Maya: Hey builders! Here's what's on the radar today on Claude Code Review. Anthropic's handing out cloud credits, a 2.1.291 patch fixes dropped permissions, cloud sessions get a field guide, Sonnet 5.5 goes default, and Haiku 5.5 does too. I'm Maya.

James: And I'm James. Let's get into it.

Maya: So Anthropic is giving existing Claude Code subscribers a one-time cloud credit: one hundred dollars on Pro, two hundred fifty on Max. You claim it by October seventh at eleven fifty-nine PM Pacific, and some accounts get it automatically.

James: Right, and the key detail, Maya, is it only pays for cloud sessions. That's Claude Code running on Anthropic's virtual machine: it clones your GitHub repo, works on its own branch, and hands back a pull request, so work survives a closed laptop.

Maya: Exactly right, and eligibility is narrow. You needed an active Pro or Max subscription when the promotion began September twenty-third. Team, Enterprise, free trials, and new signups are excluded. To claim, type slash claim credit in Claude Code or use the Usage page banner.

James: And one catch worth flagging, Maya: unpushed local commits don't reach the session, and your home CLAUDE dot M D and personal skills don't travel either. The credit is drawn first, then usage falls back to your plan, and unused credit disappears November fourth.

Maya: Onto the patch. Claude Code 2.1.291 fixes two regressions, bugs earlier releases introduced. A 2.1.290 bug meant cloud sessions could drop answers to permission prompts, so what you approved wasn't reliably carried through to the agent.

James: Yep, and the second one, Maya, dates from 2.1.288: a session's last messages could be lost on quit. So anyone on 2.1.288 through 2.1.290 had at least one of these, and 2.1.291 addresses both.

Maya: Good framing, James. A separate commentary post from an evidence-tracking maker used the release to argue teams should be able to trace an approval from the human decision through to the tool call. That's the post's framing, not Anthropic's notes.

James: That ties into story three, Maya. The ClaudeDevs account announced a field guide on October sixth for cloud sessions. Each task runs on a fresh virtual machine, a brand new virtual computer started just for that task, instead of on your laptop.

Maya: And because each task gets its own VM, James, you can run several cloud sessions at once, and they keep going when you close your laptop. The guide is built around seven repeatable workflows, plus how to connect GitHub on the first try.

James: Exactly, and it's published at claude dot dev slash blog slash claude code in the cloud. So it pairs neatly with that credit: concrete ways to actually spend those cloud hours instead of guessing.

Maya: Now the models. Claude Code 2.1.284 made Sonnet 5.5 the default Sonnet. Picking sonnet in slash model lands there, with about a million tokens of context, and the rates are unchanged: two dollars input, ten output per million tokens.

James: And the pitch, Maya, is that Sonnet 5.5 benchmarks sit very close to Opus 5.5, especially on agentic coding, at about half Opus's cost. But a hands-on test complicated that with one loose prompt for a 3D mini golf game.

Maya: Right, and the numbers surprised them: Opus took about an hour twenty-six and cost about twenty-seven dollars, Sonnet about an hour twenty and about twenty-six. Barely cheaper. The builder preferred Opus on quality, and noted Sonnet does better with precise instructions.

James: Fair to stress, Maya, it's one run of one loose prompt, so it shows how the model behaved there, not how it prices across workloads. Teams who pinned Sonnet by alias should confirm the new default behaves on their own tasks.

Maya: And the companion release, James: Claude Code v2.1.293, out October seventh, makes Haiku 5.5 the default Haiku, with a million-token context. Pricing is ten cents input, fifty cents output per million, but prompts over one hundred K jump to a higher tier.

James: Good to flag that tier, Maya. It also adds an agentType field so scripts can tell custom subagents apart. And a nice fix stops Claude from treating its last actions before a context compaction as done, which had made it retract finished work.

Maya: Exactly, and there's a memory leak fix where an HTTP MCP connection kept every request until it closed. Two changes were reverted, and Team and Enterprise orgs get faster startup since settings are fetched earlier.

Maya: That's your Claude Code Review daily brief. Check claudecodereview.ai for the full stories. Keep building.