Claude Code raises 5-hour and weekly usage limits for the weekend

• ClaudeDevs said an earlier bug hit about 3% of Claude Code Max and Pro users with an incorrect weekly limit, and some were blocked from sending messages before the fix landed.
• A few hours later, ClaudeDevs said Anthropic reset both 5-hour and weekly usage limits for everyone, across all plans, for the weekend.
• User replies show why the reset landed well: AIandDesign said they had already burned an extra $100 in overages this month, while AIandDesign said they were ready to move to the $200 plan.
• The heavier sessions were not just casual chat. In side discussion, aakashgupta and aakashgupta described Claude Code runs that kept spinning on poorly specified /goal tasks until a judge model could verify completion.

You can read ClaudeDevs' bug note, jump to aakashgupta's terminal output, and compare that with his /goal guide. The useful detail is not just that limits got reset. The weekend bump landed right as users were describing longer agentic runs, retry loops, and verification-heavy workflows that chew through turns fast.

ClaudeDevs framed the change in two steps. First, the team said a bug showed the wrong weekly limit to a small slice of Max and Pro users, about 3%, and sometimes blocked messages. Later the same day, ClaudeDevs said both the 5-hour and weekly caps had been reset for everyone, across every plan, through the weekend.

That turned a narrow fix into a broader quota holiday. The post did not mention a permanent policy change, only a weekend reset.

The replies read like a price-pressure diary. One AIandDesign post joked about upgrading to Claude's $200 tier or falling back to GLM-5.2 after running out of tokens on a $100 plan. In a follow-up, the same account said they had spent an extra $100 on overages this month and were upgrading that night.

A separate AIandDesign post pushed the phrase "token optimizing" over "tokenmaxxing," predicting that coding-model prices would rise. When the reset arrived, AIandDesign said they used the extra room after Anthropic "pulling Fable over the weekend" to ship four projects without runaway loops or insane token usage.

The clearest explanation for those swollen sessions came from aakashgupta's /goal testing. He described a checker model that decides whether a Claude Code run is finished, but only from the transcript, not by opening files or running commands, according to aakashgupta's failure log.

His examples break into three failure modes:

• Vague finish lines: "production-ready" kept the judge from ever approving completion aakashgupta.
• Impossible tasks: an external dependency stayed down, so the agent kept retrying until told to log the block and move on aakashgupta.
• Compound goals: one prompt hid four separate finish lines, which made the checker grade against the wrong thing aakashgupta.

In a separate comparison, aakashgupta said the same model produced an unsourced stakeholder memo in a normal run, then produced four properly sourced sections and one explicit DATA MISSING flag when the job ran behind /goal. The weekend reset arrived in the middle of that kind of usage pattern, where the cost is tied less to one big prompt than to long supervised runs, retries, and proof printed back into the terminal.