Claude Haiku 3.5

Sobre Anthropic

Fundado em 2021 · San Francisco, USA

Anthropic was founded in 2021 by Dario Amodei (CEO), Daniela Amodei (President) and a group of ex-OpenAI researchers. The company operates as a Public Benefit Corporation with a stated mission to build reliable, interpretable and steerable AI systems. Anthropic's research output includes 'Constitutional AI' (2022), 'Scaling Laws for Neural Language Models' (2020), interpretability work on Sparse Autoencoders and Toy Models of Superposition, and the Responsible Scaling Policy that defined ASL safety levels. The Claude product line launched in March 2023 and now spans the Haiku (small/fast), Sonnet (balanced) and Opus (frontier) tiers across multiple generations. Claude 3.5 Haiku was released in late 2024 as the speed-optimised entry tier of the 3.5 family. Investors include Google, Amazon, Spark Capital and Lightspeed, with total disclosed funding exceeding $15 billion. Anthropic is one of the most influential frontier-lab brands and a key participant in industry safety commitments with the US AISI and UK AISI institutes.

Arquitetura

Decoder-only Transformer (compact, latency-optimised)

Claude 3.5 Haiku is the lightweight, fast tier of the Claude 3.5 family, designed for low-latency and high-throughput workloads where Sonnet-level capability would be cost-prohibitive. Architecturally it is a decoder-only Transformer trained on a similar curated corpus to the larger Claude models, including web text, books, source code and licensed datasets, with a knowledge cutoff in mid-2024. Compared to Claude 3 Haiku, Anthropic reports that 3.5 Haiku outperforms the previous flagship Claude 3 Opus on many benchmarks while keeping the cost profile of a small model, achieved through improved data mixture, better post-training and architecture refinements. Post-training combines supervised fine-tuning, RLHF and Constitutional AI critiques. The model supports tool use, JSON mode, vision input and structured output but does not expose an extended thinking mode (that ships in Claude 3.7 and 4). Haiku 3.5 is heavily optimised for inference throughput, often running 2-3x faster than Sonnet on the same hardware, making it suitable for real-time customer-facing applications and high-volume data pipelines. Anthropic published a model card detailing safety evaluations and behavioural changes versus the previous Haiku generation.

Parâmetros

Undisclosed (estimated low tens of billions, dense)

Contexto

200.000 tokens

Capacidades

• Sub-second time-to-first-token for short prompts
• Strong instruction following at small model size
• 200K context window for long-document tasks
• Vision support for images, charts and PDFs
• Tool use and structured JSON output
• High throughput suitable for real-time chat workloads
• Outperforms Claude 3 Opus on several reasoning and code benchmarks despite smaller size
• Multilingual coverage of major world languages
• Constitutional-AI tuned for safe, helpful responses
• Cheap enough for use as a router/classifier in agent stacks
• Best for: high-volume chat, classification, summarisation, light coding tasks.

Treinamento & licença

Trained on Anthropic's curated mixture of web, books, code and licensed data, with a knowledge cutoff around July 2024. Post-training uses RLHF plus Constitutional AI critiques on a much smaller human-rated set.

Licença: Proprietary commercial license via Anthropic API, Amazon Bedrock and Google Vertex AI.

Testes de segurança: Internal red team plus external evaluators tested for misuse and dual-use risks; classified under Anthropic's standard ASL-2 deployment safeguards.