Important updates from Anthropic, taken from the company's own pages and rewritten in plain language. Covers the Claude models, the API, Claude Code and the Claude apps. Sorted by when we found them, and every entry links to the source.
The Python library for the Claude API has reached version 1.0. Its HTTP layer, the part that actually sends the requests, moves from httpx to httpx2, a maintained fork with the same interface. The release removes a good deal of old surface: the retired Text Completions API, the temperature, top_p and top_k parameters on the Messages methods, and compaction_control in the tool runner. On the async client, .with_raw_response now has to be followed by await response.parse().
Why it matters
This is a release that can break code that worked yesterday, so the upgrade deserves a deliberate pass rather than a routine package bump. If you use tools that trace or mock httpx, call httpx2.alias_httpx() at startup to keep them working.
Claude Code has a built-in writing style called Concise. It leads with the result and skips preamble and narration, while doing the work just as thoroughly. The same version fixes a bug where prompt caching, the mechanism that makes repeated requests cheaper, did not work for sessions running through an LLM gateway or a custom base URL.
Why it matters
The caching bug is the more important of the two: it hit the bill with no visible sign, so anyone running Claude Code through their own gateway has probably been paying more than necessary until now.
The tool that lets Claude drive a computer is out of beta as computer_toolset_20260801. It no longer needs its own header, it can perform several actions in one turn, and zoom is on by default. Alongside it comes browser_toolset_20260801, a tool for driving a browser your own application hosts. It works inside the browser viewport and reads the page directly, meaning its structure, elements, forms and tabs, rather than only looking at a screenshot and clicking.
Why it matters
Reading the page as structure rather than as an image is the difference between guessing where a button is and knowing. If you already use computer use, the upgrade changes the shape of your requests, so it is a job rather than a switch.
Platform
Claude API. Works on Claude Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8.
First detected by us
August 22, 2026
Availability
No beta header needed. Earlier beta versions still work.
Three parts of the API left beta at once. The Files API no longer needs the files-api-2025-04-14 header, and without it returns the current format with file expiry and paginated listing. Agent Skills and /v1/skills no longer need skills-2025-10-02. The user-management endpoints in the Admin API for Claude Enterprise, covering members, invites, groups and roles, are out of beta too.
Why it matters
Out of beta means the interface can no longer change without notice, and that is what decides whether something can sit in production. Note that the Files API behaves differently with and without the header: drop it and the response format changes.
Platform
Claude API. The admin change applies to Claude Enterprise organizations.
First detected by us
August 22, 2026
Availability
Requests that still send the beta header keep working as before.
You can now decide which sites the web_search and web_fetch tools are allowed to reach, using allowed_domains or blocked_domains. web_fetch also accepts max_content_tokens, and web_search accepts user_location. In the same release, Managed Agents sessions running in a self-hosted sandbox can attach memory stores, which are downloaded into the sandbox at start and synced back.
Why it matters
An agent that can fetch anything from the web is also an agent that can be reached through the web. A domain list is the simplest way to bound what it can actually read, and it belongs in any setup that runs without a person watching.
The ANTHROPIC_DEFAULT_MODEL environment variable sets the model new sessions start on, while a /model pick still overrides it and is remembered. That is what separates it from ANTHROPIC_MODEL, which cannot be overridden the same way. Cross-session SendMessage gains notify_when_idle: you can ask another Claude Code session on the machine to send one notice the next time it goes idle. The macOS sandbox also tightens, so rules like **/.env cannot be bypassed by renaming the file.
Why it matters
The notice replaces checking in on another session at intervals to see whether it has finished. It is one-shot and has to be asked for, so it does not become a stream of messages.
Platform
Claude Code, version 2.1.236. The notice works on macOS and Linux.
The tool where you try requests against Claude in the browser has been renamed from Workbench to Playground. It now supports every Messages API parameter and includes templates demonstrating API features such as code execution and web search. For each run it shows both the request as it looks in the SDK and the response the API returns.
Why it matters
Seeing the finished request beside the response makes Playground a learning tool rather than just a sandbox: you can copy what worked straight into your own code.
When you hit your plan's usage limit, Claude Code now waits and continues the session itself once the limit resets, instead of leaving you to come back and restart it. The setting is called "Continue automatically at usage limit" and can be turned off. The same version closes a security hole: file paths written with the Windows NT namespace, the \??\ form, are now rejected everywhere Claude reads files before you have approved anything, because they could be used to leak Windows login credentials.
Why it matters
The security fix is the important half, and it applies to everyone on Windows. It closes the last route into a class of attack where a file could lure Claude into connecting to a machine that steals your credential.
Platform
Claude Code, version 2.1.234
First detected by us
August 22, 2026
Availability
Applies to claude.ai plans. Can be turned off in /config.
Future Claude models will produce text containing a watermark, a pattern that can be used to estimate the likelihood that Claude helped write the text. It works by having the model use a key rather than a random number when choosing between words that are roughly equally good. Anthropic stresses that nothing is added to the text, that there are no hidden characters, that it costs no extra tokens, and that the mark cannot be traced to a person, an organization or a conversation.
