ai_sdk_anthropic 3.0.0

Anthropic provider for AI SDK Dart. Use Claude models (claude-sonnet, claude-haiku, claude-opus) for text generation, streaming, tools, and structured output.

example/example.md

ai_sdk_anthropic examples

Installation

dart pub add ai_sdk_dart ai_sdk_anthropic

The top-level anthropic factory reads const String.fromEnvironment('ANTHROPIC_API_KEY'). Use it with fvm dart run --define=ANTHROPIC_API_KEY=sk-ant-... bin/app.dart, or read Platform.environment['ANTHROPIC_API_KEY'] yourself and pass apiKey: to AnthropicProvider in server and CLI apps.

---

Text generation

import 'package:ai_sdk_dart/ai_sdk_dart.dart';
import 'package:ai_sdk_anthropic/ai_sdk_anthropic.dart';

final result = await generateText(
  model: anthropic('claude-sonnet-4-5'),
  prompt: 'Explain quantum entanglement in one sentence.',
);
print(result.text);
print('tokens used: ${result.usage}');

---

Streaming

import 'dart:io';

final result = await streamText(
  model: anthropic('claude-sonnet-4-5'),
  prompt: 'Write a short poem about Dart.',
);
await for (final chunk in result.textStream) {
  stdout.write(chunk);
}

---

System prompt

final result = await generateText(
  model: anthropic('claude-sonnet-4-5'),
  system: 'You are a concise assistant. Reply in at most two sentences.',
  prompt: 'What is the Dart programming language?',
);
print(result.text);

instructions is the current canonical top-level instruction parameter; system remains a supported fallback and is used above for continuity with older code. Prefer instructions in new code:

final result = await generateText(
  model: anthropic('claude-sonnet-4-5'),
  instructions: 'You are a concise assistant. Reply in at most two sentences.',
  prompt: 'What is the Dart programming language?',
);
print(result.text);

---

Reasoning controls

The core reasoning: parameter (LanguageModelV4Reasoning, from package:ai_sdk_provider) works across providers:

import 'package:ai_sdk_provider/ai_sdk_provider.dart';

final result = await generateText(
  model: anthropic('claude-fable-5-1'),
  reasoning: LanguageModelV4Reasoning.high,
  prompt: 'Solve step by step: if 3x + 5 = 20, what is x?',
);
print(result.text);
print(result.reasoningText);

For Anthropic-specific control over the thinking budget, use AnthropicThinkingOptions via providerOptions:

final result = await generateText(
  model: anthropic('claude-sonnet-4-5'),
  prompt: 'Solve step by step: if 3x + 5 = 20, what is x?',
  maxOutputTokens: 16000,
  providerOptions: {
    'anthropic': AnthropicThinkingOptions(budgetTokens: 10000).toMap(),
  },
);
print(result.reasoningText);

AnthropicLanguageModelOptions wraps thinking, prompt caching, and adaptive effort together:

final result = await generateText(
  model: anthropic('claude-sonnet-4-5'),
  prompt: 'Solve step by step: if 3x + 5 = 20, what is x?',
  maxOutputTokens: 16000,
  providerOptions: {
    'anthropic': AnthropicLanguageModelOptions(
      thinking: AnthropicThinkingOptions(budgetTokens: 10000),
      cacheControl: AnthropicCacheControlOptions(ttl: '5m'),
    ).toMap(),
  },
);

---

Signed reasoning replay

Anthropic's extended-thinking blocks carry a signature on LanguageModelV4ReasoningPart so the model can verify its own prior reasoning in a follow-up turn. Don't read or reconstruct the signature yourself — carry result.responseMessages through unchanged as history and Anthropic verifies it:

const question = ModelMessage(
  role: ModelMessageRole.user,
  content: 'Solve step by step: if 3x + 5 = 20, what is x?',
);
final first = await generateText(
  model: anthropic('claude-fable-5-1'),
  reasoning: LanguageModelV4Reasoning.high,
  messages: [question],
);

final second = await generateText(
  model: anthropic('claude-fable-5-1'),
  reasoning: LanguageModelV4Reasoning.high,
  messages: [
    question,
    ...first.responseMessages.map(ModelMessage.fromProvider),
    const ModelMessage(role: ModelMessageRole.user, content: 'Now check your work.'),
  ],
);
print(second.text);

---

Tagged reasoning middleware

extractReasoningMiddleware extracts <think> tags from text. It does not enable Claude's native thinking API, which the adapter handles directly:

final model = wrapLanguageModel(
  model: anthropic('claude-sonnet-4-5'),
  middleware: extractReasoningMiddleware(tagName: 'think'),
);

final result = await generateText(
  model: model,
  prompt: 'Solve step by step: if 3x + 5 = 20, what is x?',
);
print('Answer   : ${result.text}');
print('Reasoning: ${result.reasoningText}');

---

Tools (multi-step agent)

final result = await generateText(
  model: anthropic('claude-sonnet-4-5'),
  prompt: 'What is the weather in London?',
  maxSteps: 5,
  tools: {
    'getWeather': tool<Map<String, dynamic>, String>(
      description: 'Get current weather for a city.',
      inputSchema: Schema<Map<String, dynamic>>(
        jsonSchema: const {
          'type': 'object',
          'properties': {'city': {'type': 'string'}},
          'required': ['city'],
        },
        fromJson: (json) => json,
      ),
      execute: (input, _) async => 'Cloudy, 14°C in ${input['city']}',
    ),
  },
);
print(result.text);

---

Structured output

final result = await generateText<Map<String, dynamic>>(
  model: anthropic('claude-haiku-4-5'),
  prompt: 'Give me the capital and population of France as JSON.',
  output: Output.object(
    schema: Schema<Map<String, dynamic>>(
      jsonSchema: const {
        'type': 'object',
        'properties': {
          'capital': {'type': 'string'},
          'population': {'type': 'number'},
        },
      },
      fromJson: (json) => json,
    ),
  ),
);
print(result.output); // {capital: Paris, population: 68000000}

---

Middleware — default settings

final model = wrapLanguageModel(
  model: anthropic('claude-sonnet-4-5'),
  middleware: defaultSettingsMiddleware(
    temperature: 0.3,
    maxOutputTokens: 512,
  ),
);

final result = await generateText(
  model: model,
  prompt: 'Summarise the Dart language in three bullet points.',
);
print(result.text);

---

Runnable example apps

• examples/basic — Dart CLI
• examples/advanced_app — Flutter app with Anthropic provider switcher