model's reasoning capabilities for its high-tier users without prior notification. The user, who pays $200/month for Anthropic's highest individual tier, Max 20x, detailed how this unannounced change negatively impacted their development workflow.
The user, who relies on Claude Code (Anthropic's CLI tool) and a team of AI agents daily to build high-performance .NET libraries—including GPU compute transpilers, WebRTC networking, and machine learning inference engines—explained that for months, 'High' was the maximum 'effort' setting available, signifying the model's full reasoning capability. However, sometime in late March 2026, Anthropic introduced a new tier called 'Max,' positioning it above 'High.' Crucially, this change was made without any email notification, CLI banner, or announcement to subscribers about the altered meaning of their existing 'High' setting; it was discovered by chance during a configuration check.
The introduction of a tier above 'High' effectively downgraded the 'High' setting itself. While 'High' previously meant utilizing the AI's full reasoning capability, with 'Max' now existing, 'High' has been redefined to mean employing a substantial but capped reasoning budget. This shift causes the model to favor statistically likely solutions over exhaustively tracing logic. While this distinction might not matter for simpler tasks like writing a blog post or a React component, for complex operations such as transpiling C# IL to WebGPU shader code or debugging WebRTC signaling race conditions, it represents the critical difference between functionally correct code and code that appears correct but fails at runtime.
This silent downgrade incurred significant real costs for the user's team. Prior to discovering the 'Max' setting, their team generated and shipped code under what they believed was the maximum effort ('High'). Upon switching to 'Max,' the audit results were alarming:
- A single audit of their ML inference library uncovered 24 production bugs, including buffer aliasing that crashed WebGPU, memory leaks from unreturned GPU buffer rentals, and a reflection-based dispose mechanism that was destroying shared hardware accelerators.
- A complete rewrite of their WebRTC signaling was necessary. The AI agent had ignored explicit reference implementations provided by the user (their own working code from other repositories) and instead pattern-matched a generic solution, failing to reason through the architectural complexities.
- A full week was spent debugging issues that the AI should have caught during initial implementation, such as race conditions, missing BEP 10 handshakes for incoming peer connections, and a MockLoopback test infrastructure bug that only manifested due to the agent's incomplete initial implementation.
The user confirmed that the 'Max' setting successfully identified and addressed all these critical issues, which the 'High' setting had missed by 'being efficient'—a euphemism for taking shortcuts.
Comparative metrics further illustrate the impact:
- The test suite showed an improvement from 471 passing tests to 498 passing after switching to 'Max.'
- WebRTC P2P functionality transitioned from 'Broken (offer collision)' to 'Working (offers-with-announce).'
- Wire extensions went from 'Untested' to '27 new tests, all passing.'
- BEP 46 ECDSA tests progressed from having 'Gaps' to '10 tests with real cryptographic verification.'
- Crucially, while 0 production bugs were identified (and being shipped) before the 'Max' discovery, 24 were found in a single audit after the switch.
For paying premium subscribers, the expectation is to receive intelligence and high-quality reasoning, not merely an increased message count. This incident raises serious questions about Anthropic's transparency and the true value proposition of its premium AI services.