Why it matters
As of August 2 the EU requires AI providers serving its market to mark AI-generated content, and the rules reach the wider European Economic Area too. Several major model developers have signed the same Code of Practice and are adding their own watermarks, so this becomes the norm rather than a Claude quirk.
Platform
Future Claude models
First detected by us
August 22, 2026
Availability
Applies to everyone. Introduced to comply with the EU AI Act.
The tools that let Claude keep a task list as it worked, meaning TodoWrite and the TaskCreate family, are no longer available on the newest models. They remain on older models, and can be switched back on with an environment variable.
Why it matters
If you have setups, skills or workflows that assume the task list exists, they stop working quietly after an upgrade. The environment variable is the fallback until you have fixed them.
Platform
Claude Code, version 2.1.233. Affects Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and newer.
First detected by us
August 22, 2026
Availability
Can be brought back with CLAUDE_CODE_ENABLE_TODO_TOOLS=1.
Type @ in the prompt and you can mention another Claude session by name, and Claude reaches it through SendMessage. Sessions on the same machine now get guaranteed unique names, so there is no doubt about who you are addressing. At the same time, subagent forking becomes the default: a subagent of type fork inherits the whole conversation and the prompt cache, and agents spawned in an interactive session run in the background.
Why it matters
This turns several concurrent sessions into something that can collaborate instead of being separate windows. A fork inheriting the cache also makes it cheap to start, since the context does not have to be sent again.
Across three versions Claude Code has gained GitLab support on a par with GitHub. Merge requests show in the footer as MR !N with status, --worktree accepts a merge request URL, plugin marketplaces can live on gitlab.com including nested subgroups, and nine families of GitLab token are now redacted from logs. The glab config directory gets the same protection gh already had.
Why it matters
The token redaction is worth noting even if you never touch GitLab: without it an access credential could end up in a log or a shared session. The rest is parity, so your choice of code host no longer decides how well the tool works.
Platform
Claude Code, versions 2.1.232 through 2.1.234
First detected by us
August 22, 2026
Availability
The footer badge needs a GitLab remote and an authenticated glab.
The Compliance API, which organizations use to see what happens in Claude, now also returns transcripts from Cowork and Claude Code sessions running on employees' own machines. Three new endpoints list the sessions, fetch one session's metadata and fetch its transcript. The same round adds the anthropic-workspace-id response header, which reports which workspace a request resolved to.
Why it matters
For an organization with audit requirements, local sessions were a gap: invisible even though they ran on the company account. For the individual, it is worth knowing that local Claude Code sessions in an Enterprise organization can now be read by the employer.
Platform
Claude API. Beta for Claude Enterprise organizations.
First detected by us
August 22, 2026
Availability
Requires a Compliance Access Key with read:compliance_user_data.
The introductory price for Claude Sonnet 5, $2 per million input tokens and $10 per million output tokens, is now the standard price. The increase to $3 and $15 that had been scheduled for September 1, 2026 will not happen.
Why it matters
Anyone who had budgeted for a September price rise can strike it out. For anyone who was considering switching models to avoid the increase, the reason to switch has gone away.
Platform
Claude API
First detected by us
August 22, 2026
Availability
Applies to everyone using Sonnet 5 through the API.
Fable 5 has had strict safety filters around biology, and they fired too often on harmless questions. When a filter triggered, the question was routed to Opus 5, a model without the same biology capability. Anthropic has now tuned the filters so that happens far less. The company names everyday examples such as interpreting lab results, understanding symptoms and learning biology.
Why it matters
Ask Fable 5 about health and you should now more often get an answer from Fable 5 itself, rather than a weaker one with no explanation why. Anthropic stresses the filters remain: telling beneficial biology from harmful is genuinely hard, and they are keeping the restrictions where the risk is real.
A batch of fixes closes ways of slipping past approval. A crafted Bash command could hide parts of itself from the permission check, and so could commands padded with tabs or invisible Unicode characters, so the approval dialog showed something other than what ran. Workflow scripts could escape the sandbox using dynamic import(). An agent definition could put itself in bypass-permissions mode even where the organization had disabled it. On Linux and macOS, a block written with a trailing slash, such as denyRead: "~/.aws/", was silently bypassable.
Why it matters
This is the reason to update. Several of the holes come down to approving one thing while something else runs, and that is invisible to the person sitting there. The same round also makes sandbox denials appear in the result at last, so Claude sees which file or network address was refused.
Platform
Claude Code (CLI), versions 2.1.223 through 2.1.224
First detected by us
August 7, 2026
Availability
Released in Claude Code 2.1.223 (August 6) and 2.1.224 (August 7, 2026). The source flags neither as a security advisory or CVE.
The claude self-hosted-runner command turns your own machines or containers into a place Claude Code can run sessions started from web, mobile and desktop. At the same time, sessions can now message each other with SendMessage, across machines too, and find each other with ListAgents. Messages to a session running with permissions set aside are held for your approval.
Why it matters
Your code does not have to leave your infrastructure for you to steer a session from your phone. Sessions talking to each other opens the door to one job notifying another, which is also exactly why messages into a session with lowered guards have to be approved by hand.
