Claude Opus 4.5
The King Returns
Claude Opus 4.5 Has Arrived
Anthropic has launched Claude Opus 4.5 on November 24, 2025, reclaiming the throne as the world's best coding model and setting a new standard for agentic AI. With massive price reductions and powerful new capabilities, Opus 4.5 moves us from "AI Assistants" to "Autonomous Employees." This completes the Claude 4.5 model family, following Sonnet 4.5 (September) and Haiku 4.5 (October).
Claude Opus 4.5 is the world's best coding model (80.9% SWE-bench Verified). It's designed for deep work, not just chat. First model ever to break 80% on this respected benchmark.
With massive price reductions and new "Agentic" controls, this is the model that moves us from "AI Assistants" to "Autonomous Employees."
Frontier Intelligence, Now Affordable
Input / Output per million tokens
Anthropic slashed costs by 67%—down from $15/$75 with Opus 4.1. High IQ is now scalable.
The "Effort" ParameterNEW
Finally, you control the brainpower. Toggle between Low, Medium, or High effort to balance cost vs. depth.
At medium effort, Opus 4.5 matches Sonnet 4.5's SWE-bench score while using 76% fewer output tokens.
Thinking PersistenceNEW
It remembers how it thought. Opus 4.5 preserves "thinking blocks" across multi-step agentic workflows.
This solves the amnesia problem—agents can now work on tasks that take hours, not minutes.
Computer Use "Zoom"NEW
Agents can now "lean in." New zoom capabilities allow pixel-perfect interaction with complex UI elements.
The new "Zoom" action fixes the biggest issue with Computer Use: small text. Claude can now inspect detailed interfaces like a human would.
Performance Benchmarks
Claude Opus 4.5 sets new standards across coding, agentic workflows, and visual reasoning.
🏆 Coding King
SWE-bench Verified
First model ever to break 80%. Beats GPT-5.1-Codex-Max (77.9%) and Gemini 3 Pro. It doesn't just write code; it refactors entire codebases.
🤖 Agentic Surge
Tau2-bench Tool Use
This isn't a chatbot; it's a reliable worker that follows complex, multi-step instructions without getting lost.
👁️ Visual Zoom
OSWorld Computer Use
The new "Zoom" action fixes the biggest issue with Computer Use: small text. Claude can now inspect detailed interfaces like a human would.
Comparative Performance
| Feature | Claude Opus 4.5 | Sonnet 4.5 | GPT-5.1-Codex-Max |
|---|---|---|---|
| Agentic Coding (SWE-bench) | 80.9% ✓ | 77.2% | 77.9% |
| Agentic Tool Use (Tau2-bench) | 98.2% ✓ | 98.0% | -- |
| Visual Reasoning (MMMU) | 80.7% | 77.8% | 85.4% |
| Computer Use (OSWorld) | 66.3% ✓ | 61.4% | -- |
| Terminal-bench | -- | 50.0% | -- |
| Price (Input/Output) | $5 / $25 | $3 / $15 | $1.25 / $10 |
Anthropic gave Opus 4.5 the same take-home test they give engineering candidates. It scored higher than any human who's ever applied.
The test has a two-hour time limit and focuses purely on technical ability and judgment under pressure.
Key Features & Capabilities
Opus 4.5 introduces several breakthrough features that fundamentally change how AI can be deployed.
Effort Parameter (Beta)NEW
Control how much computational effort Claude allocates:
- Low: Fast responses for simple queries
- Medium: Balanced performance (matches Sonnet 4.5 with 76% fewer tokens)
- High: Extended thinking for complex problems (default)
Balance performance with latency and cost for your specific use cases.
Thinking Block PreservationNEW
Opus 4.5 maintains reasoning context across long-running tasks:
- Remembers its thought process between steps
- Maintains consistency across file edits
- Enables multi-hour agentic workflows
- Reduces redundant re-analysis
Enhanced Computer UseNEW
Significant improvements to computer use capabilities:
- Zoom action: Inspect small UI elements
- 66.3% success rate on OSWorld benchmark
- Better desktop automation reliability
- Human-like interaction with interfaces
Production-Grade Coding
Best-in-class software engineering:
- Complex refactoring and migrations
- Multi-system debugging
- Best practices and security patterns
- Efficient token usage (76% fewer tokens at medium effort)
Document Creation
Step-change improvement in creating professional documents:
- Spreadsheets with consistency and polish
- Presentations with domain awareness
- Documents with professional formatting
- Better memory for project context
Endless ConversationsNEW
Long conversations no longer hit a wall:
- Automatic context summarization when limits approached
- Continuous chat without interruption
- Better handling of extended projects
- Maintains conversation quality over time
Anthropic describes Opus 4.5 as "the most robustly aligned model we have released to date and, we suspect, the best-aligned frontier model by any developer."
Substantial progress against prompt injection attacks—harder to trick than any other frontier model in the industry.
Enterprise Use Cases
Real-world applications where Opus 4.5's capabilities create transformative value.
💻 Senior Engineer Replacement
Scenario: Development team needs to refactor legacy codebase and resolve deep bugs.
How Opus 4.5 Helps:
- 80.9% success rate on real-world software engineering tasks
- Complex refactoring across multi-system codebases
- Deep bug investigation without hand-holding
- Thinking persistence maintains context across long sessions
- Effort parameter allows cost control for different task types
🤖 Long-Running Agent Workflows
Scenario: Enterprise needs autonomous agents for multi-hour projects.
How Opus 4.5 Helps:
- Thinking Block Preservation solves amnesia problem
- 98.2% success on complex tool use (Tau2-bench)
- Can work on tasks that take hours, not minutes
- Maintains consistency across files and steps
- Affordable pricing makes scale deployment viable
🖥️ Computer Use Automation
Scenario: Automate desktop tasks with complex UIs and small text.
