Sub-agents can now spawn their own sub-agents up to 5 levels deep (Claude Code 2.1.172)
Vision documentation completely overhauled with accurate patch-based tokenization formula, coordinate handling guide, and reference Python code for computing exact resize dimensions
Claude Fable 5 and Mythos 5 added to code execution, programmatic tool calling, web search/fetch dynamic filtering, structured outputs, and tool search supported models
New advisor tool guidance: mid-conversation nudge pattern for Haiku executors and alternative system prompt optimized for coding workloads
New frontier_llm refusal category documented — blocks requests that could assist development of competing AI models
Sub-agents can now spawn their own sub-agents (up to 5 levels deep)
Amazon Bedrock now reads the AWS region from ~/.aws config files when AWS_REGION isn't set, matching AWS SDK precedence; /status shows where the region came from
Added a search bar when browsing a marketplace's plugins in /plugin
Added model attribute to the claude_code.lines_of_code.count OTEL metric
Existing feature improvements
Improved performance in long conversations by removing redundant message normalization and avoiding full message-history transforms when streaming tool-use state is unchanged
Reduced idle CPU usage: /goal status chip no longer re-renders the terminal at 5 Hz while idle, and fewer UI re-renders while subagents run in parallel
Improved Claude in Chrome tool loading: browser tools now load in a single batched call instead of one per tool
/code-review now keeps the ultra option visible when not signed in to claude.ai, with an explanation that cloud review requires a claude.ai account
Major bug fixes
Fixed sessions using 1M context without usage credits getting permanently stuck — the session now automatically compacts back under the standard context limit
Fixed background agents potentially reading another directory's project settings (.mcp.json approvals, trust) when dispatched onto a pre-warmed worker
Fixed availableModels restrictions not being applied to subagent model overrides, the agent dispatch model picker, and the advisor model
Fixed WebFetch(domain:*.example.com) wildcard domain rules never matching subdomains in allow, deny, and ask position; also fixed file permission rules with mid-pattern wildcards (e.g. Read(secrets-*/config.json)) being rejected at startup
Fixed background-session attach failing with EAUTH for sessions started on an older version after the daemon auto-updated
Fixed Fable 5 model names with a [1m] suffix not being normalized — Fable 5 includes 1M context by default, so the suffix is now stripped automatically
Fixed a spurious "sandbox dependencies missing" startup warning on Windows when sandbox was enabled in settings
Updated internal link for Fable 5 to point to the new combined model introduction page (introducing-claude-fable-5-and-claude-mythos-5.md) [line 38] [Source]
Clarified that claude-plugins-official is registered on first interactive launch; non-interactive scripts that run before that must add it explicitly with claude plugin marketplace add anthropics/claude-plugins-official [line 357] [Source]
Updated plugin submission URL for claude.ai to the new path (/admin-settings/directory/submissions/plugins/new) [line 362] [Source]
Added note that the claude.ai submission form requires a Team or Enterprise organization; individual authors without such an org can use the Console form instead [line 365] [Source]
Expanded documentation on thinking.display defaults — Claude Mythos Preview is now explicitly listed alongside Fable 5, Mythos 5, Opus 4.8, and 4.7 as defaulting to "omitted"; Claude Opus 4.6's default of "summarized" is now called out separately [lines 221-230] [Source]
Improved guidance on stripping thinking blocks when switching models — now clearly separated into main rule and two named exceptions [lines 271-279] [Source]
Added new "Mid-conversation nudge for under-calling executors" section with code example — a nudge injected before turn 2 raised Haiku task pass rates ~7 points; Sonnet showed no benefit; Opus should not receive the nudge [lines 247-310] [Source]
Added "Alternative system prompt for Haiku on coding workloads" — a specialized block that raised Haiku pass rates ~7.5 points on coding benchmarks, with a caveat that it costs ~4 points on browse/lookup workloads [lines 490-537] [Source]
Added "Increasing advisor calls on Opus executors" section with a targeted checkpoint system prompt for under-calling Opus executors [lines 539-553] [Source]
Added Claude Fable 5 and Claude Mythos 5 to the supported models table, including support for code_execution_20260120 (REPL persistence and programmatic tool calling) [lines 21-22] [Source]
Added explicit documentation of the thinking blocks cannot be modified 400 error — explains its cause (application code filtering or rebuilding the assistant message) and links to the errors reference for fix steps [line 392] [Source]
Added "Fallback credit" entry to the features table — describes the beta feature for avoiding double prompt-cache costs on refused requests [line 47] [Source]
Added "Server-side fallback" entry to the features table — describes the beta feature for retrying refused requests inside a single API call [line 49] [Source]
Added "frontier_llm" as a new refusal category value — triggered when a request could assist development of competing AI models; benign ML work can also trigger it [line 69] [Source]
Improved documentation structure for streaming vs. non-streaming fallback behavior, now with separate labeled sections [lines 214-237] [Source]
Added new error section for "thinking blocks cannot be modified" 400 error — explains the cause and links to the errors reference [lines 44-49] [Source]
Added documentation that organizations with a ZDR arrangement can enable 30-day data retention per workspace via Console, allowing Claude Fable 5 and Mythos 5 to be used in that workspace while other workspaces remain zero-retention [lines 49-51] [Source]
Replaced the approximate width * height / 750 token formula with the accurate patch-based formula: ⌈width / 28⌉ × ⌈height / 28⌉ visual tokens (one token per 28×28 pixel block) [line 45] [Source]
Updated image cost tables with exact token counts (removing approximations) and added a 4K (3840×2160) row showing that 4K and 1920×1080 images can have the same cost when both downscale to the same size [lines 64-69] [Source]
Added a major new section "Working with coordinates and bounding boxes" covering: how Claude's resize-and-pad pipeline works, a reference Python implementation of resized_size() and to_relative_coordinates(), guidance to always request absolute pixel coordinates, and separate subsections for pre-resizing before upload vs. rescaling coordinates after the fact [lines 108-230] [Source]