Platform
Claude Code (CLI), version 2.1.224. Session-to-session messaging is on macOS and Linux.
First detected by us
August 7, 2026
Availability
Self-hosted runners require a Team or Enterprise plan.
A session in Claude Managed Agents can now be given a hard cap on what it may spend. On reaching it, the session pauses with the budget_reached stop reason instead of sending further model requests. Raise or remove the cap and it continues where it left off. A deployment can apply the same cap to every session it starts.
Why it matters
An agent stuck in a loop now costs a known maximum instead of an unknown one. Because the session pauses rather than dies, setting the cap too low does not lose the work.
The inference_geo field decides where the model inference physically runs. It goes in the model object when the agent is created, and can be overridden for a single session. Which regions exist, and what they cost, is in Anthropic's data residency documentation.
Why it matters
For European organizations this is often the difference between being able to adopt a service at all and not, since regulation may require data to be processed inside a given region. Note that pricing can vary with the region you pick.
Platform
Claude API (Claude Managed Agents)
First detected by us
August 7, 2026
Availability
Available geos and pricing are listed in the data residency documentation.
A session can now be given an advisor, meaning a model at least as capable as the agent itself that the main thread can consult mid-task for guidance. It is set up as an advisor entry in the agent's roster of helpers. Separately, a session automatically discovers skills sitting in .claude/skills in a GitHub repository it has mounted, at session start.
Why it matters
A cheap agent can do the legwork and check with a stronger model when it gets stuck, rather than running everything at the expensive tier. Skills following the repository means your way of working lives alongside the code, not in a separate setup.
Enterprise organizations can turn on automatic security scanning of third-party skills and plugins. The scan runs when someone uploads or edits one, looking for malicious content.
Why it matters
A skill is instructions Claude reads and follows, so a malicious skill is a real route into an organization. The check happens at upload, meaning before anyone gets to use it.
Platform
Claude Enterprise
First detected by us
August 7, 2026
Availability
Beta, Enterprise plans only. Has to be switched on.
The Claude Opus 4.1 model has been taken out of service. Requests to the model ID claude-opus-4-1-20250805 now come back as an error rather than a result. Anthropic recommends Claude Opus 5 as the replacement. Researchers who need continued access can apply through the company's external researcher program.
Why it matters
A setup still pointing at this model ID is already broken, not merely slower or more expensive. This is one of the reasons to write down which model IDs a service depends on.
Platform
Claude API
First detected by us
August 7, 2026
Availability
Removed. All requests to claude-opus-4-1-20250805 now return an error. Researchers can request continued access through the External Researcher Access Program.
With inference hooks, an organization points Claude at its own AI security server. Every governed request is held until that server says allow or deny, before the model sees it at all. Requests are signed, the company decides what happens when the server does not answer, and every denial lands in the compliance log.
Why it matters
A business can enforce its own rules at the moment something is sent to the model, rather than auditing it afterwards. The cost is an extra hop on every request, so the failure-handling configuration is worth thinking through.
A worktree is a separate checkout of a project, meant to keep a session away from the main folder. Sessions that should have been confined to their own copy could still run destructive git commands against the shared checkout. Isolation now covers both file edits and Bash in every kind of session. The same round closes two detours around the permission system: zsh could hide commands inside regex conditionals in [[ ]], and the PowerShell check mishandled paths containing quote characters. Both now prompt for approval.
Why it matters
This is the reason to update even when nothing feels wrong. An agent working in what you believed was an isolated copy could reach your main project. A third fix stops auto-allow hooks from bypassing tool restrictions in background tasks.
Platform
Claude Code (CLI), including Windows and PowerShell, versions 2.1.221 through 2.1.222
First detected by us
August 5, 2026
Availability
Released in Claude Code 2.1.221 and 2.1.222, both August 4, 2026. The source flags neither as a security advisory or CVE.
The ultraplan feature has been taken out of Claude Code. The release note says only that it was removed, without a reason and without pointing at anything that takes over.
Why it matters
A setup or a habit built on ultraplan stops working after the update. Since the source names no replacement, it is worth checking for yourself what now covers the need.
Platform
Claude Code (CLI), version 2.1.222
First detected by us
August 5, 2026
Availability
Removed in Claude Code 2.1.222, released August 4, 2026. The source names no replacement.
A background session is a Claude Code job that keeps running without you watching it. Such sessions now commit and push to preserve the work, open a draft pull request only when the task calls for one, follow the git instructions in the project's CLAUDE.md, and always finish by reporting where the work lives.
Why it matters
The result of a long run ends up somewhere you can find it again, rather than staying inside a session you have to track down. Note that this means a background session now pushes on its own, so a project's git rules should be written down before you start one.
Platform
Claude Code (CLI), version 2.1.221
First detected by us
August 5, 2026
Availability
In Claude Code 2.1.221, released August 4, 2026. Subscription requirements are not specified.
A sandbox is a fenced-off area where commands run without full access to the machine. The new mode: "mask" setting lets a sandboxed command read a credentials file while seeing only a stand-in value. The real value is substituted by the sandbox proxy on the way out. You can mask the whole file, or only the spans a regular expression picks out.
