Claude Code Cloud Sessions GA: What $100 Buys

Anthropic moved Claude Code cloud sessions out of research preview on September 23, 2026, and every existing subscriber can claim a one-time credit to try them: $100 on Pro and $250 on Max, claimable until October 7. A cloud session is a Claude Code run on an Anthropic-managed virtual machine that keeps working after you close your laptop, and you can steer it from a browser, the Claude mobile app, the desktop app, or the terminal.

The headline number is the credit, and it is the part most people are reading wrong. The VM itself is free: Anthropic's own documentation says there is no separate compute charge. What the credit pays for is model usage, which means its real size depends on which model you run and how hard you push it. This article works out what $100 actually buys, compares cloud sessions with OpenAI Codex, Cursor and GitHub Copilot, and lays out a workflow that spends the credit on the work that benefits most from running unattended.

What changed on September 23

Cloud sessions have existed for months under the "Claude Code on the web" name and a --remote flag. The general availability announcement from the ClaudeDevs account made three practical changes: the feature is no longer labelled a research preview, the CLI flag is now claude --cloud (the old --remote spelling still works as a deprecated alias), and Pro and Max subscribers get the one-time credit, claimed from Anthropic's link or with /claim-credit in the CLI.

The launch post caused enough confusion that Anthropic followed up about four and a half hours later: cloud sessions run on your Pro or Max plan like the rest of Claude Code, and the promo is an optional credit that cloud sessions spend first before falling back onto normal plan usage. In other words, the credit is a buffer in front of your existing limits, not a new metered product. AlphaSignal's launch summary adds that the credit applies automatically when a cloud session starts and is kept separate from the account's regular allowance.

Eligibility, per the cloud sessions documentation, covers Pro, Max and Team plans, plus Enterprise users with premium seats or Chat + Claude Code seats. You need GitHub connected to start a session, and organizations with Zero Data Retention enabled cannot use cloud sessions at all.

What a cloud session actually runs on

Each session gets its own isolated VM with your repository cloned onto its own branch. Anthropic's cloud environments page lists the resources: 4 vCPUs, 16 GB of RAM and 30 GB of disk. Docker is available, so docker compose up works for projects that need a database or a local API alongside the code. The docs warn that the VM may stop tasks needing significantly more memory, such as large build jobs.

Every session runs inside a saved "environment" that sets three things: network access, environment variables and a setup script. Network access comes in four levels. None blocks outbound traffic, Trusted (the default) allows package registries, GitHub and cloud SDKs, Full allows any domain, and Custom takes your own allowlist. The setup script should finish in roughly five minutes, because Anthropic snapshots the filesystem after it runs and reuses that snapshot for later sessions. The cache rebuilds when you change the script or the allowed hosts, and otherwise expires after roughly seven days.

Two details matter for anyone with secrets in their workflow. On Pro and Max, API keys you add to an environment stay outside the sandbox and are attached to matching requests after they leave the session, so Claude can call an API without ever seeing the key. And git credentials and signing keys never enter the VM: a proxy authenticates on the session's behalf.

The credit is model usage, not VM time

The limitations section of the docs is explicit: cloud sessions share rate limits with all other Claude and Claude Code usage in your account, running tasks in parallel consumes limits proportionately, and "there is no separate compute charge for the cloud VM." So the $100 is not buying machine hours. It is buying model tokens that would otherwise come out of your five-hour and weekly limits.

Anthropic has not published the rate at which the promo credit draws down. The closest reference is the dollar-denominated usage it already sells to Pro and Max subscribers: usage credits are billed at standard API rates. If the promo credit works the same way, the arithmetic with Claude Opus 5.5 list prices ($4 per million input tokens, $20 per million output, $0.20 per million cached reads) looks like this:

The last column uses the $5.98 average cost per task that Artificial Analysis measured for Opus 5.5 on its Intelligence Index, a useful yardstick for a hard, self-contained problem. A long agentic coding session that runs tests, reads logs and iterates will usually cost more than one of those tasks, so treat 16.7 as a ceiling on what Pro gets from Opus, not a floor.

Caching is what stretches the budget. Agentic sessions re-read the same context on every turn, and a cached read is 20 times cheaper than fresh input. Re-reading a 200,000-token context costs about $0.04 from cache against $0.80 as fresh input. That gap is the reason long, focused sessions on one branch are far cheaper per useful change than many short sessions that each rebuild context from scratch. It is also why a cloud session is a better use of the credit than you might expect: Claude Code manages compaction and caching itself, and cloud sessions set their own auto-compaction threshold.

Coverage of the terms, including KuCoin's summary of the announcement, puts the credit's expiry at November 4. That gives a subscriber who claims on launch day six weeks to use it, and one who claims on October 7 four.

Cloud coding agents compared

Every major coding tool now has a "keep working while I am away" mode. They differ less in capability than in how they bill and where they stop.

