Claude Code's Big Update: Opus 5, Cross-Session Chat, and a Rate Cut
Claude Code's last two months in one place: Opus 5 as default, cross-session messaging, /design, a security plugin, and a rate limit change worth understanding.
What’s new in Claude Code right now?
Claude Code shipped an unusually large batch of updates over roughly two months: Opus 5 became the default Opus model, sessions gained the ability to message each other, a new /design command brought a visual UI workflow into the CLI, a security scanning plugin arrived, and auto mode became the default permission setting. Alongside all of that, Anthropic also announced a rate limit change that looks like an increase on paper but functions as a cut for anyone used to the boosted limits of the past few months.
TL;DR
- Opus 5 rolled out in late July as the default Opus model across Pro, Max, Team, Enterprise, pay-as-you-go, and the API, with a 1 million token context window available on the API and on Max, Team, and Enterprise plans.
- Cross-session messaging lets separate Claude Code sessions on the same machine discover each other and pass messages, so one session can tell another about a renamed database column or an API change without you copying context manually.
- The new /design command, in research preview, brings a canvas of editable UI artboards into the CLI and desktop app so you pick a direction visually instead of describing it in text.
- Auto mode is now the default permission mode for new sessions on Pro, Max, and Team, using a background classifier to approve safe actions and block risky ones, and those classifier calls no longer count against usage limits.
- A new Claude security plugin runs a multi-agent vulnerability scan (architecture mapping, threat modeling, vulnerability hunting, independent review) across a repo, branch, diff, PR, or commit.
- Anthropic’s rate limit change raises weekly limits by a permanent 25% starting September 14, but because it replaces a temporary 50% boost, it works out to roughly a 17% reduction compared to current usage, something Anthropic itself acknowledged after initially framing it as a straightforward increase.
- Smaller additions include fork mode for background sub-agents that inherit full context, a concise output style, GitLab support, a /doctor command, screen reader mode, and self-hosted environments in public beta for Team and Enterprise.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
What does Opus 5 actually change for developers?
Opus 5 is now the default Opus model in Claude Code, available across Pro, Max, Team, Enterprise, pay-as-you-go, and the API. The headline spec is context length: a 1 million token window on the API and on Max, Team, and Enterprise plans, which matters for anyone working across large codebases where the model previously had to lose track of earlier files or decisions. Fast mode has also moved to Opus 5, priced at $10 per million input tokens and $50 per million output tokens.
The practical difference shows up most on long agentic tasks, the kind where a coding agent runs for many steps in a row. Earlier Opus versions had a tendency to drift off track the longer a session ran. Opus 5 holds context better over these extended runs, which is the main reason it’s worth switching to if you haven’t already. You can select it through the model picker or by typing the model command directly.
How does cross-session messaging work?
If you run multiple Claude Code sessions in parallel on the same machine, which is now common for anyone juggling separate features or services, those sessions can find each other and send messages directly. Say you’re renaming a database column in one session while another session is deep in payments API work. Instead of switching windows and explaining the change yourself, you tell the first session to notify the other, and it does.
You can also address a specific session by name using the @ symbol in a prompt. The feature started on macOS and Linux in early August and has since reached native Windows. It pairs naturally with fork mode, which is now on by default: instead of spinning up a sub-agent from scratch, Claude can fork a sub-agent that inherits the full conversation history and prompt cache, so side tasks don’t require re-explaining context. You can trigger this manually with a subtask command, the fork runs in the background, and results land back in your main conversation when finished. Anthropic also removed the previous cap of 200 sub-agents per session, with sessions now running up to 20 concurrently by default.
Together, these changes point at the same idea: Claude Code is being built around running several agents at once rather than one long linear conversation.
What is the /design command and is it useful?
/design is a research preview feature that brings Claude’s artboard-style design workflow into the CLI and desktop app. You give it a brief, something like “redesign the composer,” and Claude generates a canvas of multiple editable artboards showing different UI directions. You open the canvas, pick the option closest to what you want, adjust it, then tell Claude which one to implement.
One coffee. One working app.
You bring the idea. Remy manages the project.
The value here is straightforward: instead of Claude guessing at a single interpretation of a vague instruction and you iterating through several rounds of text feedback, you get visual options up front and choose. For front-end work in particular, this cuts down the back-and-forth considerably. It’s available on Pro, Max, Team, and Enterprise plans.