Why it matters
A tool that needs an API key to work can run without that key sitting readable to the code executing it. On macOS the masking does not exist, and the setting becomes a plain refusal instead.
Platform
Claude Code (CLI) on Linux and WSL. On macOS the setting falls back to deny.
First detected by us
August 5, 2026
Availability
In Claude Code 2.1.221, released August 4, 2026. Subscription requirements are not specified.
Focus view is a new mode in the VS Code extension that hides tool activity behind a per-turn summary you can expand when you want the detail. An indicator shows which tool is running right now. Toggle it with Ctrl+Alt+F or the "Claude Code: Toggle Focus view" command.
Why it matters
The conversation stays readable when you are following what Claude is thinking, instead of drowning in lists of files being read. The detail is not gone, just folded away until you ask for it.
Platform
The VS Code extension
First detected by us
August 5, 2026
Availability
In Claude Code 2.1.221, released August 4, 2026. Subscription requirements are not specified.
Dreams is a feature under trial in Anthropic's service for agents hosted on their side. It can now run on Claude Opus 5, the company's newest flagship model. The release note says no more than this, and points to the documentation for the limits that apply.
Why it matters
Anyone trying Dreams gets the strongest model available inside it. That it is still a research preview means the behavior can change without notice.
Model Context Protocol (MCP) is an open standard for connecting AI models to data and tools that live outside the model. Its fifth release, named MCP 2026-07-28, moves from a two-way connection held open for the whole session to plain request-and-response calls. That means an MCP server can run on serverless infrastructure, meaning computing power that starts on demand and is billed per call. Authorization now follows ordinary OAuth 2.0 and OIDC setups, so servers plug into an existing corporate identity system such as Entra or Okta without workarounds.
Why it matters
Running an MCP server gets considerably cheaper and simpler, and companies avoid custom detours for controlling who gets access. Anthropic says MCP has passed 400 million monthly SDK downloads, and that Claude now lists over 950 MCP servers. Support is still rolling out across the Claude products, so not everything is in place everywhere yet.
Platform
Model Context Protocol (an open standard), with support rolling out across the Claude products
First detected by us
July 30, 2026
Availability
The spec went live on July 28, 2026. Support in the Claude products is still rolling out, and no completion date is given.
Claude Opus 5 is Anthropic's new flagship model, described as a step change over Claude Opus 4.8. The price is unchanged at $5 per million input tokens and $25 per million output tokens, a token being the small chunk of text a language model counts in. The model has a 1 million token context window, which is how much text it can hold in mind at once, and thinking is now on by default, so it reasons before it answers. Anthropic reports leading scores on several coding and agent benchmarks, while stating itself that the model still trails Mythos 5 on cybersecurity work.
Why it matters
Because the price did not change, existing setups can move across without a cost shift, and the migration is essentially swapping the model name from claude-opus-4-8 to claude-opus-5. Two things are worth checking afterwards: answers are longer by default, and inherited instructions to verify its own work should be removed, since the model now does that unprompted. Claude Opus 4.8 remains available.
Platform
The Claude apps, Claude Code, Claude Cowork, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry
First detected by us
July 30, 2026
Availability
Available on all platforms on launch day. The new default model on Claude Max and the strongest model on Claude Pro. The API model ID is claude-opus-5. No waitlist or staged rollout is mentioned.
Claude Code is Anthropic's command line tool, where Claude works directly in a code project from the terminal. Version 2.1.219 adds Claude Opus 5 as the new default Opus model, with a 1 million token context. At the same time Opus 4.7 is dropped from fast mode, the quicker run mode, which now covers Opus 5 and Opus 4.8. The /model picker was tidied up as well, so the merged row reads "Opus (1M context)" rather than just "Opus".
Why it matters
Claude Code users get the newest model without doing anything, and the bundled claude-api skill now points at Claude Opus 5 with a stated path from Opus 4.8. The source only quotes the fast mode price of $10 and $50 per million tokens, and says neither what the ordinary price is nor which of the two figures covers input and which covers output.
Platform
Claude Code (CLI), version 2.1.219
First detected by us
July 30, 2026
Availability
Added in Claude Code 2.1.219, released July 24, 2026. Plan and subscription requirements are not specified. The quoted price of $10 and $50 per million tokens applies explicitly to fast mode.
Thinking means the model works through a task in a separate reasoning step before answering, and effort is the dial that sets how much work goes into it. On Claude Opus 5, thinking can only be turned off at effort high or below. Disabled thinking combined with effort xhigh or max returns a 400 error, meaning the request is rejected outright. That is a breaking change from Claude Opus 4.8, where thinking could be disabled at any effort level.
Why it matters
Integrations that ran with thinking off at a high effort level stop working, and must either lower effort to high or drop the thinking field. Anthropic recommends keeping thinking on and controlling cost through lower effort instead, partly because with thinking off the model sometimes writes a tool call as ordinary text rather than as a proper tool call.
Platform
Claude Opus 5 (no platform list is specified for this change)
First detected by us
July 30, 2026
Availability
Generally available behavior from Claude Opus 5 onward, enforced on every request. No beta header, no staged rollout.
