Version 2.0.51

What: Opus 4.5 is now available as the most capable model for complex work, replacing Opus 4.1 as the default for first-party API users.

How to use:

# Select Opus 4.5 via the model command
/model opus

# Or specify directly
claude --model opus

Details:

• Opus 4.5 becomes the default for Max and Team subscribers using first-party API
• Bedrock/Vertex/Foundry users continue to use Opus 4.1
• Model descriptions updated: "Most capable for complex work"
• One-time migration automatically resets model selection for eligible users
• A notification informs users about Opus 4.5 availability on first use
• Evidence: Model registry at rcA() line 163104, default logic at QUA() line 224121, selector at rJ6 line 224397

What: A new tool that allows Claude to proactively request permission to enter plan mode for complex tasks requiring careful exploration before implementation.

How to use:

# Claude will automatically suggest entering plan mode for complex tasks
# You'll see a prompt: "Enter plan mode?"
# Press 'y' to approve or 'n' to decline

Details:

• Requires user approval before entering plan mode
• Designed for tasks with multiple valid approaches, architectural decisions, or unclear requirements
• Includes extensive documentation on when to use vs. not use plan mode
• Cannot be used from agent contexts (main conversation only)
• Evidence: Tool definition at GRA variable in _50 at line 425698, prompt at vb2 line 425575

What: Rate limit warnings now show exact utilization percentages and use intelligent thresholds based on both token usage and time remaining.

How it appears:

75% of 5-hour limit used · resets in 2h
50% of weekly limit used · resets in 3d

Details:

• Shows percentage of limit used instead of generic "Approaching limit" messages
• Smart thresholds: warns at 75%/60%, 50%/35%, 25%/15% (tokens/time) for weekly limits
• 5-hour limits now show warnings (previously silent)
• Considers usage rate vs. time remaining in window
• Evidence: Function N49() at line 467497, threshold config at j$3/S$3 line 467531

What: When multiple Task agents run simultaneously, the UI now shows a collapsed summary view that can be expanded with ctrl+o.

How it appears:

Running 3 agents… (ctrl+o to expand)
├─ investigator (Check auth flow) · 5 tool uses · 2.3k tokens
├─ investigator (Find routes) · 3 tool uses · 1.8k tokens
└─ investigator (Review tests) · 7 tool uses · 3.1k tokens

Details:

• Reduces visual clutter when multiple agents run in parallel
• Shows agent type, description, tool use count, and token usage
• Displays "X agents finished" when complete
• Same-type agents show grouped label (e.g., "Running 3 investigator agents…")
• Evidence: Main renderer i49() at line 469450, item renderer g49() at line 469102

The extra usage feature now has deeper integration with account state and re-authentication flows.

• Added hasExtraUsageEnabled account property tracked via OAuth
• /extra-usage command now triggers re-authentication to sync state
• Pro users with extra usage enabled get access to Opus features
• Evidence: Account property at DC1() line 67554, API mapping at line 67521

Rate limit error messages now provide actionable guidance specific to your subscription tier.

• Messages suggest specific actions: /upgrade to Max, turn on /extra-usage, /model opus
• Different guidance for Pro, Max, Team, and Enterprise users
• Sonnet weekly limit now has distinct messaging from general weekly limit
• Evidence: Function YLA() at line 336478

• Sonnet 4.5: "Best for everyday tasks" (was "Smartest model for daily use")
• Opus: "Most capable for complex work"
• Haiku 4.5: "Fastest for quick answers" (was "Fastest model for simple tasks")
• Dynamic pricing display using jc() helper function
• Evidence: Model options at qh1, kzB, xzB around line 224380

Max and Team users see an informational notice about limit changes:

• Opus cap removed - can use Opus 4.5 up to overall limit
• Sonnet now has its own separate limit
• Evidence: Function TN3() at line 477797

Assistant messages ending with thinking blocks now have those blocks automatically filtered out before display.

• Prevents display of incomplete or dangling thinking blocks
• Logs telemetry about filtered blocks for debugging
• Replaces content with "[No message content]" if all blocks were thinking
• Evidence: Function GT3() at line 504394, helper oX9() at line 504391

Max and Team subscribers now default to Opus 4.5 instead of requiring manual model selection.

• First-party API users get Opus 4.5 by default
• Cloud provider (Bedrock/Vertex) users continue with Opus 4.1
• Respects ANTHROPIC_DEFAULT_OPUS_MODEL environment variable override
• Evidence: Function PnA() at line 224154, QUA() at line 224121

The model detection functions now correctly recognize Opus 4.5 model IDs.

• claude-opus-4-5 pattern now properly identified
• Fixes potential issues with model-specific features not activating for Opus 4.5
• Evidence: Detection at KAB() line 163004, Pw() line 224165