Detailed Analysis
Claude Opus 4.6, Anthropic's flagship model released on February 5, 2026, has become the subject of mounting user complaints regarding sharp performance degradation, particularly in response speed. Reports concentrated on platforms such as Cursor's developer forum describe token generation rates as low as 3–4 tokens per second and wait times exceeding 20 minutes for prompts that would ordinarily resolve in seconds. The complaints are notably model-specific — users indicate that other Claude models available through the same interfaces remain unaffected — which has amplified speculation about the root cause. Among the theories circulating in developer communities is the allegation that Anthropic may be deliberately redirecting compute resources toward a successor model, though no evidence has emerged to substantiate that claim, and Anthropic has not confirmed any such policy.
A significant portion of the performance issues may instead be attributable to deliberate architectural decisions Anthropic made in designing Opus 4.6. The model was engineered to "think more deeply," revisiting its reasoning chains before delivering responses, and to handle an extended context window of up to one million tokens — currently in beta. These capabilities carry an inherent latency cost, particularly on simpler prompts that do not warrant the model's full deliberative capacity. Anthropic has acknowledged this tradeoff and recommends that developers use the `/effort` parameter — reducing the setting from high to medium — as a practical mitigation for tasks where deep reasoning is unnecessary. This design philosophy reflects a broader industry shift toward test-time compute scaling, in which models are given more processing "budget" at inference time to improve output quality, at the expense of raw speed.
Beyond the architectural overhead, user reports and independent monitoring point to several additional degradation vectors. High API utilization within tools like Cursor appears to trigger throttling, slowing generation speeds as usage percentages climb. A separate and technically distinct issue — sometimes called "context rot" — sees performance drop substantially before even 50% of the one-million-token context window is consumed, a problem documented in Anthropic's own GitHub issue tracker. Independent model quality trackers such as MarginLab's Claude Code monitor have been watching Opus 4.6 alongside its predecessor Opus 4.5 for statistically significant benchmark regressions on metrics like SWE-Bench-Pro, but have not reported confirmed degradation tied to compute reallocation as of mid-April 2026. Objectively, Opus 4.6 continues to post strong benchmark scores — 90.2% on BigLaw Bench and 76% on MRCR v2 for long-context retrieval — suggesting that quality has not materially declined on standardized evaluations.
The episode highlights a persistent and widening tension in frontier AI deployment: the gap between benchmark performance and real-world user experience. Anthropic's decision to push the boundaries of context length and reasoning depth in Opus 4.6 represents a genuine technical achievement, but those same capabilities introduce latency profiles that feel punishing to developers relying on the model for rapid iterative workflows. The compute-redirection allegation, while unsubstantiated, speaks to a broader anxiety in the developer community about the opacity of inference infrastructure decisions at major AI labs — users have limited visibility into whether slowdowns stem from model design, capacity constraints, or deliberate prioritization. As competition among frontier model providers intensifies and infrastructure costs remain a central variable, Anthropic and its peers will face growing pressure to communicate more transparently about the operational factors that shape day-to-day model performance, particularly for enterprise and developer users whose workflows depend on consistent throughput.
Read original article →