Anthropic Prompting Best Practices

Chapter 1: Introduction & Model Overview

Overview of the Claude Model Family

Claude is a family of state-of-the-art large language models developed by Anthropic. All current Claude models support text and image input, text output, multilingual capabilities, and vision. Models are available via the Claude API, AWS Bedrock, and Google Vertex AI.

The current generation consists of five models, each optimized for different use cases:

Claude Opus 5

The most capable model, built for complex agentic coding and enterprise work, with particular strengths in long-horizon agentic tasks. Opus 5 completes full tasks rather than leaving stubs or placeholders, and it performs best when given the complete task specification up front and left to run. Thinking is on by default, and it ships with a 1M token context window as both the default and the maximum.

• API ID: claude-opus-5
• Context Window: 1M tokens (default and maximum)
• Max Output: TBD
• Pricing: TBD
• Latency: Moderate
• Thinking: On by default; can be disabled only at effort high or below
• Default Effort: high

Claude Opus 4.7

The most capable model, with particular strengths in long-horizon agentic work, knowledge work, vision, and memory. Opus 4.7 builds on the strengths of previous Opus models with improved reasoning depth and state tracking across extended sessions. It supports effort levels up to max and introduces the new xhigh effort level.

• API ID: claude-opus-4-7
• Context Window: 200K tokens
• Max Output: 128K tokens
• Latency: Moderate

Claude Opus 4.6

A highly intelligent model for building agents and coding. Opus 4.6 excels at long-horizon reasoning, complex multi-step tasks, and agentic workflows. It features adaptive thinking, 128K max output tokens, and supports effort levels up to max.

• API ID: claude-opus-4-6
• Context Window: 200K tokens (1M tokens in beta)
• Max Output: 128K tokens
• Pricing: $5/input MTok, $25/output MTok
• Latency: Moderate
• Knowledge Cutoff: May 2025 (reliable), Aug 2025 (training data)

Claude Sonnet 4.6

The best combination of speed and intelligence. Sonnet 4.6 is ideal for everyday coding, analysis, content tasks, and agentic workflows where fast turnaround matters.

• API ID: claude-sonnet-4-6
• Context Window: 200K tokens (1M tokens in beta)
• Max Output: 64K tokens
• Pricing: $3/input MTok, $15/output MTok
• Latency: Fast
• Knowledge Cutoff: Aug 2025 (reliable), Jan 2026 (training data)

Claude Haiku 4.5

The fastest model with near-frontier intelligence. Best for high-volume, latency-sensitive workloads.

• API ID: claude-haiku-4-5-20251001 (alias: claude-haiku-4-5)
• Context Window: 200K tokens
• Max Output: 64K tokens
• Pricing: $1/input MTok, $5/output MTok
• Latency: Fastest
• Knowledge Cutoff: Feb 2025 (reliable), Jul 2025 (training data)

Model Comparison Table

Feature	Opus 5	Opus 4.7	Opus 4.6	Sonnet 4.6	Haiku 4.5
API ID	claude-opus-5	claude-opus-4-7	claude-opus-4-6	claude-sonnet-4-6	claude-haiku-4-5-20251001
Pricing (input/output)	TBD	TBD	$5/$25 per MTok	$3/$15 per MTok	$1/$5 per MTok
Extended Thinking	Yes (on by default)	Yes	Yes	Yes	Yes
Adaptive Thinking	Yes	Yes	Yes	Yes	No
Context Window	1M (default and max)	200K	200K (1M beta)	200K (1M beta)	200K
Max Output	TBD	128K tokens	128K tokens	64K tokens	64K tokens
Latency	Moderate	Moderate	Moderate	Fast	Fastest
Default Effort	high	—	—	—	—
Effort: xhigh	Yes	Yes	No	No	No
Thinking can be disabled	Only at effort high or below	Yes	Yes	Yes	Yes

Platform Availability

All models are available on:

• Claude API (direct)
• AWS Bedrock (IDs: anthropic.claude-opus-4-6-v1, anthropic.claude-sonnet-4-6, anthropic.claude-haiku-4-5-20251001-v1:0)
• GCP Vertex AI (IDs: claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5@20251001)

Starting with Claude Sonnet 4.5 and later, AWS Bedrock and Google Vertex AI offer global endpoints (dynamic routing for maximum availability) and regional endpoints (guaranteed data routing through specific geographic regions).

How to Choose the Right Model

Choose Opus 5 when:

• You need the absolute most capable model available
• Difficult agentic coding: multi-file features, larger refactors, end-to-end feature work
• Code review and bug-finding, where you want high precision and recall in a single pass
• Long-context work — instruction following, tool calling, and reasoning stay consistent across the full 1M token window
• Multi-agent workloads that coordinate teams of subagents
• Vision-heavy work: charts, documents, diagrams, and UI/frontend visual replication
• Office and document tasks: multi-sheet spreadsheets with non-trivial formulas, structured slide decks
• You want strong quality at low token cost — low and medium effort hold up well on Opus 5

Choose Opus 4.7 when:

• You need the absolute most capable model available
• Long-horizon agentic work requiring deep memory and state tracking
• Complex knowledge work, research, and analysis
• Vision-heavy tasks requiring strong image understanding
• You need the xhigh or max effort levels for maximum capability
• Note: effort is more important for Opus 4.7 than any prior model

Choose Opus 4.6 when:

• You need very high intelligence for complex reasoning tasks
• Building autonomous agents that run for extended periods
• Working on large-scale code migrations or deep research
• You need long-horizon reasoning across multiple context windows
• You need the max effort level for absolute maximum capability
• Cost is less important than quality

Choose Sonnet 4.6 when:

• You need a balance of speed and intelligence
• Building agentic coding workflows with fast turnaround
• Working on everyday coding, analysis, and content tasks
• Cost efficiency matters alongside strong performance
• You need adaptive thinking with faster response times

Choose Haiku 4.5 when:

• Speed is the top priority
• Running high-volume, latency-sensitive workloads
• Simple classification, routing, or subagent tasks
• Cost optimization is critical
• Tasks don't require deep reasoning

Decision Framework

1. Start with the task complexity: If it requires deep multi-step reasoning or long-horizon agentic work, start with Opus 5
2. Consider latency requirements: If you need fast responses, use Sonnet 4.6 or Haiku 4.5
3. Factor in volume: High-volume tasks favor Haiku 4.5 for cost efficiency
4. Evaluate with the effort parameter: You can often get Opus-level quality from Sonnet at lower cost by adjusting effort levels
5. Test and iterate: The best model for your use case depends on your specific requirements - prototype with Opus, then see if Sonnet or Haiku can meet your quality bar