Is Claude's AI Performance Declining? Users Report Frustration, Seek Alternatives

The performance of AI models, specifically Anthropic's Claude, has become a hot topic among users, with many reporting a noticeable degradation in quality and reliability. This perceived decline affects various use cases, from coding and code review to creative writing and complex problem-solving.

Perceived Quality Degradation

Many users, including long-term subscribers, have observed a clear decline in Claude's capabilities over the last few weeks. Specific issues frequently cited include:

Incomplete Responses and Streaming Stops: The model often stops streaming mid-response, hitting internal tool-call or turn limits. This forces users to repeatedly press "Continue," which re-feeds context and rapidly consumes tokens and quota.
Increased Errors and Reduced Accuracy: Users report more mistakes, a failure to fully complete tasks that were previously handled reliably, and a loss of creative thinking. For code, this manifests as nonsensical architectural suggestions, ignoring user-defined stop hooks, and even destructive actions on codebases.
Token Inefficiency: The need for frequent continuations makes the model, especially Opus, token-hungry and quickly depletes weekly quotas, becoming impractical for regular use.

One frustrated user, a professional engineer, described Claude as going from "impressive and helpful" to "unusable" in a matter of weeks, highlighting its new tendency to get confused and argue against necessary instructions.

The Business Imperative: Profitability vs. Performance

A significant portion of the discussion revolves around the potential reasons for this perceived downgrade. The prevailing theory suggests that AI companies, including Anthropic, are under immense pressure to justify eye-watering venture capital investments and become profitable. This financial imperative could lead to:

Cost-Cutting Measures: Degraded performance might be a result of technical adjustments like lower quantization or reduced computational resources per query, aimed at stemming operational costs.
"Bait and Switch" or "Enshittification": Some users explicitly accuse the company of attracting subscribers with a high-performing product, securing yearly subscriptions, and then intentionally degrading the service. This practice, often termed "enshittification," is seen as a way to potentially drive enterprise clients towards more expensive, higher-performing models.
Limiting Usage: The increased token consumption due to streaming issues might be an indirect way to limit usage and manage computational resources under high demand.

A Counter-Narrative: Speed Still Wins for Complex Tasks

Despite the widespread complaints, one user offered a nuanced perspective. While acknowledging subtle errors and mid-stream self-corrections, they found Claude 4.6 (Extended thinking) via the web interface still highly effective for tearing through extremely complex Kubernetes setups, reviewing code, and drafting plans. The near-instant answers and speed at which it reconciles concerns across dozens of dimensions remain a significant advantage, often outweighing the occasional need to correct minor errors. For high-velocity troubleshooting, this speed is considered a critical win, even if it means sacrificing absolute perfection.

Navigating AI Volatility: User Strategies and Alternatives

Facing inconsistency, users are adopting several strategies:

Voting with Their Feet: Many are canceling subscriptions or downgrading plans, stating that consistent negative experiences make the service unreliable and unacceptable.
Subscription Flexibility: A common recommendation is to opt for monthly subscriptions over yearly ones. This allows users to easily discontinue payments if the service quality deteriorates, avoiding the feeling of being locked into a sub-par product.
Exploring Alternatives: Users are actively seeking other AI tools. Mentioned alternatives include DeepSeek, Z.ai (GLM), GPT-5.4, Qwen 3.5 (often run locally for more control), and Codex. Some are even considering investing in powerful local hardware (like AMD 395 or Mac Studio) to run models and gain more reliable, self-controlled workflows.

The Broader Challenge of AI Reliability

The discussion highlights a fundamental challenge with current commercial AI services: their non-deterministic nature and the fact that performance can change at a vendor's whim. With models constantly being tweaked, updated, or resource-allocated differently on the backend, achieving consistent results becomes nearly impossible. This lack of transparency and reliability makes it difficult for professionals to integrate these tools into critical workflows, underscoring a call for greater stability and trustworthiness in AI product development.