Anthropic’s Downfall

Recently, the global developer community’s anger erupted as Claude seemed to fall from grace overnight. On April 6, Anthropic launched Claude Code v2.1.92, introducing a new feature called Ultraplan. However, the launch coincided with an epic service outage.

Developers eagerly awaiting Ultraplan were met with “authorization failure” and “internal server error” messages. Claude Code bombarded users with login prompts, only to inform them they couldn’t log in. Many developers took to Reddit to express their frustration, calling the company a joke.

Additionally, Anthropic faced backlash for a controversial move: if developers attempted to modify system prompts, the backend would throw a “400 error.” This may have been a patch for a previous source code leak incident but sparked significant controversy.

Furthermore, developers reported on GitHub that Claude Code’s performance had severely declined since the February update, with the model failing to handle complex engineering tasks, producing unstable outputs, and even becoming unusable.

In just a few days, the service experienced three major failures, demonstrating that Anthropic’s rush to market has led to a focus on flashy new features at the expense of basic stability.

New Feature Launch

Major Outage on Launch Day

The Ultraplan feature, described as exciting, is not a new subscription plan but rather a new capability of Claude Code that shifts the planning process from local terminals to the cloud.

During the planning generation, users can free up their local terminals and not wait continuously. The new feature also provides a richer web interface to view various parts of the plan and add comments. Finally, users can choose to execute the plan in the cloud (requiring code to be in a GitHub repository) or send it back to the local terminal for execution.

This advancement aimed to elevate AI-assisted development, allowing AI not just to complete code but to help plan the entire execution path of a project.

However, the anticipated surprise quickly turned into a disaster, leading developers to joke that “Ultraplan” should be called “Ultralogin,” with some even dubbing it “OnlyPlans.”

Many users reported issues with Claude Code, with Opus being “dumbed down” and receiving prompts about reaching a five-hour usage limit after just two rounds of dialogue.

On Reddit, developers expressed their frustration, suggesting that rather than rushing to release new features, Anthropic should focus on creating a stable product. Frequent system crashes are unacceptable.

Service Status Update

Anthropic’s official status page later recorded that at 15:45, login errors were detected, confirmed to affect Claude Code by 15:54, with repairs completed by 17:16. The outage lasted for 90 minutes, impacting various functionalities across Claude.ai and Claude Code.

By April 7 at 01:16, Anthropic announced the issue was resolved but did not explain the cause. According to third-party monitoring platform IsDown, 1,212 user reports about Claude were received in the past 24 hours.

Even after claiming to have resolved the issue, many users continued to face access problems.

Token Consumption Issues

This is not the first incident. The source code leak incident is still fresh in memory, and now the launch of a new feature has led to a collective outage. While promoting cloud planning as a productivity booster, the core stability has become a joke.

Developers seem to be gambling on whether the servers will hold up. Despite the issues, Ultraplan has shown some effectiveness. A few early users acknowledged that the browser-based plan review is indeed enjoyable.

Inline comments, emoji feedback, and side navigation have significantly improved iteration efficiency compared to scrolling through text in the terminal.

However, user feedback is highly divided: during execution, the probability of failure in non-Git repository environments is high. While cloud planning is powerful, it can completely fail due to network fluctuations, token limits, or server issues.

Anthropic has labeled this as a “research preview,” indicating potential problems at any time.

Clearly, Anthropic bets on “cloud computing power + Opus model can outpace all local solutions,” while developers prioritize stability, cost, and control.

Major Concerns: Token Consumption

Moreover, Ultraplan has other significant drawbacks, such as being a complete token consumption machine. Previously, Claude faced scrutiny for an unexplained increase in consumption speed after updates, and this time is even more exaggerated.

The reason is that Ultraplan is a “cloud-first” workflow that generates detailed plans in the cloud, consuming tokens at an alarming rate. Hard-core video bloggers have reported that when processing complex architectures, cloud planning is twice as fast as local execution.

The cost? An exponential increase in token consumption!

What could have been lightweight local code validation is now forced onto the cloud pipeline. Each time a user generates a satisfactory plan with Ultraplan, tokens from their plan are rapidly consumed, leading to financial strain on developers.

Additionally, Anthropic has been accused of another alarming move.

New Restrictions on Custom Prompts

Peter Steinberger, known as the “father of Lobster,” revealed that Anthropic has begun cracking down on “custom system prompts.” Developers attempting to modify the system prompt for a more personalized experience encountered a “400 error” from Anthropic’s backend.

Some believe this is a response to the previous source code leak incident. However, users paying a premium subscription fee are unable to freely define AI behavior, leading to significant community skepticism.

Declining Performance Since February Update

The most distressing issue for core engineers is not just the outages or costs, but that Claude Code has genuinely become less capable. Developers on GitHub have reported that since the February update, Claude Code has struggled with complex engineering tasks.

One developer, whose team operates in a highly complex development environment, compared several months of historical logs and concluded that since February 2026, Claude has shown significant performance degradation in handling multi-file logic, long context retention, and engineering reasoning.

For example, it has ignored explicit instructions or claimed to have completed tasks when no code was executed, and performed operations contrary to requests.

This feedback resonated strongly in the comments section. Many enterprise users utilizing Claude in real production environments shared similar sentiments: the AI has begun to exhibit a form of “engineering laziness,” which is quite disappointing.

Developer Patience Running Thin

Faced with a barrage of bugs, the community has reached a consensus: Anthropic is too hasty. In the race for AI supremacy, they prefer to pile on new features around the clock rather than resolve core product stability issues.

Perhaps it is indeed easier to attract attention this way than to fix bugs. However, for a paid productivity tool, developers seek a stable product rather than flashy features that cause system crashes.

Currently, Claude’s reliance on being “more user-friendly and engineering-savvy” is starting to backfire. Some teams have reduced their usage of Claude from 80% to 20%, shifting to more stable and cost-effective local solutions, resulting in API costs dropping by hundreds of dollars.

For AI giants, this is a dangerous signal. As they rush to market, we hope Anthropic does not forget its original intentions.