Anthropic Provider

Setup

Set your Anthropic API key:

export ANTHROPIC_API_KEY=sk-ant-...

llmist will automatically discover and use Anthropic.

Available Models

Model	Alias	Best For
claude-opus-4-5	opus	Complex reasoning, creative tasks
claude-sonnet-4-5	sonnet	Balanced performance (recommended)
claude-haiku-4-5	haiku	Fast, cost-effective tasks

Tip

Sonnet is the recommended default—it offers the best balance of capability and cost for most use cases.

Usage Examples

Basic

import { LLMist } from 'llmist';

const answer = await LLMist.createAgent()
  .withModel('sonnet')
  .askAndCollect('Write a haiku about TypeScript');

With System Prompt

import { LLMist } from 'llmist';

const answer = await LLMist.createAgent()
  .withModel('sonnet')
  .withSystem('You are a helpful coding assistant. Be concise.')
  .askAndCollect('How do I read a file in Node.js?');

Long Context

import { LLMist } from 'llmist';
import { readFileSync } from 'fs';

const codebase = readFileSync('src/main.ts', 'utf-8');

const answer = await LLMist.createAgent()
  .withModel('sonnet')
  .withSystem('You are a code reviewer.')
  .askAndCollect(`Review this code:\n\n${codebase}`);

Extended Thinking

All Claude 4+ models support extended thinking (reasoning). llmist maps ReasoningEffort to Anthropic’s thinking.budget_tokens:

Effort	Budget Tokens
"none"	1024 (minimum)
"low"	2048
"medium"	8192
"high"	16384
"maximum"	32768

const answer = await LLMist.createAgent()
  .withModel('opus')
  .withReasoning('high')
  .askAndCollect('Explain the Riemann hypothesis.');

You can also specify an explicit token budget:

.withReasoning({ enabled: true, budgetTokens: 10000 })

Temperature

Anthropic forbids temperature when thinking is enabled — llmist automatically strips it.

For multi-turn tool use conversations, enable interleaved thinking so the model can reason between tool calls:

.withReasoning({ enabled: true, effort: 'high', interleaved: true })

See the Reasoning Models guide for full details.

Prompt Caching

Anthropic’s prompt caching reduces cost and latency by reusing previously computed context across requests. llmist enables caching automatically by adding cache_control markers to system prompts and the last user message.

No configuration needed — caching is on by default:

// Caching is automatic — repeated context is served from cache
const agent = LLMist.createAgent()
  .withModel('sonnet')
  .withSystem(largeSystemPrompt)    // Cached automatically
  .ask('Analyze this...');

To disable caching (removes all cache_control markers):

const agent = LLMist.createAgent()
  .withModel('sonnet')
  .withoutCaching()                 // No cache_control markers sent
  .ask('Quick question');

Tip

Cached input tokens are ~90% cheaper than uncached tokens. For long system prompts or multi-turn conversations, this adds up fast.

Vision (Image Input)

Claude models support image input:

import { LLMist, imageFromBuffer } from 'llmist';
import { readFileSync } from 'fs';

const imageBuffer = readFileSync('diagram.png');

const answer = await LLMist.createAgent()
  .withModel('sonnet')
  .askWithImage(
    'Describe this diagram',
    imageFromBuffer(imageBuffer, 'image/png')
  )
  .askAndCollect();

Supported image formats: JPEG, PNG, GIF, WebP

Model Characteristics

Claude Opus 4.5 Most Capable

• Best for complex reasoning and creative tasks
• Highest quality outputs
• Higher latency and cost
• 200K context window

Claude Sonnet 4.5 Recommended

• Best balance of capability and speed
• Great for coding, analysis, and general tasks
• Good cost efficiency
• 200K context window

Claude Haiku 4.5 Fastest

• Fastest response times
• Lowest cost
• Good for simple tasks and high-volume use
• 200K context window

Configuration Options

import { LLMist, AnthropicMessagesProvider } from 'llmist';

const client = new LLMist({
  autoDiscoverProviders: false,
  adapters: [
    new AnthropicMessagesProvider({
      apiKey: process.env.ANTHROPIC_API_KEY,
      baseUrl: 'https://api.anthropic.com', // Custom endpoint
    }),
  ],
});

Best Practices

1. Use Sonnet as default - Best balance for most tasks
2. Reserve Opus for complex reasoning - When quality matters more than speed
3. Use Haiku for high-volume - Simple tasks, classification, summarization
4. Leverage long context - Claude handles 200K tokens well

Cost Tracking

for await (const event of agent.run()) {
  if (event.type === 'llm_call_complete') {
    console.log('Input tokens:', event.usage?.promptTokens);
    console.log('Output tokens:', event.usage?.completionTokens);
    console.log('Estimated cost:', event.cost);
  }
}