Effort is the setting that decides how much work Claude puts into an answer, and Claude Opus 5 supports the whole ladder: low, medium, high, xhigh and max. The default is high, and max is the top rung, meant for work that needs the deepest possible reasoning. Anthropic says the model turns extra effort into better results more reliably than earlier Opus models, so the choice carries more weight than before. The setting goes in output_config, for example output_config: { effort: "max" }.
Why it matters
Developers can dial effort down where quality holds up, saving tokens and latency, and dial it up for the hardest work. At xhigh and max, set a generous max_tokens, since the thinking draws on the same budget as the answer itself.
Platform
Claude Opus 5 (no platform list is specified for this change)
Fast mode is a run mode where the model answers more quickly for a price premium. For Claude Opus 5 it is a research preview, meaning an early version still being tested, and it exists only on the Claude API. The price is $10 per million input tokens and $50 per million output tokens, double the model's ordinary rate. In the launch post, Anthropic says fast mode runs at around 2.5 times the default speed.
Why it matters
Teams that need shorter response times can pay for them, but only when calling the Claude API directly. Running Claude through Amazon Bedrock, Google Cloud or Microsoft Foundry, the option is not there for now.
Platform
Claude API only. Explicitly not Amazon Bedrock, Google Cloud or Microsoft Foundry.
First detected by us
July 30, 2026
Availability
Research preview. The source page carries no date of its own; the date comes from the release note published the same day, which points to it for what is new in Claude Opus 5.
Prompt caching means the API stores a piece of text that gets sent in over and over, so it does not have to be processed afresh each time. That makes repeated calls both faster and cheaper. On Claude Opus 5 the minimum cacheable length drops to 512 tokens, down from 1,024 tokens on Claude Opus 4.8.
Why it matters
Short instructions that were too small to cache before can now create cache entries without a developer changing a single line of code.
Platform
Claude Opus 5 (no platform list is given for this change)
First detected by us
July 30, 2026
Availability
Not specified. The source page carries no date of its own; the date comes from the release note published the same day.
Fast mode, the quicker run mode, is gone for Claude Opus 4.7. Requests asking for speed: "fast" on that model now come back as an error. Unlike what happened when the same change hit Claude Opus 4.6, they do not fall back to standard speed. The model itself is still available at standard speed.
Why it matters
A setup still asking for fast mode on Claude Opus 4.7 breaks with an error rather than simply running a little slower. Anthropic recommends moving to Claude Opus 5 or Claude Opus 4.8 to keep fast mode.
Platform
claude-opus-4-7 (no platform list is given for this change)
First detected by us
July 30, 2026
Availability
Removed. Requests to claude-opus-4-7 with speed: "fast" now return an error. The removal was announced on June 25, 2026 for July 24, 2026, and that date held.
A tool is a function the developer allows the model to use, such as a search or a database lookup. Until now the full tool list had to be resent unchanged for the life of a session, or the cached context was thrown away. Tools can now be added or removed between turns while the prompt cache survives. The feature is in beta and has to be switched on with its own request header.
Why it matters
An agent can be offered fewer choices at a time, which often improves accuracy, without paying to rebuild its context. Beta status means the behavior can still change.
Platform
Claude Fable 5, Claude Mythos 5, Claude Opus 4.8 and Claude Opus 5
First detected by us
July 30, 2026
Availability
Beta. Requires the beta header mid-conversation-tool-changes-2026-07-01 on the request. No date for general availability is given.
A fallback is a backup model that takes over when a request is stopped by the safety classifiers. Until now the developer had to list which models to use. A new "default" mode applies Anthropic's own recommended fallback models, chosen by the reason the request was stopped. In the Claude apps, Claude Code and Claude Cowork this already happens automatically, with Claude Opus 4.8 as the backup.
Why it matters
With automatic fallbacks on, a request is routed to the best available model instead of being blocked. The trade is that an answer may come from a different model than the one you asked for, so plan for that in logging and testing.
Platform
Claude API
First detected by us
July 30, 2026
Availability
Beta. The "default" mode requires the beta header server-side-fallback-2026-07-01. The older server-side-fallback-2026-06-01 header accepts explicit model lists only. The entire fallbacks parameter is in beta.
A subagent is a helper Claude that the main session assigns a narrow piece of work. Version 2.1.217 turned off the ability for subagents to spawn subagents of their own. Version 2.1.219 turned it back on, three levels deep by default, and it can be disabled with CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1. There is also a new cap of 20 concurrently running subagents, and the --max-budget-usd spend limit now halts background subagents too.
Why it matters
Big jobs can be broken into several layers again, while the cap and the spend limit stop one message from fanning out into an unbounded pile of agents. The direction reversed twice inside four days, so it is worth checking which version you are actually running.
Platform
Claude Code (CLI), versions 2.1.216 through 2.1.219
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.217 (July 21) and 2.1.219 (July 24, 2026). Subscription requirements are not specified.
