Cloud sessions are now generally available, letting Claude Code continue running on Anthropic-hosted infrastructure even when the user's laptop is closed. Existing Pro and Max subscribers receive a one-time credit of $100 and $250 respectively.
Cloud sessions run on your existing Pro or Max plan subscription. The one-time promo credit ($100 for Pro, $250 for Max) is an optional top-up that gets spent first, with cloud session usage falling back to normal plan limits once the credit is exhausted.
The one-time cloud session credit can be claimed at the link or by running /claim-credit in the CLI. GitHub must be connected to start a session, and the offer expires October 7.
Cloud sessions run on Anthropic-hosted infrastructure, keeping work going with the laptop closed. You can start one from claude.ai, the Claude mobile app's Code tab, the desktop app, or with claude --cloud in the CLI.
Projects, which previously ran threads only as cloud sessions, now also supports local execution so threads can run directly on your own machine.
The team is working through the Projects early access waitlist and granting access to more Pro and Max users. Users who previously used Claude projects are still being migrated over.
A user reports that each subagent spawns its own Node process, causing severe resource exhaustion that led them to switch to Codex. The team acknowledged the problem and says a few solutions are being evaluated, each with tradeoffs.
After community feedback split between "don't need it" and "love the brainstorming UX," plan mode will live on as a built-in mod with customizable prompts, user-created modes, and shift+tab rebinding, rather than being removed.
An Anthropic team member responded to speculation about plan mode's future, saying their probability of killing it is higher than the 10% someone else suggested.
A user reports using plan mode as their default for non-trivial tasks, essentially replacing manual research-and-summarize workflows. A team member replies that custom planning patterns are emerging (via interviews, artifacts, etc.) and that the default behavior may need updating.
Blocked requests in three narrow categories (biology, distillation attacks, and frontier LLM development) are now billable again, after coordinated attacks on Anthropic systems. The classifiers have a sub-0.1% false positive rate, and 99.7% of users are unaffected. Users who believe a request was wrongly blocked can report it via /feedback in Claude Code.
The workflow uses Opus to model specific state machines and race-prone code paths, find counter-examples (suspected bugs), reproduce them, and fix them. It's not full codebase verification but targeted formal modeling of the hairiest parts, combined with TLA+ for data flow and concurrency checks.
Echoing advice to stop fixating on model releases, adocomplete argued that Opus 5.5 paired with a well-built harness and environment unlocks almost unlimited capability, and that environment quality matters more than which model you're running.
Cloud sessions in Projects are not separately paid. They draw from your existing plan's rate limits, with a bonus one-time credit applied first to let you try them easily.
Cloud sessions are not a paid add-on. They are included in Pro and Max plans and count against normal usage limits. The one-time credit for Pro/Max users is just a trial buffer, not a new billing tier.
A bug flagged by a user related to cloud session behavior has been resolved in version 2.1.281, released the same day.
