When developers collectively complained that Claude Code had suddenly become “dumber,” most people’s first instinct was to switch to a stronger, more expensive model. But a recent official technical blog post from Anthropic has completely upended that assumption: the culprit may not be the model itself, but a hidden toggle called “Effort” — and the trap was dug by Anthropic’s own hand.
The origins of this uproar trace back to March 2026. At the time, a large number of Claude Code users discovered that the AI assistant, which had previously been capable of independently completing complex engineering tasks, had begun frequently “slacking off”: skipping files it should have read, ignoring tests it should have run, and stopping halfway through tasks to demand more information. The sharpest criticism came from AMD’s head of AI, Stella Laurenzo. After analyzing 6,852 session logs on GitHub, she publicly pointed out that Claude’s reasoning output had plummeted 67% compared to pre-February levels, bluntly stating that “Claude can no longer be trusted to do complex engineering work.”
Initially, the community attributed the problem to poorly designed prompts or individual configuration errors. It wasn’t until the truth surfaced that people realized the problem lay with Anthropic itself. On March 4, in order to reduce response latency, Anthropic quietly lowered the default Effort setting in Claude Code from “High” to “Medium.” Although this change was documented in the update log, the vast majority of users never noticed — they simply felt, viscerally, that the model in their hands had inexplicably “gotten stupid.” After enduring nearly a month of public pressure, Anthropic finally restored the default setting to High on April 7 and reset usage quotas for all paying users.
In its latest official explanation, Anthropic distilled the core concept into a single sentence: “The Model changes the brain; Effort changes the attitude.” Behind every model lies a set of “frozen weights” — its knowledge and capabilities are locked in the moment training concludes and cannot be modified through subsequent prompts or context. Switching models essentially means swapping out an entire set of weights to address the question of “whether it can.” Effort, on the other hand, determines how much actual work Claude is willing to invest in a given task — whether it proactively reads through multiple files, executes test validations, and pushes a multi-step task through to completion.
Official data shows that under the same prompt, the High Effort path generates roughly 7 times the number of tokens compared to Low Effort. The additional computational resources are not spent on “smarter thinking,” but entirely on reading files, running verifications, and repeatedly confirming results. This leads to a counterintuitive conclusion: a smaller model like Sonnet paired with High Effort can absolutely outperform a larger model like Opus running on Low Effort in specific tasks.
Anthropic has painted clear “personas” for its different models. Sonnet is portrayed as “the all-rounder with an entire afternoon” — it will read code from start to finish, execute tests, and thoroughly digest the task. Opus is “the expert who only gives you five minutes” — it strikes at the heart of the problem based on experience, but lacks the time to scan every file. As for the most expensive model, Fable, that’s “the expert you only call in when everyone else is stuck,” reserved for the toughest challenges. Official testing shows that in long-duration tasks and multi-step workflows, Fable pulls ahead by the widest margin, with certain tasks proving insurmountable for Opus and Sonnet at any Effort level.
Beneath this “brain fog” episode lies a shift in the competitive landscape of AI-powered programming. The race is moving from “whose model is stronger” to “who is better at orchestrating agents.” Developers need to act like project managers: assigning Sonnet on Low Effort for simple code changes to achieve sub-second responses and cost control, deploying strong models with High Effort for major refactoring, and equipping agent tasks that require extended autonomous operation with ample firepower. The new “ultracode” tier that Anthropic added to Claude Code is a direct product of this philosophy — it grants the AI the authority to autonomously decide whether to spin up a team of agents to decompose and parallelize a task.
Meanwhile, competitor OpenAI is accelerating its pursuit on the same track. According to reports from multiple tech media outlets including Synced, on July 9, OpenAI released its new GPT-5.6 model along with the brand-new ChatGPT Work feature, directly targeting Claude Cowork. In simulated real-world task tests, GPT-5.6 Sol, operating at extremely high reasoning intensity, delivered results with higher completion quality and could even generate publicly shareable web links. In a separate comparison test involving creating a “user manual” for public figure Luo Yonghao, ChatGPT Work produced a web design with a distinct Smartisan product-launch aesthetic, complete with print stylesheets and local progress-saving functionality, while Claude Cowork leaned more toward a standard viral-content template.
However, OpenAI does not hold the advantage across the board. Testing revealed that ChatGPT Work suffers from excessively rapid Agent quota consumption — just a few tasks can exhaust a subscription’s quota, with billing hidden outside the subscription. Anthropic’s Claude Cowork, on the other hand, retains unique strengths in desktop workflows such as handling local files and organizing meeting materials. In a competitive analysis slide that ChatGPT Work generated about itself, it objectively noted: “Cowork is more mature; Work is broader,” and recommended that production-oriented pilot programs prioritize Cowork.
This industry-wide shockwave triggered by the “Effort” setting ultimately points to a clear conclusion: the era of judging by model leaderboards alone is passing. For developers and enterprise users, learning how to “assign work” to AI — precisely orchestrating combinations of different models and Effort levels — is becoming the core craft that determines productivity. Otherwise, even the most expensive model is just a pricier search box.