A dynamic workflow is Claude Code assembling a team of agents on its own to solve a task. The default is now a medium size guideline, aiming for fewer than 15 agents. The guideline is advisory rather than a hard limit, and can be changed or lifted under "Dynamic workflow size" in /config. A new workflowSizeGuideline settings key lets the choice be set from any settings file, in which case the /config row is hidden.
Why it matters
The default keeps resource use down without the user having to think about it, and the current size shows in the status line while a workflow runs. Anyone who needs larger runs can pick another size or remove the restriction.
Platform
Claude Code (CLI), version 2.1.219
First detected by us
July 30, 2026
Availability
In Claude Code 2.1.219, released July 24, 2026. Subscription requirements are not specified.
A sandbox is a fenced-off area where commands run without full access to the machine and the network. The sandbox.network.strictAllowlist setting denies hosts that are not on the approved list, without prompting the user first. The sandbox.filesystem.disabled setting enables the opposite choice: skip filesystem isolation but keep control over outbound network traffic.
Why it matters
Teams that want strict network rules stop getting repeated prompts, and teams that need open file access do not lose network control. Turning off filesystem isolation lowers the protection, so that choice should be a deliberate one.
Platform
Claude Code (CLI), versions 2.1.216 through 2.1.219
First detected by us
July 30, 2026
Availability
sandbox.filesystem.disabled in 2.1.216 (July 20, 2026), sandbox.network.strictAllowlist in 2.1.219 (July 24, 2026). Subscription requirements are not specified.
A hook is a small script that runs automatically when something specific happens in Claude Code. The new DirectoryAdded hook fires when a new working directory is registered part-way through a session. Two things can trigger it: the /add-dir command, and the register_repo_root control request from the Agent SDK.
Why it matters
Setup that has to happen per directory, such as loading rules or checking permissions, can now run automatically instead of being remembered by hand.
Platform
Claude Code (CLI) and the Agent SDK, version 2.1.219
First detected by us
July 30, 2026
Availability
In Claude Code 2.1.219, released July 24, 2026. Subscription requirements are not specified.
MCP servers are the external connections that give Claude access to data and tools. When one fails to connect, claude mcp list and /mcp now show the HTTP status and the error text instead of merely reporting failure. Claude Code also warns about config values with hidden leading or trailing whitespace, a classic cause of trouble that is invisible to the eye. In headless mode, the init event lists skipped configs under mcp_server_errors.
Why it matters
Debugging a connection that will not come up moves from guesswork to reading an actual cause. A fix in the same period also removes a trap where re-authenticating to an MCP server could revoke working credentials before the new sign-in succeeded.
Platform
Claude Code (CLI, including headless and stream-json), versions 2.1.216 through 2.1.219
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.218 (July 22) and 2.1.219 (July 24, 2026). Subscription requirements are not specified.
Four consecutive releases improve how Claude Code behaves for people using a screen reader, meaning software that reads the screen aloud. Word and line deletions are now announced, and screen reader mode echoes only the character typed instead of rewriting the whole input line on every keystroke. Plugin and settings panels move the terminal cursor to the focused row, so screen readers and magnifiers follow along during arrow-key navigation. In the VS Code extension, right-to-left text such as Arabic, Hebrew and Persian no longer renders in the wrong order when mixed with English or code.
Why it matters
The tool becomes usable for more people. For a screen reader user, the difference between hearing each character and hearing the entire line repeated is the difference between a steady workflow and an unreadable stream of audio.
Platform
Claude Code (CLI, --ax-screen-reader) and the VS Code extension, versions 2.1.216 through 2.1.219
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.217 (July 21), 2.1.218 (July 22) and 2.1.219 (July 24, 2026). Subscription requirements are not specified.
Headless mode is Claude Code run as a command inside a script rather than as a conversation in the terminal. Previously the whole answer could vanish if the connection to the API broke part-way through, even when the text had already been written. What was produced is now kept. The same round removes spurious messages claiming the user interrupted the request, and a state where an aborted tool call left half an entry in the transcript.
Why it matters
Automated runs no longer lose a finished result because of an error that arrived at the very end. It also makes the transcript more trustworthy when something has to be debugged after the fact.
Platform
Claude Code (headless claude -p, stream-json and the Agent SDK), versions 2.1.218 through 2.1.219
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.218 (July 22) and 2.1.219 (July 24, 2026). Subscription requirements are not specified.
Windows paths where a folder starts with the letter u, such as C:\Users\unicorn, were read as character codes and turned into Chinese, Japanese or Korean characters. That made the files unreachable for Claude. It is now fixed. The same round fixes auto-updates that could leave a machine without claude.exe, and CLAUDE_CODE_GIT_BASH_PATH pointing at something that is not bash is now ignored with a warning instead of exiting.
Why it matters
Windows users whose account name starts with u are spared a fault that was both confusing and hard to trace back to its cause. A failed update now also restores the preserved executable automatically, rather than leaving a machine with no Claude Code at all.
Platform
Claude Code (CLI) on Windows, versions 2.1.217 through 2.1.219
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.217 (July 21), 2.1.218 (July 22) and 2.1.219 (July 24, 2026). Subscription requirements are not specified.