What changed with permissions and auto mode?
Auto mode became the default permission mode for new Claude Code sessions on Pro, Max, and Team starting August 14. Rather than prompting you for approval before every action, a background classifier evaluates each action, lets the safe ones through automatically, and blocks or flags the risky ones. If you’d already set a custom default permission mode, that setting is preserved unless you actively accept the switch.
This is the kind of change that splits opinion. Automating permission decisions clearly speeds up workflows, and Anthropic addressed one common complaint by making classifier calls free, meaning they no longer count against usage limits, which had previously acted as a hidden cost of running in this mode. But handing more judgment calls to an automated classifier won’t sit well with everyone, especially people who want tight control over exactly what an agent touches. You can still switch modes at any time.
What other features shipped that are easy to miss?
A handful of smaller updates round out the update cycle:
- Claude Security plugin: installed via the plugin marketplace, this runs a multi-agent scan that maps a codebase’s architecture, builds a threat model, hunts for vulnerabilities, and has agents independently review each other’s findings before producing a report. It can scan a full repository, a branch, a diff, a PR, or a single commit, and turns findings into patches you review before applying.
- Background code review:
/code-reviewnow runs as a sub-agent with its own context window, keeping review output out of your main conversation, and/reviewnow works as an alias for the same command. - Desktop app improvements: a built-in browser lets Claude pull up documentation or designs directly, and on macOS there’s a public beta iOS simulator pane that streams a live view of the app while Claude interacts with it.
- Remote control: out of research preview, letting a desktop session appear as a device card in the Claude mobile app so you can pick up a session from your phone.
- GitLab support: merge request URLs now work with the worktree flag, and plugin marketplaces can clone directly from GitLab.
- Other additions include a
/doctorcommand for diagnosing and fixing setup issues, a screen reader accessibility mode, a concise output style that skips preamble, and self-hosted environments in public beta for Team and Enterprise, letting sessions run on internal infrastructure.
Why are people upset about the new rate limits?
In May, Anthropic doubled 5-hour session limits, removed peak-hour throttling, and added a temporary 50% boost to weekly limits, attributing the extra capacity to new compute. That boost kept getting extended through several deadlines over the following months. On August 30, Anthropic announced that starting September 14, weekly limits would rise by a permanent 25% over the original baseline for Pro, Max, Team, and seat-based Enterprise plans.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
The catch is that the temporary 50% boost expires the day before, on September 13. If your baseline was 100 units, the 50% boost meant you were getting 150. After September 14, the permanent 25% increase means you’ll get 125. Compared to what heavy users have grown used to over the past several months, that’s roughly a 17% reduction, not an increase. Anthropic’s own follow-up post acknowledged this math directly after an initial announcement was posted, deleted, and reposted with clarified language.
To be fair, a permanent 25% bump over the original baseline is better than simply letting the temporary boost expire with no replacement, and the doubled 5-hour limits from May remain in place. Claude Code also now auto-continues a session once a usage limit resets, and usage credits are available beyond the cap at API rates for anyone who needs more. Still, the framing drew criticism, since presenting a real-terms cut as a permanent increase is the kind of thing that erodes trust faster than the limit itself.
Frequently Asked Questions
Is Opus 5 available on the free tier of Claude Code?
The rollout covers Pro, Max, Team, Enterprise, pay-as-you-go, and the API. The 1 million token context window specifically applies to the API and to Max, Team, and Enterprise plans.
Do I need to enable cross-session messaging manually?
No specific opt-in was described beyond running multiple sessions on the same machine; sessions can discover each other automatically, and you can address a session directly by name using the @ symbol.
Does auto mode make Claude Code less safe?
Auto mode uses a background classifier to approve routine actions and block risky ones automatically instead of prompting for every action. It’s optional if you’d rather keep manual approval, and you can switch modes at any time even after auto mode became the default.
Will my Claude Code usage actually go down after September 14?
If you were relying on the temporary 50% weekly boost introduced in May, yes. The permanent 25% increase replaces that boost and results in roughly 17% less weekly usage than what boosted-tier users currently have, according to Anthropic’s own clarification.
What’s the easiest way to try the new features?
Opus 5 can be selected from the model picker or by typing the model command directly. Features like /design and the security plugin are invoked with their respective slash commands once you’re on a supported plan.
