Claude 3.5 Haiku

Base model

Haiku 3.5 base was pretrained on data scraped from the internet up to July 2024.1 This model has made an appearance in Anthropic’s later research.234

Further reading

Appearances in Anthropic’s subsequent research:

  • Interpretability:2356
  • Alignment:748

Footnotes

  1. Claude 3 Model Card October 2024 Addendum (Anthropic, Oct 22, 2024)

  2. On the Biology of a Large Language Model (Lindsey et al., Mar 27, 2025) 2

  3. Circuit Tracing: Revealing Computational Graphs in Language Models (Ameisen et al., Mar 27, 2025) 2

  4. Unsupervised Elicitation (Anthropic, Jun 11, 2025) 2

  5. When Models Manipulate Manifolds: The Geometry of a Counting Task (Gurnee et al., Oct 21, 2025)

  6. Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations (Fraser-Taliente et al., May 7, 2026)

  7. Modifying LLM Beliefs with Synthetic Document Finetuning (Wang et al., Apr 24, 2025)

  8. Abstractive Red-Teaming of Language Model Character (Rahn et al., March 2026)