Voice mode means talking to Claude instead of typing. Until now it ran only on Claude Haiku, the fastest and simplest model, which kept conversations quick but not always deep. It can now also run on Claude Opus and Claude Sonnet, and the model can be switched mid-conversation. Claude can additionally reach tools you have connected, such as Gmail and Slack, and asks permission before using one.
Why it matters
You can talk your way through a hard problem instead of typing it, and ask Claude to do practical things like moving a meeting or summarizing the day's email. The language list has grown to eleven, and Claude does not detect your language on its own: you have to ask out loud or pick the language in the voice mode settings.
Platform
The Claude apps on mobile (iOS and Android), desktop and web
First detected by us
July 30, 2026
Availability
Beta for all chat users, though it works best from a phone. The Free plan gets Claude Haiku, one connected tool and every available language. Paid plans get more models and all connected tools. Voice conversations count against normal usage limits.
The Anthropic Economic Index is the company's ongoing measurement of how Claude is actually used at work. A new connector, meaning a link Claude can pull data through, gives the model direct access to the numbers behind the index. Answers are then grounded in the real data rather than the model's memory. It is switched on from the connectors menu in claude.ai.
Why it matters
Questions about automation trends by occupation, AI adoption by region or task-specific usage patterns can be answered with a source, and Claude points to where the data comes from along the way. Anthropic notes itself that the index reflects patterns in Claude usage, not the labor market as a whole.
Platform
claude.ai (the connectors directory)
First detected by us
July 30, 2026
Availability
Available in claude.ai today. No plan restriction is stated. It works in any conversation with any Claude model, and there is nothing to install.
Claude Managed Agents is Anthropic's service for running agents hosted on their side rather than on your own server. The effort level, meaning how much work the model puts into each answer, can now be set in the agent's model configuration. The value is passed as effort inside the model object when the agent is created.
Why it matters
Effort can be decided once per agent instead of on every request. An agent doing simple work can be kept cheap, and one doing heavy work can be set high.
A webhook is a message the service sends to your system when something happens, instead of your system having to ask again and again. Webhooks for Claude Managed Agents now cover the environment and memory store lifecycle too, with four event types starting with environment. and three starting with memory_store.. The release note does not list the exact names and points to the documentation instead.
Why it matters
Your system can react when an environment or a memory store changes state, without polling for it. That saves both calls and delay.
When a session is created in Claude Managed Agents, it can now be seeded with events straight away. The field is called initial_events, goes on POST /v1/sessions, and takes up to 50 events of the types user.message and user.define_outcome. If the list is not empty, the agent loop starts in that same call.
Why it matters
Getting started drops from two calls to one, since there is no separate send-events request needed to put the agent to work.
Updating an agent in Claude Managed Agents used to require the version field. It is now optional. Supply it and it acts as a guard against two changes overwriting each other, where a mismatch returns a 409 error. Leave it out and the update is applied unconditionally.
Why it matters
Simple updates get simpler to write, while anyone who needs the guard against concurrent changes can still use it.
An event delta is a small update sent continuously while the model writes, instead of the whole answer arriving in one piece at the end. This already existed at session level in Claude Managed Agents, and now covers the stream from a single thread as well. The endpoint GET /v1/sessions/{session_id}/threads/{thread_id}/stream accepts the same event_deltas[] query parameter as the session-level stream. Each connection previews only the thread it is reading.
Why it matters
An interface can show what a subagent is doing as it happens, rather than sitting and waiting for a finished result.
/code-review is the command that asks Claude to read through your changes and look for problems. It now runs as a background subagent, so the review work no longer fills up your conversation. Several /ultrareview bugs were fixed too: descriptive arguments like "review my auth changes" now work, and get attached as a note to the findings instead of stopping the command.
Why it matters
The conversation stays readable while the review runs, and error messages now say what is actually wrong, such as how large the diff is when it is too large. In non-interactive sessions, /code-review ultra now launches the cloud review instead of quietly running a local one.
Platform
Claude Code (CLI), versions 2.1.216 through 2.1.218
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20) and 2.1.218 (July 22, 2026). /code-review ultra is described as a cloud review, but what it takes to use it is not specified.
/deep-research puts Claude Code to work on a thorough, time-consuming research job. The command now starts only when the user invokes it. Claude no longer launches it on its own.
Why it matters
An expensive, slow operation no longer begins unexpectedly in the middle of something else. The user keeps control over when time gets spent on deep research.
Platform
Claude Code (CLI), version 2.1.218
First detected by us
July 30, 2026
Availability
In Claude Code 2.1.218, released July 22, 2026. Subscription requirements are not specified.
Auto mode is the setting where Claude Code decides for itself whether a command is safe enough to run without asking. The checks for dangerous delete commands, backgrounding with & and suspicious Windows paths no longer open a permission dialog. The classifier, meaning the model that judges the risk, adjudicates them instead. Plan mode with auto also stops prompting for Bash commands the static analyzer cannot prove are read-only.
Why it matters
Fewer interruptions for commands that turn out to be safe. In exchange, more of the judgment rests on the classifier, and a fix in the same period removes a bug where an expired sign-in made auto mode deny commands outright.
