Claude Code fixes parallel subagent overload and resets 5-hour limits

• Anthropic says ClaudeDevs' reset notice reset both 5-hour and weekly Claude Code limits for Pro and Max users after a bug burned usage faster than expected.
• According to ClaudeDevs' follow-up, the fault sat in how Opus 4.8 requests were handled, where the model triggered more parallel tool calls than intended.
• The company also said in the same follow-up that the issue was unrelated to Claude Code's new dynamic workflows feature.
• Anthropic's own Claude Code cost docs already note that usage can swing with multiple instances or automation, which helps explain why a subagent fan-out bug would hit subscription caps so hard.

Anthropic had already positioned Claude Opus 4.8 around more efficient tool use and Claude Code's new dynamic workflows. A few days later, the ClaudeDevs account ended up resetting everyone's buckets after a request path started spawning too many parallel workers. You can also trace the practical fallout in Anthropic's own usage docs, which spell out that automation patterns can change consumption fast.

Anthropic's fix came with compensation first. The official reset post says Pro and Max users got both their 5-hour and weekly limits refreshed after sessions started chewing through usage unexpectedly.

That is a bigger gesture than a quiet patch note, because Claude Code subscriptions are normally bounded by plan usage bars rather than pay-as-you-go billing, as Anthropic explains in its cost docs.

Anthropic's explanation is narrow: the follow-up post says Opus 4.8 request handling caused more parallel tool calls than intended. The company did not describe a model quality regression, a pricing change, or a general outage.

The wording matters because Claude had just marketed Opus 4.8 as more efficient at tool calling in its launch post. What broke here was the request path around the model, not the broader promise that tool use got better.

Anthropic explicitly separated the bug from dynamic workflows. In the Opus 4.8 announcement, the company described that feature as Claude Code's way to tackle very large-scale problems by orchestrating bigger jobs.

So the official line is two-part:

• Dynamic workflows are a real product feature in Claude Code, per the Opus 4.8 launch post
• The runaway usage came from unintended parallel tool calls in Opus 4.8 request handling, per ClaudeDevs' incident clarification

That distinction should save some confusion, because both stories involve parallel work but only one was the bug.

Anthropic's Claude Code cost documentation adds one useful implementation detail: Pro and Max subscribers see usage bars and activity stats, while the session dollar figure in /usage is only a local estimate and is not what subscription users are billed on.

The same doc also says per-developer usage can vary widely with model choice, codebase size, multiple instances, and automation. That makes this incident unusually concrete: a parallel-subagent mistake was exactly the kind of behavior the usage model is sensitive to, and Anthropic ended up resetting the caps after fixing it.