Claude 3.5 Haiku

Generates text responses with claimed sub-second latency across Anthropic-managed inference infrastructure, supporting a 200,000-token context window that enables processing of entire documents, codebases, or conversation histories in a single request. Uses proprietary transformer architecture optimized for throughput rather than parameter count, allowing rapid token generation without sacrificing context retention. Streaming output is supported for progressive response delivery.

Generates, completes, and debugs code across multiple programming languages by leveraging transformer-based pattern recognition trained on diverse codebases. Matches Claude 3 Opus performance on coding benchmarks (MMLU) and achieves 73.3% on SWE-bench Verified, indicating capability for real-world software engineering tasks including bug fixes, test generation, and refactoring. Supports tool use for executing code or querying documentation, enabling iterative debugging workflows.

Implements safety guardrails through Constitutional AI (CAI) training, which aligns the model with a set of principles to reduce harmful outputs, bias, and misuse. The model has been extensively tested and evaluated with external experts to identify and mitigate safety risks. Safety mechanisms are built into the model itself rather than as post-hoc filters, enabling safer outputs across diverse use cases.

Available through multiple deployment channels including Anthropic's native Claude Platform API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, enabling integration with diverse cloud ecosystems and enterprise infrastructure. Each deployment option provides native API integration, reducing friction for teams already invested in specific cloud providers. Pricing and availability may vary by platform.

Provides Claude Code, an integrated environment for coding tasks that combines the model with code execution, testing, and debugging tools. Enables developers to write, test, and refactor code within a single interface without switching between tools. Supports iterative development workflows where the model generates code, executes it, receives feedback, and refines based on results.

Processes images and visual documents through a multimodal transformer architecture, enabling analysis of photographs, diagrams, charts, screenshots, and scanned documents. Integrates vision encoding with text generation to produce descriptions, extract structured data, answer questions about visual content, or identify objects and text within images. Supports multiple image formats (JPEG, PNG, GIF, WebP) and can process multiple images in a single request.

Enables the model to invoke external tools or functions by parsing structured function definitions (JSON schema format) and generating function calls as part of its output. Supports native integration with Anthropic's tool-use API, allowing developers to define custom functions that the model can call autonomously. Integrates with broader agentic workflows where Haiku acts as a sub-agent executing specific tasks (classification, data extraction, API calls) orchestrated by a larger model.

Classifies text into predefined categories and extracts named entities (people, organizations, locations, dates, etc.) using transformer-based pattern recognition. Leverages structured output mode to return results in JSON or other machine-readable formats, enabling direct integration with downstream systems without parsing unstructured text. Optimized for high-throughput classification pipelines where speed and cost are critical.

Implements token-level caching of frequently-used prompts, system instructions, or document context, reducing the number of tokens billed on subsequent requests that reuse the same cached content. Caching operates at the API level and is transparent to the application; developers specify which parts of the prompt should be cached, and Anthropic's infrastructure stores and reuses them across requests. Provides up to 90% cost savings on cached tokens compared to standard pricing.

Processes multiple API requests asynchronously in batches, reducing per-request costs by 50% compared to standard API pricing. Requests are queued, processed during off-peak hours or when capacity is available, and results are delivered asynchronously via webhook or polling. Designed for non-latency-sensitive workloads (e.g., overnight data processing, bulk classification) where cost optimization is prioritized over response time.

Designed to function as a fast, cost-effective sub-agent within larger multi-agent systems, handling specific tasks (classification, extraction, API routing) while a larger model (Opus, Sonnet) orchestrates the overall workflow. Haiku's speed and cost efficiency make it ideal for high-frequency sub-tasks, while its tool-use capability enables it to execute actions autonomously. Integrates with broader agentic frameworks via standard API patterns.

Combines vision capability with tool use to enable the model to interact with computer interfaces, including web browsers, desktop applications, and command-line tools. The model can view screenshots, identify UI elements, and generate tool calls to click buttons, type text, or execute commands. This enables automation of repetitive UI-based tasks without requiring explicit programming of interaction sequences.

Generates and understands text in multiple languages, enabling global applications and cross-lingual workflows. The model can translate between languages, answer questions in non-English languages, and generate content in diverse linguistic contexts. Specific language coverage is not detailed in documentation, but the platform supports multilingual capabilities.