How Opus 4.5 Helps:
- New "Zoom" action for pixel-perfect interaction
- 66.3% success rate on OSWorld computer use benchmark
- Can inspect detailed interfaces like spreadsheets
- Reliable automation of desktop workflows
- Human-like interaction with complex UI elements
📊 High-Stakes Enterprise Tasks
Scenario: Financial analysis, legal research, strategic planning requiring frontier intelligence.
How Opus 4.5 Helps:
- Effort parameter lets you dial up thinking for critical tasks
- Extended reasoning for complex trade-off analysis
- Step-change improvement in creating professional documents
- Domain awareness and consistency
- Affordable pricing means you can deploy at scale
Previous Opus models were "the real SOTA" but cost-prohibitive. Opus 4.5 is now at a price point where it can be your go-to model for most tasks.
It's the clear winner and exhibits the best frontier task planning and tool calling we've seen yet. This isn't an incremental improvement—it's a fundamental shift in what's economically viable.
Technical Specifications
Model ID for API: claude-opus-4-5-20251101 or claude-opus-4-5
Context Window: 200,000 tokens
Thinking Budget: 64,000 tokens (extended thinking capability)
Max Output: 64,000 tokens per response
Knowledge Cutoff: May 2025 (use web search for current info)
Pricing
API Pricing
Per million tokens (input / output)
Additional Savings:
- Up to 90% with prompt caching
- 50% with batch processing
Pricing Comparison
Opus 4.5 represents a dramatic price reduction:
- Opus 4.1: $15 / $75 per million tokens
- Opus 4.5: $5 / $25 per million tokens
- Reduction: 67% cost decrease
Makes frontier intelligence accessible for production use at scale.
Access Methods
| Platform | Availability | Best For |
|---|---|---|
| Claude.ai Web & Apps | Pro, Max, Team, Enterprise (Default model) | Individual users and teams |
| Claude API | Available now | Developers building AI solutions |
| Amazon Bedrock | Available now | AWS customers |
| Google Cloud Vertex AI | Available now | GCP customers |
| Microsoft Azure Foundry | Available now | Azure customers |
| GitHub Copilot | Paid plans | Developers using GitHub |
Product Updates
Claude Code DesktopNEW
Available on Windows, macOS, and Windows (Arm 64):
- Run multiple coding sessions in parallel
- Upgraded plan mode with precise execution
- Clarifying questions upfront
- User-editable plan.md files
- Auto-compaction for long contexts
Claude for ChromeEXPANDED
Now available to all Max users:
- Handle tasks across browser tabs
- Automated web workflows
- Research and data gathering
- Form filling and navigation
Claude for ExcelGA
Now generally available to Max, Team, and Enterprise:
- Direct Excel integration
- Data analysis and manipulation
- Formula assistance
- Spreadsheet automation
Usage LimitsIMPROVED
Increased capacity for Opus users:
- Opus-specific caps removed
- Max and Team Premium limits increased
- Can run Opus at Sonnet-tier levels
- Extended thinking on by default
- API: Use model ID
claude-opus-4-5-20251101orclaude-opus-4-5 - Documentation: Visit docs.anthropic.com
- Web Interface: Available at claude.ai
- System Card: Full safety and evaluation details in the model card
The Claude 4.5 Model Family
Opus 4.5 completes the Claude 4.5 model family, joining Sonnet 4.5 (September 2025) and Haiku 4.5 (October 2025). Each model serves different use cases across the intelligence-speed-cost spectrum.
Frontier Intelligence
Released: November 24, 2025
Pricing: $5 / $25 per million tokens
SWE-bench: 80.9%
- World's best coding model
- Highest reasoning capability
- Complex agentic workflows
- Effort parameter for cost control
- Best for: Deep work, enterprise tasks, complex refactoring
Balanced Performance
Released: September 29, 2025
Pricing: $3 / $15 per million tokens
SWE-bench: 77.2%
- Default model for most users
- 30+ hours autonomous operation
- 61.4% OSWorld (computer use)
- Checkpoints in Claude Code
- Best for: Daily use, production workloads, agents
Speed & Efficiency
Released: October 15, 2025
Pricing: $1 / $5 per million tokens
SWE-bench: 73.3%
- Matches Sonnet 4's performance
- 2x+ faster, 1/3 the cost
- Available to all free users
- Sub-agent orchestration
- Best for: Real-time apps, chatbots, high-volume tasks
Full Model Comparison
| Feature | Opus 4.5 | Sonnet 4.5 | Haiku 4.5 |
|---|---|---|---|
| Release Date | Nov 24, 2025 | Sep 29, 2025 | Oct 15, 2025 |
| SWE-bench Verified | 80.9% | 77.2% | 73.3% |
| OSWorld (Computer Use) | 66.3% | 61.4% | ~Sonnet 4 level |
| Context Window | 200K | 200K | 200K (1M for developers) |
| Input Price (per 1M) | $5 | $3 | $1 |
| Output Price (per 1M) | $25 | $15 | $5 |
| Best For | Complex reasoning | Daily production use | Speed-critical apps |
The Claude 4.5 family is designed to work together. Opus 4.5 can break down complex problems into multi-step plans, then orchestrate a team of multiple Haiku 4.5s to complete subtasks in parallel.
This enables sophisticated workflows where the lead agent (Opus) handles strategic decisions while sub-agents (Haiku) execute tasks at high speed and low cost.
• Use Opus 4.5 for: Complex refactoring, deep bugs, high-stakes decisions, frontier reasoning
• Use Sonnet 4.5 for: Daily coding, production workloads, balanced performance
• Use Haiku 4.5 for: Real-time chat, high-volume APIs, cost-sensitive applications, sub-agents