Developers reported Claude suddenly felt less capable. Anthropic confirmed it quietly reduced the default effort level to save compute. The controversy raised transparency questions ahead of a potential IPO.
Anthropic is dealing with a wave of complaints from developers who say Claude has recently felt less capable, especially on complex coding and workflow tasks. The reaction has been loud enough to trigger speculation across GitHub, Reddit, and X, with some users posting side-by-side examples of older and newer outputs that they say show a clear decline.
This matters because Claude is not just a chatbot for casual questions, it is also used for coding and other work where mistakes can cause real problems. “Tokens” are chunks of text the AI processes, and “compute” means the computer power needed to run the model. If a company quietly changes how much effort an AI uses, people may get worse results without understanding why.
If you run Claude in an agent workflow, verify your model or tool settings now and check whether the effort level is set to medium by default. Test one representative task at low, medium, and high effort, then compare accuracy, latency, and token usage so you know the tradeoff for your own jobs. If your setup supports it, pin higher effort for code generation, multi-step reasoning, and tool-using agents, and keep a changelog note so future model changes do not surprise your team.
The issue centers on Claude Code, Anthropic’s coding tool that can read, write, and execute code in a developer’s environment. Users say it has become more likely to miss instructions, take shortcuts, and make basic mistakes on multistep jobs, according to Fortune’s reporting. One AMD senior director wrote online that Claude had regressed to the point where it could not be trusted for complex engineering work.
Anthropic later acknowledged that it had changed how Claude operates. Boris Cherny, who leads Claude Code, said the company reduced the model’s default “effort” level to medium after receiving feedback that Claude was using too many tokens, the units of data the model processes when it generates a reply. More tokens usually mean more compute, which is the processing power needed to run the model.
That explanation did not calm users. Many said the company had not made the change obvious enough, and some argued they only noticed after their agent started performing worse. Anthropic has said publicly that the change was listed in the changelog, a running update log, and Cherny said users of Claude Code can manually adjust effort levels. But the criticism has focused less on the setting itself than on how quietly it was introduced.
The controversy has also fed wider speculation about whether Anthropic is under pressure on compute, meaning the scarce and expensive GPU capacity needed to run AI models at scale. The company has not said it is degrading models because of capacity limits, and an Anthropic spokesperson has pushed back on that idea. Still, Fortune reported that Anthropic has faced recent outages, tighter usage limits during busy periods, and intense demand following the rapid growth of products like Claude Code.
The timing matters because Anthropic is now one of the most closely watched private AI companies in the market. Fortune reported that the company was recently valued at $380 billion and is believed to be moving toward an IPO, which would bring even more scrutiny to product quality, pricing, and communication with customers. Cherny said Anthropic will test setting Teams and Enterprise users to high effort by default so they get extended thinking, even if that means more tokens and more latency.
For now, the concrete question is not whether Claude can still work, but how much trust users are willing to give it after a change that many felt they should have been told about more clearly. Anthropic has already signaled that it may shift defaults again for some customer groups, and the next update to those settings will likely be watched as closely as any new model release.