Sources: OpenAI's Codex pricing and cloud pages,Cursor's Cloud Agents docs and GitHub's Copilot cloud agent docs, all read on September 24, 2026.

The split is clear. Codex is the most widely available, because it rides every ChatGPT plan. Cursor is the most flexible on source hosts but charges API rates from the first token, which is why it makes you set a spend limit before you start. Copilot draws on two meters at once and stops every session at 59 minutes. Claude Code and Codex both bill cloud work against the subscription you already pay for, with no second meter. Where Claude Code differs is that Anthropic publishes the VM size, and it is the only one of the four currently handing out a credit to try it. For the self-hosted end of the market, our look at Cursor Cloud Agents on your own infrastructure covers the option Anthropic also offers through self-hosted environments.

How to spend the credit well

The credit is best spent on work that is long, well specified and testable, because that is where running unattended saves you the most time. A workflow that follows the documented tips:

1. Claim first. Run /claim-credit in the CLI or use Anthropic's claim link before October 7. The credit does nothing until claimed.
2. Plan locally. Start claude --permission-mode plan on your machine. In plan mode Claude reads files and proposes an approach without editing code, which costs little and catches misunderstandings before they burn cloud usage.
3. Commit the plan and push. Save the plan to a file in the repo and push. The cloud VM clones your GitHub remote at your current branch, not your local checkout, so unpushed commits are invisible to it.
4. Send it to the cloud. Run claude --cloud "Execute the plan in docs/plan.md". Each call creates its own session, so independent tasks can run in parallel, each consuming limits in proportion.
5. Steer from anywhere. Check progress on claude.ai/code or the mobile app, or queue a follow-up from any logged-in machine with claude -p "your message" --cloud <session-id>.
6. Bring it home. When the work is done, open a pull request from the session, or run claude --teleport to pull the branch and full conversation into your terminal.

Two settings change the cost of the same job. Cloud sessions accept /model and /effort as arguments, so a mechanical refactor does not need Opus at maximum effort; our Opus 5.5 breakdown covers what the effort setting changes. And auto-fix, which watches a pull request and pushes fixes for CI failures and review comments, can keep a session working, and consuming usage, for as long as the PR stays open. Turn it on for PRs you actually want babysat.

Who should skip it

Cloud sessions are not for every setup. Teams whose code lives on GitLab or Bitbucket can upload a bundle, but the session cannot push results back to that remote. Anything that needs more than 16 GB of memory will hit the VM ceiling. Organizations with IP allowlisting on their Claude account will see every Anthropic-hosted session fail authentication until support exempts the service. And if you use Claude Code through Amazon Bedrock or Google Cloud rather than an Anthropic account, --cloud is not available to you.

For those cases the docs point to two alternatives. Remote Control keeps the session on your own machine while letting you steer it from a phone or browser, and self-hosted environments run cloud sessions on compute your organization operates. Anyone already watching their weekly limits should also read our note on the September weekly limit change, because cloud sessions draw from the same pool once the credit is gone.

What to do next

If you are on Pro or Max, claim the credit now even if you do not plan to use it this week, because the claim window closes October 7. Then pick one real task you have been putting off, such as a flaky test suite or a dependency upgrade across a repository, plan it locally, and send it to a cloud session before you log off. That single run will tell you more about how fast the credit drains on your codebase than any per-token arithmetic.

Frequently asked questions

Is Claude Code cloud sessions free?

There is no separate charge for the cloud VM. Sessions use the same Pro, Max, Team or Enterprise plan limits as the rest of Claude Code. The one-time launch credit ($100 on Pro, $250 on Max) is spent first, then usage falls back onto your normal plan limits.

How do I claim the $100 or $250 credit?

Run /claim-credit in the Claude Code CLI or follow Anthropic's claim link, by October 7, 2026. You need GitHub connected to start a session, and the credit applies automatically once a cloud session starts.

What hardware does a cloud session run on?

Each session runs in an isolated Anthropic-managed VM with 4 vCPUs, 16 GB of RAM and 30 GB of disk. Docker is available, and a setup script can pre-install tools, which Anthropic caches for roughly seven days.

Can I start a cloud session from the terminal?

Yes. claude --cloud "task description" creates a new session for your current repository. The older --remote flag still works as a deprecated alias. Use claude --teleport to pull a finished session back into your terminal.

Do cloud sessions work without GitHub?

Partly. From a repository with no GitHub remote, Claude Code uploads a bundle of your local repository, which must be under 100 MB. The session can work on it, but it cannot push results back to a non-GitHub remote.

How does it compare with Codex cloud and Copilot's cloud agent?

Codex cloud is included in every ChatGPT plan, including Free, and counts against the plan allowance. Copilot's cloud agent uses GitHub Actions minutes plus AI credits and stops each session at 59 minutes. Claude Code cloud sessions share your plan limits with no separate VM charge.