Skip to main content
Anthropic provides the Claude family of models, known for strong reasoning, long context windows (200K tokens), and extended thinking capabilities.

​
Installation

The Anthropic client is included in the core Koog library. No additional dependencies required.

​
Quick Start

import ai.koog.prompt.executor.clients.anthropic.*
import ai.koog.agents.core.*

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5
)

val agent = AIAgent(
    executor = executor,
    tools = toolRegistry {
        // Your tools here
    }
) {
    // Define your agent strategy
}

val result = agent.execute("Analyze this document...")

​
Authentication

​
API Key Setup

export ANTHROPIC_API_KEY=sk-ant-...

​
Programmatic Configuration

val client = AnthropicLLMClient(
    apiKey = "sk-ant-...",
    settings = AnthropicClientSettings(
        baseUrl = "https://api.anthropic.com",
        apiVersion = "2023-06-01",
        timeoutConfig = ConnectionTimeoutConfig(
            requestTimeoutMillis = 120_000
        )
    )
)

​
Available Models

Best model for complex agents and coding tasks. 
AnthropicModels.Sonnet_4_5       // 200K context, highest intelligence
Capabilities:
  • Extended thinking with signatures
  • Function calling
  • Vision (images and PDFs)
  • Document processing
  • 200K token context window
  • 64K token output

​
Claude Haiku 4.5

Fastest model with near-frontier intelligence. 
AnthropicModels.Haiku_4_5        // 200K context, blazing speed
Use cases:
  • High-throughput applications
  • Real-time interactions
  • Cost-sensitive deployments
  • Quick analysis tasks

​
Claude Opus Series

Most capable models for specialized complex tasks. 
AnthropicModels.Opus_4_5         // Premium model, best for coding/agents
AnthropicModels.Opus_4_6         // Frontier model with 1M output tokens
AnthropicModels.Opus_4_1         // Specialized complex tasks
Use cases:
  • Software engineering
  • Agentic tasks
  • Long-context reasoning
  • Financial analysis
  • Multi-step research workflows

​
Legacy Models

AnthropicModels.Sonnet_4         // Previous generation
AnthropicModels.Opus_4           // Previous flagship
AnthropicModels.Haiku_3          // Deprecated, use Opus_4_6 instead

​
Code Examples

​
Basic Chat Completion

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5
)

val result = executor.execute(
    prompt = prompt {
        system("You are a helpful assistant.")
        user("Explain quantum computing in simple terms.")
    }
)

println(result.first().content)

​
Function Calling with Tools

data class SearchArgs(val query: String)

val searchTool = tool<SearchArgs, String>(
    name = "search",
    description = "Search the web for information"
) { args ->
    "Results for: ${args.query}"
}

val agent = AIAgent(
    executor = simpleAnthropicExecutor(
        apiKey = System.getenv("ANTHROPIC_API_KEY"),
        model = AnthropicModels.Sonnet_4_5
    ),
    tools = toolRegistry { tool(searchTool) }
) {
    defineGraph<String, String>("search-agent") {
        val response = callLLM()
        finish(response)
    }
}

val result = agent.execute("Find information about renewable energy")

​
Vision - Image Analysis

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5
)

val result = executor.execute(
    prompt = prompt {
        user {
            text("What's in this image?")
            image(
                url = "https://example.com/image.jpg"
                // or: bytes = imageBytes
            )
        }
    }
)

​
Document Processing

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5
)

val result = executor.execute(
    prompt = prompt {
        user {
            text("Summarize this PDF")
            file(
                bytes = pdfBytes,
                mimeType = "application/pdf",
                fileName = "document.pdf"
            )
        }
    }
)

​
Extended Thinking

Claude models support extended thinking for complex reasoning: 
val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Opus_4_6,
    params = AnthropicParams(
        thinking = AnthropicParams.Thinking(
            type = "adaptive",
            effort = "high" // low, medium, high, xhigh
        )
    )
)

val result = executor.execute(
    prompt = prompt {
        user("Solve this complex math problem: ...")
    }
)