Platform
Claude Code (CLI), versions 2.1.216 through 2.1.218
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20) and 2.1.218 (July 22, 2026). Subscription requirements are not specified.
A skill is a package of instructions Claude Code can pull in for a particular kind of task. Skills with context: fork now run in the background by default, and can be opted out per skill with background: false. Frontmatter, meaning the configuration fields at the top of the file, now accepts yes, no, on, off, 1 and 0 alongside true and false. Agent files reject names containing a colon, since that character is reserved for plugin namespacing.
Why it matters
Skills and commands edited mid-session now show up in the menu immediately, without a restart. A fix in the same period also removes a bug where a config with many brace groups could eat all the memory and stall the tool at startup.
Platform
Claude Code (CLI), versions 2.1.216 through 2.1.218
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.217 (July 21) and 2.1.218 (July 22, 2026). Subscription requirements are not specified.
A batch of fixes closes gaps where work could escape the area it was meant to stay inside. Subagents that were supposed to be confined to their own working copy could point git back at the shared checkout. Workflow saves could follow a symbolic link at .claude, meaning a filesystem shortcut, and write files outside the project. On Windows, read-only commands could reach network paths without a permission prompt, and hooks in agent files could run from folders the user had never approved.
Why it matters
The boundaries around a project hold better, both for what can be read and for where things can be written. Trust dialogs now also name the repository root the grant covers, so it is easier to know what you are agreeing to.
Platform
Claude Code (CLI), including Windows and PowerShell, versions 2.1.216 through 2.1.218
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.217 (July 21) and 2.1.218 (July 22, 2026). The source flags none of these as a security advisory or CVE.
The context window is how much of a conversation Claude can hold in mind at once, and compaction (/compact) is condensing it when it gets too long. The /context command now shows an explicit warning once the conversation exceeds the window, and a failed /compact displays as an error instead of looking like it worked. Auto-compaction, which never triggered for Claude Opus 4.8 on Amazon Bedrock, is fixed as well.
Why it matters
Users find out that a conversation has grown too large, rather than running into unexplained errors. A retry loop that re-sent the same doomed request over and over is gone too.
Platform
Claude Code (CLI), including Amazon Bedrock, versions 2.1.216 through 2.1.218
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.217 (July 21) and 2.1.218 (July 22, 2026). Subscription requirements are not specified.
The left arrow key could discard an entire conversation with no undo. Claude Code now asks for confirmation when the key is pressed right after an edit, and Esc in the agent view returns you to the conversation it backgrounded. Claude Code also warns when transcript writes are failing, for instance because the disk is full, rather than losing them silently. Further fixes stop a resumed session from failing every turn or crashing because of a malformed entry in its history.
Why it matters
A mistyped key no longer costs the whole session, and storage problems become visible while they can still be fixed.
Platform
Claude Code (CLI), including headless and SDK sessions, versions 2.1.217 through 2.1.218
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.217 (July 21) and 2.1.218 (July 22, 2026). Subscription requirements are not specified.
A batch of fixes targets companies that manage Claude Code centrally. Server-managed settings no longer trigger an approval prompt for benign feature and cost toggles. Spend metering now prices Bedrock inference profiles and other config-mapped model IDs at the configured model's rates. Corporate settings for mTLS, TLS verification, OAuth and proxies were being ignored in Claude Desktop sessions, and that is fixed.
Why it matters
Centrally managed setups behave more predictably, and the spend figures line up better with what is actually being used. Telemetry now goes where the managed setting says, instead of a narrower setting redirecting it somewhere else.
Platform
Claude Code (CLI), Amazon Bedrock and Claude Desktop, versions 2.1.216 through 2.1.219
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20), 2.1.217 (July 21), 2.1.218 (July 22) and 2.1.219 (July 24, 2026). Subscription requirements are not specified.
The Claude Code prompt input now understands emoji shortcodes. Type :heart: and ❤️ is inserted; type :hea and you get suggestions. The feature can be turned off with the emojiCompletionEnabled setting. The source does not say what the default value is.
Why it matters
A small convenience for anyone writing commit messages or notes straight in the terminal, and it can be switched off by anyone who would rather do without.
Platform
Claude Code (CLI), version 2.1.217
First detected by us
July 30, 2026
Availability
In Claude Code 2.1.217, released July 21, 2026. Subscription requirements are not specified.
In long sessions, the cost of normalizing messages grew quadratically with the number of turns, meaning it quadrupled every time the count doubled. The result was multi-second stalls and slow resumes. That is now fixed. A memory leak, where truncated MCP tool output was still held in memory at full length for the rest of the session, has been plugged as well.
Why it matters
Sessions that run for a long time become noticeably more pleasant to work in, rather than getting slower and slower as the day goes on.
Platform
Claude Code (CLI), versions 2.1.216 through 2.1.217
First detected by us
July 30, 2026
Availability
Released in Claude Code 2.1.216 (July 20) and 2.1.217 (July 21, 2026). Subscription requirements are not specified.
This log starts on Jul 20, 2026. Anything older is deliberately left out, so it is not the vendor's full history.
