Anthropic is softening the most inconvenient scenario in Claude Code: when the 5-hour limit is reached, the agent will no longer cut off a task mid-edit. Instead, it will attempt to bring it to a logical point of completion, spending a small, fixed reserve that is deducted from the user's weekly quota. The reserve works differently across tiers, and continuing work after the session ends automatically will require separate usage credits.
What happened
On September 25, 2026, the Anthropic team announced the change in its official ClaudeDevs account. Claude Code—an agentic terminal tool for coding—previously cut off a session immediately upon reaching the 5-hour usage window, even if it was in the middle of rewriting code. Now, in such a situation, the agent will attempt to find a graceful stopping point—that is, a logical point of completion—and finish the current task by spending a small, fixed reserve. This reserve is deducted from the user's weekly quota, so hitting the 5-hour window no longer means an immediate loss of work progress. The rollout is gradual and depends on the tier: on Pro, the reserve for finishing a task is available once a week; for Max and Team Premium users, it is available every time the limit is reached. If the session does end automatically, additional usage credits will be required to continue working.
Context
A hard time-based cutoff has long been considered the main practical drawback of agentic tools: a session would die mid-refactor, the agent would lose context, and unfinished edits would remain in the codebase, requiring manual cleanup. In this case, Anthropic is not releasing a new model or claiming a scientific result—the company is changing its limit policy and the very lifecycle logic of an agentic session, replacing an emergency cutoff with a planned stopping point. At the same time, the structure of the limits becomes more complex: in addition to the 5-hour window and the weekly quota, there is now a fixed reserve for finishing a task and purchasable usage credits—four levels of consumption, each of which must be tracked separately. The workflow for limits, the /usage command, and usage credits is described in the official Claude Code documentation on managing costs.
Why this matters for the industry
For the industry, the very model of limits for agentic coding is changing: the hard cutoff at the 5-hour window is giving way to a soft-completion mechanism paid for from the weekly quota. The agent now optimizes not for 'squeezing out the maximum before the cutoff' but for 'reaching a logical point of completion' while accounting for the remaining budget—this is a new pattern in the agent's lifecycle, not a cosmetic tweak. Long, multi-hour runs benefit most: migrations and large refactors are less likely to end in a broken state of the codebase. At the same time, Anthropic is building a monetization point beyond the subscription: the notification that the reserve has been exhausted becomes a natural moment to offer purchasing usage credits. The tier dependence is also telling: a weekly reserve on Pro and a reserve every time the limit is hit on Max and Team Premium indicate that this is primarily a pricing decision, not a technical necessity, so the behavior of the same tool differs across subscriptions. If the mechanism takes hold, it is expected that competitors will start copying 'soft' limits with completion reserves, and hard cutoffs will become entrenched as an antipattern.
Why this matters for users
For readers, this changes how work with the tool is planned. The rollout is gradual, so first wait for the feature to appear in your account: until then, hitting the 5-hour window triggers the previous behavior of cutting off the session. A large refactor still should not be started 'right on the edge' before the weekly quota resets, and the longest run of the week is better planned closer to the start of a new weekly window rather than its end. The /usage command shows both the remaining time in the 5-hour session and weekly consumption—make it a mandatory part of your routine, because the weekly quota now acts as a budget from which both regular work and the reserve for finishing tasks are paid, and it should be allocated consciously.
What is still unknown / limitations
How exactly Claude Code determines a logical point of completion has not been disclosed; it is also not explained whether the fixed reserve is guaranteed to be enough to finish a specific task, or what will happen if the reserve is insufficient. For now, this is a statement about product behavior, not a verifiable result with a described methodology. The rollout is gradual, so some users will continue to work under the previous model of session cutoffs for some time, and there is no accumulated statistics yet on how often a graceful stopping point actually saves a task rather than spending the reserve without a result.
Sources
- ClaudeDevs (Anthropic) — X post on graceful stopping point at the 5-hour limit
- Claude Code Docs — Manage costs effectively
Author
Look at AI, editorial team