Anthropic is facing a significant backlash from enterprise users over claims its flagship Claude Opus 4.6 model has been quietly “nerfed,” or made less capable, with one analysis showing a 67% reduction in the model's reasoning depth for complex coding tasks. The controversy threatens to erode trust in the $380 billion startup, particularly as it competes with OpenAI’s enterprise offerings and reportedly heads toward an IPO.
"When thinking becomes shallow, the model tends to take the lowest-cost action," Stella Laurenzo, a senior director of AI at AMD, wrote in a widely circulated GitHub analysis. "Modifying without reading, stopping before it's done, shirking responsibility for its mistakes, and choosing the simplest, not the most correct, solution."
The core of the user complaints, which have spread across GitHub, Reddit, and X, is that Claude has become less reliable for the complex, multi-step workflows it was initially praised for. An analysis by Laurenzo of over 6,800 Claude Code sessions found that from late February to early March, the model’s "reads-per-edit"—a proxy for how much context it considers before writing code—plummeted from 6.6 to 2.0. In response, Anthropic’s head of Claude Code, Boris Cherny, stated the company had not secretly degraded the model but had changed the default "effort" level to "medium" to balance intelligence, latency, and cost for most users.
This controversy highlights the opaque nature of the "token economy," where customers pay for a seemingly standard unit of AI processing without guarantees on the quality of intelligence delivered. While token prices have fallen roughly 300-fold in three years, enterprise AI budgets are becoming harder to control. A survey from Mavvrik and Benchmarkit found 84% of enterprises report AI costs have eroded gross margins more than expected, with only 15% able to control budget variance to within 10%. The issue is compounded by technical factors like caching; one analysis showed that a change in Claude Code's caching behavior could increase input costs by 5.7 times.
The "Shrinkflation" Problem
The heart of the user revolt is a sentiment that they are paying the same price for a less capable product, a phenomenon some have dubbed "AI shrinkflation." The issue gained traction after a viral post on X from developer Om Patel summarized the perceived decline as a 67% drop in capability, echoing the findings from Laurenzo's GitHub analysis.
Anthropic has pushed back, attributing the perceived changes to product and interface choices, not a secret downgrade. Cherny noted that on February 9, Opus 4.6 enabled "adaptive thinking" by default, and on March 3, the default effort level was set to "medium." While users of the Claude Code terminal can manually set the effort to "high," Pro and Enterprise users on other platforms cannot. In response to the backlash, Cherny said the company will test defaulting Teams and Enterprise users to "high effort."
A Question of Trust and Compute
The debate occurs as Anthropic experiences surging demand, which has led to stricter usage limits during peak hours and fueled speculation that the company may be facing a compute shortage. OpenAI's revenue chief, in a reported internal memo, claimed Anthropic made a "strategic misstep" by not securing enough compute capacity. Anthropic has denied it degrades models to manage demand.
The situation creates a critical trust gap for a company that has branded itself as more transparent and aligned with user interests than its rivals. As Anthropic competes with offerings like OpenAI's Codex and eyes a potential IPO, the perception that it would silently reduce model quality—even if for cost-balancing reasons—could damage its standing with the enterprise developers who have been key to its growth. The company's challenge now is to reconcile the fixed price of a token with the variable value of the "intelligence" it contains.
This article is for informational purposes only and does not constitute investment advice.