// Access reasoning
result.forEach { message ->
    when (message) {
        is Message.Reasoning -> println("Reasoning: ${message.content}")
        is Message.Assistant -> println("Answer: ${message.content}")
    }
}

​
Streaming Responses

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5
)

executor.executeStreaming(
    prompt = prompt { user("Write a detailed analysis of...") }
).collect { frame ->
    when (frame) {
        is StreamFrame.TextDelta -> print(frame.text)
        is StreamFrame.ReasoningDelta -> print("[Thinking: ${frame.text}]")
        is StreamFrame.End -> println("\nComplete")
        else -> {}
    }
}

​
Long Context Processing

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5 // 200K context
)

val longDocument = File("large_document.txt").readText() // Can be very large

val result = executor.execute(
    prompt = prompt {
        system("You are analyzing a long document.")
        user("Document content: $longDocument\n\nQuestion: What are the main themes?")
    }
)

​
Advanced Configuration

​
Custom Parameters

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5,
    params = AnthropicParams(
        temperature = 0.7,
        maxTokens = 4096,
        topP = 0.9,
        topK = 40,
        stopSequences = listOf("\n\nHuman:"),
        thinking = AnthropicParams.Thinking(
            type = "adaptive",
            effort = "medium"
        )
    )
)

​
Model Version Mapping

Claude uses version-specific model IDs. Koog handles this automatically: 
val settings = AnthropicClientSettings(
    modelVersionsMap = mapOf(
        AnthropicModels.Sonnet_4_5 to "claude-sonnet-4-5-20250929",
        AnthropicModels.Opus_4_6 to "claude-opus-4-6"
        // Custom mappings
    )
)

​
Tool Choice Control

val executor = simpleAnthropicExecutor(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    model = AnthropicModels.Sonnet_4_5,
    params = AnthropicParams(
        toolChoice = LLMParams.ToolChoice.Required // Auto, None, Required, Named
    )
)

​
Model Capabilities

ModelContextOutputVisionDocsToolsThinking
Sonnet 4.5200K64K✅✅✅✅
Haiku 4.5200K64K✅✅✅✅
Opus 4.6200K1M✅✅✅✅
Opus 4.5200K64K✅✅✅✅
Sonnet 4200K64K✅✅✅❌

​
Pricing

Pricing varies by model. See Anthropic Pricing for current rates. Example costs (per 1M tokens):
  • Haiku 4.5: 1(input)/1 (input) / 5 (output)
  • Sonnet 4.5: 3(input)/3 (input) / 15 (output)
  • Opus 4.5: 5(input)/5 (input) / 25 (output)
  • Opus 4.6: 5(input)/5 (input) / 25 (output)

​
Best Practices

  1. Use Sonnet 4.5 for most tasks - best balance of capability and cost
  2. Use Haiku 4.5 for high-throughput scenarios
  3. Use Opus 4.6 for maximum output length needs
  4. Leverage extended thinking for complex reasoning tasks
  5. Utilize long context for document analysis
  6. Enable streaming for better UX with long responses

​
Limitations

  • No embeddings: Use OpenAI or Google for embeddings
  • No moderation API: Implement custom content filtering
  • No structured outputs: Parse responses manually or use tool calling
  • No audio: Text and vision only

​
Troubleshooting

​
Rate Limits

val client = AnthropicLLMClient(
    apiKey = System.getenv("ANTHROPIC_API_KEY"),
    settings = AnthropicClientSettings(
        timeoutConfig = ConnectionTimeoutConfig(
            requestTimeoutMillis = 180_000 // Longer timeout for complex tasks
        )
    )
)

​
Error Handling

try {
    val result = executor.execute(prompt { user("Hello") })
} catch (e: LLMClientException) {
    when {
        e.message?.contains("overloaded") == true -> {
            // Retry with exponential backoff
        }
        e.message?.contains("invalid_request_error") == true -> {
            // Check request format
        }
        else -> throw e
    }
}

​
Resources

Build docs developers (and LLMs) love

Get started for freeTalk to us