Claude Code v2.1.154 Major Updates - Opus 4.8 is here!
Claude Code v2.1.154 has been released. The main highlights are the introduction of the new top-tier model Opus 4.8 and the introduction of dynamic workflows that coordinate multiple agents. There are also important security-related fixes.
2026.05.29
This page has been translated by machine translation. View original
This is Ishikawa from the Cloud Business Division. Claude Code v2.1.154 (released May 28, 2026) has been released. It's a major update with 44 total changes, centered around the introduction of the new top-tier model Claude Opus 4.8 and the introduction of "dynamic workflows" that coordinate multiple agents in the background.
Update Summary
v2.1.154 includes 44 changes: 7 new features, 5 improvements, 3 security fixes, 24 bug fixes, 1 deprecation, and more. For practical purposes, the overall picture is easiest to grasp by focusing on three points: (1) the new top-tier model Opus 4.8 and its effort settings and pricing, (2) dynamic workflows that orchestrate dozens to hundreds of agents, and (3) security-related changes such as dangerous path blocking and data exfiltration detection.
Notable Updates
- New Feature: The new top-tier model Claude Opus 4.8 has arrived (high effort by default, with
/effort xhighalso available for especially difficult tasks). Additionally, dynamic workflows have been introduced, allowing multiple agents to coordinate in the background (check execution status with/workflows). - Bug Fix: Fixed an issue where
rm -rf $HOMEwas not being blocked as a dangerous path whenHOMEhad a trailing slash, which could lead to home directory deletion. - Security: Detection of data exfiltration by the auto mode classifier, such as bulk transfers of repository contents, has been improved. The dangerous path fix mentioned above is also an important security fix.
Update Details
New Features
- Introduction of Claude Opus 4.8: A new top-tier model is now available. It operates with high effort (high thinking capacity) by default, and
/effort xhighcan also be selected for especially difficult tasks. Support for Opus 4.8 and migration guidance from 4.7 → 4.8 have also been added to the/claude-apiskill.
- Dynamic Workflows: When you ask Claude to create a workflow, it can now orchestrate dozens to hundreds of agents in the background to tackle larger and more complex tasks. Execution status can be checked with
/workflows.
- Background Execution of Shell Commands: In
claude agents, entering! <command>now allows shell commands to be run as attachable/detachable background sessions. The same is available withclaude --bg --exec '<command>'. - (Chrome) Browser Selection: In Claude in Chrome, you can now select which connected browser to use from
/chrome→ "Select browser…". When multiple browsers are connected, you can also select from within the chat. - (Plugin) Declaring Default Disabled: Plugins can now declare
defaultEnabled: falseinplugin.jsonor in marketplace entries. They can be enabled with/pluginorclaude plugin enable, and dependencies of enabled plugins continue to be automatically enabled. - (Plugin) Directory-Based Recommendations: In the Discover tab of
/plugin, plugins highly relevant to the current directory are now pinned with a "suggested for this directory" annotation. - (MCP) Passing Environment Variables to Subprocesses: The environment variables
CLAUDE_CODE_SESSION_IDandCLAUDECODE=1are now passed to subprocesses of stdio-based MCP servers.
Improvements
- Opus 4.8 Fast Mode at Lower Cost: Opus 4.8 Fast Mode is now available at a fraction of the previous cost. According to the official description, it offers 2.5x speed at 2x the standard rate.
- Lean System Prompt Now Default: The lean (lightweight) system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier.
- Suppression of Unnecessary Questions: When Claude has sufficient context to proceed on its own, it will no longer present multiple-choice questions, reserving them only for decisions it truly cannot make independently.
- Clarified Role of
/simplify:/simplifynow runs a cleanup-focused review (reuse, simplification, efficiency, abstraction) and applies fixes, rather than a full review with bug hunting like/code-review --fix. - Streaming Tool Execution Always Enabled: Streaming tool execution is now always enabled (including when telemetry is disabled and for Bedrock/Vertex/Foundry). Previously it was behind a feature flag.
- (VS Code) Improved Auto Mode Display: Auto mode now appears in the mode picker without requiring bypass-permissions configuration, and on the new session screen, a dismissible notification explaining auto mode is shown the first time it becomes available.
- Additionally, minor operational improvements have been made, including changing the
/effortslider labels from "Speed"/"Intelligence" to "Faster"/"Smarter", and having/remote-controlautocomplete display "Disconnect Remote Control" while connected.
Security
- Improved Data Exfiltration Detection: Detection of data exfiltration by the auto mode classifier has been improved, particularly for bulk transfers of repository contents.
- Fixed Dangerous Path Blocking Gap: Fixed an issue where
rm -rf $HOMEwas not being blocked as a dangerous path whenHOMEhad a trailing slash. This is an important practical fix as it could lead to home directory deletion. - Preventing Auto-Approval of Unapproved MCP Servers: With
claude mcp list/get, unapproved.mcp.jsonservers now display⏸ Pending approvalwhen output is piped, instead of being automatically approved and connected.
Bug Fixes
Here are the key fixes related to stability and operability.
- Fixed
$TMPDIRResolution Mismatch: Fixed an issue where$TMPDIRresolved to different directories for sandboxed Bash commands versus non-sandboxed commands within the same session. - Fixed Subagent Worktree Isolation Bypass: Fixed an issue where subagents in background sessions bypassed worktree isolation guards and wrote to shared checkouts.
- Fixed Full Policy Discard in Managed Settings: Fixed an issue where a single invalid
allowedMcpServers/deniedMcpServersentry in managed settings caused the entire managed-settings policy to be discarded. Invalid entries are now ignored with aclaude doctorwarning. - Fixed False Blocking in Auto Mode: Fixed an issue where auto mode incorrectly blocked actions as "could not evaluate this action" when the safety classifier ran out of output tokens during inference.
- Fixed Agent View Launch Compatibility: Opening the agent view with
←←(two left arrows) now works on Bedrock, Vertex, Foundry, and when telemetry is disabled. - Fixed Multiple Background Session Stability Issues: Fixed issues including premature "out of context" behavior on some 1M context models due to completion notifications, losing objectives during scheduled execution, process multiplication from per-minute regeneration after updates, failure to terminate from blocked/running/working states, and 100% CPU usage by
claude --bg-pty-hostprocesses on macOS. - Additionally, numerous minor bugs have been fixed in UI display, notifications, plan and workflow display,
/mcptool listing, error display during Windows updates, terminal rendering, and more.
Deprecation
There are no breaking changes, but there is one environment variable scheduled for removal. It is handled separately as migration is required.
- Deprecation of
CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE: This environment variable has been deprecated (scheduled for removal on June 1, 2026). To use Fast Mode with Opus 4.6, switch to/model claude-opus-4-6[1m]and then run/fast on.
Closing
v2.1.154 is an update centered around two major topics — the arrival of the top-tier model Opus 4.8 and the introduction of dynamic workflows — along with security fixes (dangerous path blocking gap, data exfiltration detection) and a deprecation requiring action (CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE deletion on June 1). If you are using that environment variable in particular, please confirm your migration before the scheduled deletion date. For environments using the relevant features, why not update early and verify the behavior?
Recent Updates
References
